Biotech is entering its AI gold-rush moment.

This is not a criticism. Prospectors can get rich. Sometimes they find gold. Sometimes they get dysentery. Sometimes they buy a lot of equipment from people who get rich either way.

You have heard it before: during a gold rush, sell picks and shovels.

Fine. AI-bio needs picks and shovels: better sequencing, cheaper synthesis, automated labs, cleaner data infrastructure, better assays, model tooling, workflow software, and all the dull-but-essential plumbing that turns biology from craft into industry.

But the point of a gold rush is not to own a shovel.

The point is to find gold.

And in AI-bio, the durable value may not come from selling tools to every miner. It may come from finding the next vein before anyone else realizes the mountain contains one.

The center is crowded

AI is good at exploiting abundant data.

That is one reason protein folding became the great proof-of-concept for modern AI in biology. There was a clear problem, a benchmark, structural data, sequence data, and decades of scientific effort making the problem legible. Then the models arrived, and the results were spectacular.

But a solved structure is not a drug.

A fold is not a therapy. A predicted binder is not a medicine. A pretty interface is not a clinical effect. Biology, being biology, reserves the right to humiliate those that confuse molecular plausibility with functional reality.

This is the central problem in AI-bio. The field has data. Lots of it. But not all data is equally valuable.

Sequence databases, structures, transcriptomes, single-cell atlases, CRISPR screens, and binding datasets all matter. They will continue to matter. But the middle of the map is crowded with companies mining the same public resources and making similar claims about foundation models, design loops, and programmable biology.

Some of these efforts will matter. Some will be extraordinary.

But the center is crowded.

The frontier is where things get interesting.

A lesson from bacteria

Oskar Hallatschek and colleagues once mixed bacteria labelled with different fluorophores and let them grow on an agar plate.

At the start, the colony was just a jumbled mix of colors. But as it expanded, the edge formed striking radial sectors each spreading outward like a wedge.

The lesson was not that one bacterium had a better pitch deck.

The lesson was that frontiers are weird.

At the frontier, small early advantages compound. A cell that happens to sit at the growing edge gets access to open space. Its descendants inherit that position. What begins as a tiny wedge becomes an enduring advantage.

That is a useful picture for AI-bio.

At the center, there is density, activity, and competition. At the edge, there is uncertainty and open space. It is harder to measure, noisier, and riskier. But if you get a position there, the advantage can compound.

That is what a proprietary data engine can be: a frontier position that turns into a huge territory.

The data that matters

The mistake is to say AI-bio needs more data.

True, but not enough.

AI-bio needs the right kind of data.

A useful dataset answers two questions.

First: does it tell you what to build?

Does it help choose a sequence, receptor, binder, perturbation, RNA element, protein interface, chemical scaffold, pathway design, or cell construct?

Second: does it tell you what that design actually does?

Does it connect the molecular change to a functional biological outcome?

Most datasets do one better than the other.

Structures tell you about shape and plausible interfaces. But structure alone does not prove function.

Binding data tells you what sticks to what. Important. But binding is not function.

CRISPR screens tell you what genes matter. Powerful. But knowing that a gene matters does not necessarily tell you what molecule to build.

Transcriptomes and single-cell atlases show biological state. Rich, useful, often beautiful. But mostly observational.

The frontiers that matter do both: data that links molecular design variables to causal functional outcomes.

Not just what exists.

Not just what binds.

Not just what correlates.

What designed molecular change causes the function we want?

That is the good stuff.

Why TCR-pHLA is interesting

At Synteny we focus on TCR-pHLA therapeutics.

A TCR is a designable molecule. A peptide-HLA complex is a target. The output is not merely whether the two touch. The output is whether a T cell does something therapeutically meaningful.

That puts TCR-pHLA in an unusually valuable part of the map.

TCR specificity data already has molecular design value because it connects receptor sequence to antigen recognition. But specificity is not the finish line. A TCR that recognizes an antigen is not automatically a good therapeutic TCR.

It may bind with the wrong kinetics. It may cross-react. It may fail to signal. It may signal too strongly. It may work in an artificial assay and fail in a real cellular context.

The therapeutic question is not simply:

Does this TCR bind this pHLA?

The therapeutic question is:

Does this TCR, in this cellular context, cause the immune function we want while avoiding the functions we fear?

That is a much harder question.

It is also a much more valuable one.

Because if you can generate data that connects TCR sequence, peptide-HLA recognition, and immune-cell function at scale, then you are no longer just building an assay.

You are building a frontier data engine.

Hard is the point

This is the part where optimism needs to put on a seatbelt.

TCR-pHLA is a monster combinatorial problem.

There are many possible TCRs. Many peptides. Many HLA alleles. Many antigen densities. Many cellular contexts. Many functional outputs. Many safety liabilities. Many near-misses that are scientifically interesting and therapeutically useless.

This is not a simple search problem.

Which is exactly why it matters.

Small, obvious spaces do not create durable AI companies. If the answer can be found with a modest screen and a spreadsheet, the moat is shallow.

The valuable frontiers are too large for intuition and too complex for one-off experiments. They require data engines. They require systematic exploration. They require assays that can return functional labels at scale. They require models that suggest what to test next. They require metadata, controls, repeatability, and negative examples.

Everyone likes to talk about the model.

The model is not enough.

The model needs data at the frontier to feed it.

The gold

So yes, AI-bio is a gold rush.

And yes, picks and shovels matter.

But the biggest question is still where the gold is.

Existing biological data will keep powering models. Public datasets will improve. Foundation models will get better. The center of the colony will become more crowded, more active, and more sophisticated.

But enduring value may come from frontier positions where new data can be generated before the rest of the field arrives.

Hallatschek’s bacteria remind us that the frontier is not just the edge of the picture. It is where small early advantages become lasting structure.

For Synteny, the opportunity is to occupy one of those frontier positions in TCR-pHLA: to connect receptor design, antigen recognition, and immune-cell function in a scalable data engine.

Amazing if solved.

Hard to solve.

Which is another way of saying: worth paying attention to.

The importance of this distinction is underscored by cases such as MAGE-A3/Titin. In that case, the engineered TCR recognised the intended target, but also cross-reacted with an unrelated peptide derived from Titin, leading to severe toxicity.

We have seen a similar challenge in our own work on KRAS-targeting TCRs. Our aim is to engineer receptors that bind mutant KRAS G12D and G12V peptide-HLA complexes, whilst avoiding recognition of the corresponding wild-type peptide. In practice, this is a difficult specificity problem because the mutant and wild-type peptides are highly similar, and improvements in mutant binding can easily come with increased wild-type recognition. Simple optimisation for mutant binding alone is insufficient.

Early attempts to optimise across multiple KRAS mutants showed that the design objective had to be specified carefully. Rather than encouraging binding to G12D and G12V, the model also had to preserve the right dependence on the mutant KRAS context, instead of finding superficially high-scoring solutions that did not reflect the specificity we wanted. In practice, this meant revising the guidance so that multiple objectives were handled as a constrained design problem, rather than by naïvely summing scores across desired targets.

Aaron Sim’s recent discussion of inference-time control for diffusion models (Synteny substack, Sept 2025) is a perceptive approach for exactly this reason. He argues that conditional generation should be understood in terms of the distribution being sampled, rather than in terms of heuristic edits made during generation. In diffusion models, the aim is to sample from a distribution that reflects both the learnt prior over valid biological objects and a set of constraints that describe the desired output. For TCR design, the prior captures the statistical regularities of plausible receptor sequences or structures, while the constraints describe the programme-specific requirements imposed by the target, the off-target landscape, and any additional properties needed for development.

Let u denote a TCR sequence, Y+ a set of peptide-HLA targets, and Y- a set or distribution of peptide-HLA complexes that should be avoided. A useful abstraction for the desired sampling distribution is

where p(u) is the learnt receptor prior and the conditional terms encode target recognition, off-target avoidance, and other design constraints. The terms in this expression need not correspond to a single assay or model. They may be built from affinity predictors, activation models, structural compatibility scores, peptide scanning data, self-peptide panels, tissue expression information, or multiplexed measurements of receptor binding. The main point is that these quantities influence the distribution from which candidates are generated, rather than being reserved solely for ranking candidates after generation.

This distinction matters because TCR datasets contain substantial structure that can easily become conflated with the intended biological condition. Available data are shaped by antigen choice, HLA representation, assay format, disease context, and the historical emphasis on experimentally convenient targets. A receptor intended for a tumour associated peptide, for example, should be guided by the conditional distribution relevant to that peptide HLA complex and its associated liabilities, rather than by the most common specificities represented in the training set.

This is the same statistical issue that appears in simpler constrained sampling examples. When a condition selects a region that differs substantially from the marginal distribution learned from data, heuristic conditioning methods can drift toward samples that are likely under the prior while failing to represent the desired conditional distribution. In TCR generation, this failure mode would appear as plausible looking receptors that reflect the broad composition of the dataset more than the target and safety profile required for a therapeutic programme. The problem is particularly relevant when the peptide or HLA is under-represented, or when the important negative constraints are defined by rare but clinically meaningful cross reactivities.

Training-free guidance is attractive in this setting because it separates the learning of the receptor prior from the specification of the design objective. The base model learns the grammar of receptor space from available sequence or structure data. At inference-time, the sampling process is modified by conditional terms that are chosen for a particular programme. One term may favour a desired affinity range for the intended peptide HLA complex. Another may penalise predicted binding to a panel of self peptides, related peptides, or experimentally observed cross binders. Further terms may preserve similarity to a known parent receptor, restrict changes to selected CDR regions, discourage sequence features associated with poor expression, or favour diversity among candidates that satisfy the same specificity profile.

This modularity is useful because the design objective changes over the course of a TCR programme. Early stages may require broad exploration around a target peptide HLA complex, especially when few natural receptors are known. Lead optimisation may require local changes around a characterised clone while preserving parts of the binding mode. A specificity screen may identify a new liability that should be incorporated into the next design round. Developability measurements may impose additional constraints after a promising specificity profile has been found. A framework that imposes these requirements at inference-time allows the same generative prior to be reused as the conditional information changes.

The approach also fits naturally with high throughput experimental measurement. One concrete example of this was discussed in an earlier post on ARLO’s self-calibrating student-teacher training, in which the ranking and calibration objectives used to turn noisy MYRIAD sequencing counts into affinity estimates for TCR–pMHC interactions were derived. Multiplexed library assays can generate interaction data across large sets of receptors and peptide-HLA complexes, while models trained on these data can provide affinity or activation estimates for combinations that have not been directly measured. These learnt scoring functions can then serve as guidance terms during sampling. As new off-targets are discovered experimentally, they can be added to the negative condition. As target binding models improve, the positive condition can be sharpened. As peptide presentation or tissue expression estimates become available, the off-target distribution can be reweighted to reflect biological risk.

For TCR discovery, this hints at a shift from generating plausible receptors followed by increasingly elaborate rejection steps toward sampling from programme specific distributions that already encode the desired recognition profile. Such a framework remains dependent on the quality of the underlying assays, the calibration of the scoring models, and the biological relevance of the negative set, but it provides a principled way to connect those sources of information to candidate generation. As datasets become larger and more informative, the value of generative models will depend increasingly on how precisely they can be conditioned on the full therapeutic objective, including the interactions that must be avoided.

Imatinib, the first precision medicine approved for limited oncology indications, targets the BCR-ABL fusion protein produced by a chromosomal translocation, provides 10-year survival with many patients achieving deep molecular remission, and validated precision medicine for genetic cancer driver hypothesis. The broadly applicable precision medicine approach of KRAS targeting has the best claim to becoming the backbone for precision oncology, analogous to PD-1 inhibitors, not because KRAS is “one drug for everyone,” but because KRAS is the most universal molecular anchor around which scalable, multi-tissue precision strategies can be developed.

While KRAS has long been pursued intensely as a target for small molecule inhibition, its role in precision oncology is unlikely to be amplified through this modality alone due to the multitude of resistance mechanisms tumours deploy against precision therapies. The “next PD-1” moment for KRAS will come when it is exploited as a universal intracellular antigen, enabling scalable immune redirection via TCR T cell engagers (TCR TCEs). The biology of KRAS almost inevitably points toward immune-based solutions.

Recent acquisition speculation around Revolution Medicine, the number of completed acquisition, major licensing deals and strategic partnerships, and the number of large pharma with late-stage programs targeting KRAS hints at the significant role of KRAS. KRAS is not a single target, but rather the foundational node for multi-year, multi-combination precision oncology franchises within large pharma.

The universality of KRAS

The defining feature of KRAS is its reach. Across solid tumours, KRAS is one of the most frequently mutated oncogenes, with mutations concentrated in a small number of recurrent hotspots. Pancreatic ductal adenocarcinoma is dominated by KRAS mutations, colorectal cancer shows KRAS alterations in roughly half of patients, and lung adenocarcinoma, uterine cancers, and others contribute meaningfully to the total KRAS-mutant population.

Critically, this universality is not driven by a single allele. While much attention has been paid to G12C with the first approved therapies, this mutant represents a minority of KRAS-mutant disease overall. G12D and G12V together account for the largest by fraction of KRAS mutations across tumour types, with G13D and G12R adding further breadth depending on tissue context. KRAS is not one mutation, but rather a family of dominant oncogenic states that recur across lineages.

It is precisely this universality that makes KRAS so valuable as a foundation for precision oncology: few molecular targets combine similar degrees of prevalence, recurrence, and biological centrality.

Small-molecule KRAS inhibitors proved feasibility, but exposed the limitations of the modality

For decades KRAS was the canonical undruggable oncoprotein. The breakthrough was allele-specific covalent inhibition of KRAS G12C. Approved therapies sotorasib and adagrasib demonstrated that KRAS can be drugged safely at scale, and switching it off produces real tumour regressions in defined populations.

However, those same trials also revealed the structural limitation of the approach. Responses are often partial. Durability is inconsistent. Resistance emerges through pathway reactivation, feedback signalling, and lineage-specific escape mechanisms.

This is not a failure of medicinal chemistry but a reflection of KRAS biology. KRAS is a signalling hub embedded in redundant and adaptive networks. Turning it off transiently is rarely clinically sufficient. Importantly, this dynamic was visible early, and one of the reasons why positioning inhibitors as the ultimate solution is misguided.

Immune redirection against KRAS, however, can achieve the durability of response that has plagued the inhibitors.

KRAS as an intracellular antigen: the universal kill-signal

KRAS mutations generate shared, tumour-specific neoantigens. These mutant peptides are processed intracellularly and presented on HLA molecules at the tumour cell surface. Unlike many lineage markers or differentiation antigens, KRAS-derived neoepitopes are:

• tumour-specific

• recurrent across patients

• biologically constrained (tumours cannot easily discard KRAS without paying a fitness cost).

This makes KRAS an unusually attractive target for TCR-based immunotherapies, which can recognize peptide–HLA (“pHLA”) complexes derived from intracellular proteins, an antigen repertoire uniquely designed for TCR targeting which have a strong clinical track record and the strongest efficacy signal relative to less effective antibodies and CARs in this target space.

Clinical and translational work over the past several years has validated the core premise: KRAS mutant peptides are naturally presented, and T cells can be isolated that specifically recognize them. This creates a path not merely to personalized therapies, but to off-the-shelf immune redirection against one of the most universal oncogenic drivers in cancer.

TCR T cell engagers as combination partners to inhibitors are the real platform

TCR T cell engagers (TCR TCEs) represent a fundamentally different way to attack KRAS. Rather than suppressing KRAS signalling and hoping the tumour does not adapt, TCR TCEs convert KRAS into a kill signal, redirecting endogenous T cells to selectively eliminate KRAS-mutant cells.

While the inhibitors provide a transient block in carcinogenic signalling that cells find many ways of overcoming, immune targeting carries the promise of completely eliminating the cells that carry the oncogenic mutation.

Furthermore, remodelling of the tumour microenvironment (TME) is observed in otherwise cold and immunosuppressive KRAS-driven tumours following treatment with (K)RAS inhibitors. Upon KRAS inhibition, tumour cells upregulate MHC class I-related pathway components and antigen presentation, and cytotoxic T cells infiltrate the tumour. Cancer cell death further increases antigen presentation by antigen-presenting cells (APCs) now present in the TME, in turn enhancing T cell recognition and activation.

This additionally supports the universality of a TCR-TCE modality: as the modulation of the TME and enhanced antigen presentation occur regardless of the mechanisms of KRAS inhibition, the TCR TCE modality is positioned as a highly promising combination therapy partner to most inhibitors. It capitalises on the vulnerability of the tumour exposed by KRAS inhibition and avoids potential toxicities from co-targeting a pathway in the vertical signalling axis. It may indeed prove a better combination therapy partner than immune checkpoint blocking agents, as their systemic effect often leads to toxicities, and TCR-TCE are guiding tumour T cells to specifically kill tumour cells rather than broadly activating the immune system.

Several features thus make TCR TCEs uniquely suited to become the KRAS backbone:

1. Mechanistic durability
   Immune-mediated killing does not require sustained pathway suppression. Once a KRAS-mutant cell presents the target peptide, it becomes vulnerable to cytotoxic elimination.

2. Pan-tumour applicability
   Because KRAS mutations recur across tissues, the same KRAS-derived antigens can be targeted in pancreatic, colorectal, lung, and other cancers.

3. Combination leverage
   TCR TCEs naturally complement targeted agents, checkpoint inhibitors, and microenvironment modulators, mirroring how PD-1 became more powerful as combinations matured.

4. Platform extensibility
   KRAS can serve as the anchor target in a broader intracellular antigen strategy, enabling expansion into additional peptides that address heterogeneity and resistance.

Addressing the HLA question head-on

HLA restriction is the most common objection raised against TCR-based therapeutics. Historically, each TCR recognized a single peptide presented by a single HLA allele, fragmenting patient populations.

With the right data and model, we can now leverage frontier AI to program TCRs for multiple HLA recognition. Synteny has explicitly taken on this challenge by designing TCR T cell engagers that bind multiple KRAS-derived peptides presented across multiple HLA alleles. Rather than building a one-TCR-per-HLA paradigm, our approach treats HLA specificity as a design constraint to be solved upfront.

Multi-HLA binding transforms KRAS from a collection of niche programs into a genuinely scalable target that can support population-level deployment similar to PD-1.

pan-KRAS as Synteny’s first anchor in oncology

Aside from being an unarguably attractive biological and commercial target, our decision at Synteny to make pan-KRAS our first target in oncology was premised on KRAS being the anchor of a large collection of off-the-shelf TCR molecules as part of a precision oncology strategy.

In service of building a large library of TCRs targeting differentially expressed intracellular peptides, our first program must:

• validate our platform’s ability to program potent and safe biomolecules

• establish regulatory and clinical credibility

• anchor the platform in a target with maximal relevance.

KRAS satisfies all three conditions better than almost any other intracellular antigen family. Its mutations are universal enough to justify investment, constrained enough to prevent easy escape, and central enough to tumour biology to remain relevant across lines of therapy.

Once KRAS is established as the anchor, additional intracellular targets increase in value, not only as standalone programs, but as cocktails in a coherent precision immunotherapy strategy.

The strategic backdrop: why large pharma cares

The industry’s interest in KRAS platforms is evident. Ongoing discussions around the potential acquisition of Revolution Medicines and a recent standing ovation at ASCO following data on inhibition of KRAS in pancreatic cancer underscore how strategically valuable KRAS has become.

However, ownership of KRAS signalling inhibitors is only one dimension. The larger opportunity is to own the KRAS axis itself: signalling, immune recognition, combinations, and resistance management. In that framework, immune redirection against KRAS is the inevitable next layer.

KRAS as the next PD-1

KRAS is the most universal, biologically constrained, and scalable intracellular target in solid tumours. Small-molecule inhibitors proved KRAS could be targeted but they also revealed the limits of suppression alone. The future belongs to strategies that weaponize KRAS mutations as immune targets.

TCR T cell engagers, especially those engineered to span multiple HLA alleles, convert KRAS from a signalling node into a precision immunotherapy backbone. Any comprehensive, precision oncology roadmap will ultimately start with KRAS at its foundation.

If PD-1 defined the first era of immuno-oncology, KRAS, through TCR-based immune redirection, is poised to define the next.

With thanks to for her input in this blog.

Different sequencing technologies offer distinct advantages, while suffering from technology-specific shortcomings. Short read approaches offer unparalleled accuracy and output levels, but handle longer molecules (>1kb) and in particular long amplicons quite poorly. Long read technologies, in contrast, offer the ability to sequence longer molecules, but at much lower throughput and in some cases accuracy. Costs and availability of sequencing approaches also differ widely. Identifying and deploying the most suitable sequencing technology and approach for a particular assay ensures reliable data outputs and fast turnaround times.

Given the size of our libraries which routinely have ~1M individual sequences, we need our sequencing outputs to be extremely reliable in order to consistently deliver qualitatively and quantitatively meaningful outputs required for ML model training. Due to the nature of our TCR constructs, multi-cystronic ~2kb synthetic molecules, we have initially found ourselves in a catch-22 situation: the molecule we had to sequence as our assay output was far too long for the most commonly used short read approaches, and the number of samples and the coverage required were too high for any long read approach to handle without incurring astronomical costs and unacceptable delays. For months, our outputs were falling short of quality and quantity targets. Data was unreliable, both quality and coverage varied widely. While making some adjustments to the sequencing library preparation protocol and the choice of the sequencing instrument offered improvements to data quality, the necessity to use external facilities and the costs, waiting times and data transfer issues associated with that were considerably slowing us down.

To address the problem, we introduced changes to the assay and the way we normalise the data. This has drastically reduced the required coverage, thus enabling us to switch fully to Nanopore sequencing. This technology handles the length and, crucially, amplicon nature of our constructs with ease. Nanopore sequencers are very accessible, and from this point on we were able to bring all our sequencing in-house. This gives us full control over the process and has drastically reduced our timelines whilst significantly improving the quality and coverage of our outputs.

Bringing sequencing in-house was only part of the story — we also needed a robust, efficient pipeline to process the data it generates. We invested heavily in optimising every stage of our Nanopore data processing, from raw signal to variant counts, with the goal of maximising both throughput and accuracy. A key area of investigation was basecalling: the step in which the electrical signals produced by the sequencer are translated into nucleotide sequences. Nanopore offers several basecalling models that sit on a spectrum of speed versus accuracy, and choosing the right one has a material impact on the quality of downstream results.

We systematically compared the High Accuracy (HAC) and Super Accurate (SUP) basecalling models across our sequencing runs. As shown in the figure below, the per-read quality score distributions differ markedly: SUP basecalling produces a pronounced shift toward higher Q-scores, substantially outperforming HAC. Importantly, the accuracy achieved by SUP is not far off what we previously obtained with Illumina short-read sequencing. On the basis of these results, SUP basecalling is now the default in our production pipeline, giving us confidence that the switch to fully in-house Nanopore sequencing comes with no compromise on data quality.

This uplift in per-read accuracy has a direct and significant impact on how we analyse the data downstream. Because SUP basecalling produces reads with very few errors, the vast majority of our sequences are close enough to the expected reference to be identified through exact string matching. In practice, this means we can rely on algorithms like Aho-Corasick — a highly efficient multi-pattern search method — to match reads against our construct libraries in a single pass. Compared to alignment-based approaches such as BLAST, which must account for insertions, deletions, and substitutions at every position, exact matching is orders of magnitude faster and requires a fraction of the compute. For a pipeline that routinely processes millions of reads across dozens of samples, this difference is transformative: what would otherwise demand expensive GPU-accelerated alignment completes in minutes on a modest CPU. The higher the basecalling quality, the larger the proportion of reads that resolve through this fast path, making SUP not just an accuracy improvement but an infrastructure cost saving as well.

Originally conceived as a strategy to redirect cytotoxic T cells against tumors expressing tumor-associated antigens in oncology, TCR-TCEs have demonstrated a compelling therapeutic logic that extends well beyond cancer: their exquisite sensitivity to intracellular peptide targets presented on HLA (Human Leukocyte Antigens) molecules makes them uniquely suited to autoimmune indications, where aberrant self-reactive T and B cell populations and dysregulated antigen presentation drive pathology.

In oncology, TCR-TCEs enable the targeting of intracellular neoantigens and tumor-specific peptides invisible to conventional antibody-based therapeutics, dramatically expanding the druggable proteome. In autoimmunity, they open the possibility of selectively eliminating autoreactive T and B cell clones or tolerizing pathogenic immune responses with a precision that small molecules and broad biologics cannot match.

This era of precision targeting has emerged from converging advances in the TCR–peptide–HLA field. High-resolution structures of ternary TCR–peptide–HLA complexes have revealed the molecular grammar of antigen recognition, while pHLA multimer technologies and high-throughput single-cell TCR sequencing now allow systematic mapping of antigen-specific T cell repertoires.

Computational predictors of pHLA binding (NetMHCpan/NetMHCIIpan and successors MHCflurry 2.0, HLAthena, HLApollo, and MixTCRpred), combined with deep learning models trained on structural, immunopeptidomic, and functional datasets, are accelerating the discovery of targetable peptide epitopes.

Advances in HLA typing and population-scale immunopeptidomics are clarifying shared versus patient-specific targets, informing both TCR-TCE design and patient stratification. Together with improved understanding of the immunological synapse and the very low antigen density required for T cell activation, these tools are revealing the molecular grammar of antigen recognition with unprecedented resolution.

Against this backdrop, TCR-TCEs represent an information-dense modality in drug discovery. The non-linear relationship between sequence, structure, and function in the TCR–pHLA axis means that small changes in CDR loop composition, peptide identity, or HLA alleles can produce dramatic, non-intuitive shifts in binding affinity, cross-reactivity, stability, and T cell activation. The triadic system of TCR-peptide-HLA introduces combinatorial complexity making TCR-TCEs the ultimate AI challenge: classical structure-activity relationship models break down, and success demands the kind of multi-modal, context-aware learning that next-generation AI platforms, like Synteny’s, are built to provide.

Critically, TCR-TCEs also offer an opportunity to extend and generalize antibody-focused AI models, as they share antibody-like developability risks - including aggregation propensity, off-target binding, and stability concerns - while simultaneously imposing small molecule-like specificity constraints driven by the precise geometry of peptide presentation. Their pharmacokinetics and pharmacodynamics are deeply biology-driven, governed by T cell trafficking, MHC density, and the kinetic parameters of TCR engagement and T cell activation, demanding mechanistic PK/PD modeling integrated with molecular design. Together, these features make TCR-TCEs both a rigorous test case and a generative frontier for AI-driven drug discovery. Mastering TCR prediction would mean mastering high-dimensional immune recognition.

As immunologist drug developers, we have spent years watching the autoimmune treatment landscape evolve from broadly immunosuppressive therapies to increasingly targeted approaches. TCR-based therapeutics represent not just an incremental improvement, but a paradigm shift in how we should think about treating autoimmune diseases, especially in otherwise healthy individuals.

In cancer, TCR-TCEs have delivered efficacy in the clinic with an improved safety profile relative to surface antigen-TCE in solid tumors. We now have that same opportunity in autoimmune diseases, often driven by oligoclonal expansions of T cells recognizing specific self-peptides presented in defined HLA contexts, to address the disease at its informational root.

A fundamental question we should ask is this: if you’re an otherwise healthy person whose only medical problem is an overactive immune response to a specific set of self-antigens, should your treatment carry the same infection risks and systemic toxicities as chemotherapy? The answer, increasingly, is no. TCR-based therapies may finally allow us to deliver on that promise.

The Current Landscape: Broad Hammers for Precise Problems

Today’s autoimmune treatment arsenal is dominated by broadly acting biologics. Adalimumab (Humira), the world’s best-selling drug for much of the past decade, works by blocking tumor necrosis factor-alpha (TNF-α) across the entire immune system. While revolutionary when introduced, its boxed warning tells the story: serious infections including tuberculosis reactivation, invasive fungal infections, and a two-fold increase in serious infection rates compared to control populations. Analysis of over 23,000 patients across clinical trials showed serious infection rates of 5.1 per 100 patient-years in rheumatoid arthritis patients, with cellulitis and pneumonia being the most common serious infections. The drug carries additional warnings for lymphoma, congestive heart failure, and paradoxically, even the development of new autoimmune conditions like lupus-like syndrome. Nevertheless, this and other anti-TNF-α biologics like Infliximab (Remicade), Certolizumab pegol (Cimzia), Golumumab (Simponi) and Etanercept (Enbrel, soluble TNFR) are approved for RA, AS, PsA, Plaque PsO, JIA and some IBD, share similar risks of increased infection, malignancy, induction of additional autoimmune disease with Enbrel showing the lowest risk.

Dupilumab (Dupixent), a recently approved IL-4/IL-13 blocker approved for atopic dermatitis and other type 2 inflammatory conditions, represents a step towards greater specificity. By targeting the IL-4 receptor alpha subunit, it blocks type 2 inflammation more selectively than broad immunosuppressants. However, recent pharmacovigilance data from over 37,000 adverse events reveal an interesting pattern: while dupilumab avoids traditional humoral autoimmune complications, it appears to skew immune responses toward IL-23/IL-17 pathway-related diseases. Seronegative arthritis showed a 9.61-fold increased odds ratio, while enthesitis and enthesopathy showed a striking 12.65-fold increase. The message is clear: even “targeted” therapies that broadly modulate major cytokine pathways can create unexpected immune imbalances.

Guselkumab (Tremfya), an IL-23 inhibitor approved for psoriasis and psoriatic arthritis with trials ongoing in lupus nephritis, exemplifies the limitations of even modern targeted approaches. In the ORCHID-LN Phase 2 trial for lupus nephritis, guselkumab failed to demonstrate superior reduction in proteinuria versus placebo, despite IL-23’s well-established role in autoimmune pathology. The IL-23/Th17 pathway clearly plays a role in systemic autoimmunity, but simply blocking this cytokine across all immune compartments may not be sufficient - or may even be counterproductive - in complex autoimmune diseases.

The pattern is consistent: we keep learning that immune biology is more nuanced than our therapies can accommodate. Blocking entire cytokine pathways affects every immune cell expressing those receptors, whether those cells are causing disease or protecting against it. The difference is profound: cytokine blockade modulates consequences while TCR targeting modulates causality.

Why TCR Specificity Changes Everything

T cell receptors are nature’s solution to the precision problem. Each TCR recognizes a specific peptide-MHC (major histocompatibility complex) combination with exquisite specificity, analogous to how antibodies recognize antigens, but with a crucial difference. While antibodies bind to surface proteins, TCRs can recognize any protein that gets processed and presented by MHC molecules, including intracellular proteins. This allows TCR-based therapies to access the vast proteome that antibodies cannot reach.

For autoimmune diseases, this specificity offers transformative potential. Rather than broadly suppressing TNF-α everywhere or modulating IL-23 signaling across all T cells, TCR-based approaches can theoretically target only the T cells driving pathology while leaving protective immunity intact. This kind of profile has the potential to lead to a significantly safer therapy.

This idea is not theoretical: we’re seeing proof of concept in oncology right now. Tebentafusp (Kimmtrak), the first TCR bispecific T-cell engager approved by the FDA, demonstrates what’s possible. This molecule consists of a soluble TCR specific for gp100 peptide presented on HLA-A*02:01 fused to an anti-CD3 effector domain. Despite redirecting T cells to kill target cells - a mechanism that historically produced severe cytokine release syndrome with other bispecific formats - tebentafusp showed a remarkably manageable safety profile. In the pivotal Phase 3 trial in metastatic uveal melanoma, while 89% of patients experienced cytokine release syndrome, only 0.8% had grade 3 or 4 events, and importantly, there were no grade 4 or 5 CRS events. The majority of CRS events occurred with the first three infusions, and the median time to resolution was just 2 days. This is striking when compared to some CAR-T therapies or antibody-based bispecifics targeting solid tumors, where CRS can be life-threatening and prolonged.

The tebentafusp experience teaches us several crucial lessons applicable to autoimmune disease. First, TCR-based recognition does not inherently cause unmanageable toxicity - the specificity of the targeting matters enormously. Second, dose escalation strategies can significantly mitigate adverse events, with tebentafusp using a stepped dosing approach of 20 μg, then 30 μg, then 68 μg weekly to reduce toxicity. Third, even with a bispecific format designed to robustly activate T cells, on-target toxicity can be acceptable if the target is truly specific to pathogenic cells.

The Safety Imperative for Autoimmune Patients

This brings us to perhaps the most important consideration: the patient population. Oncology patients facing metastatic disease have a different risk-benefit calculus than autoimmune patients. A patient with metastatic uveal melanoma and a median survival of 16 months without treatment will rationally accept higher risks for a therapy offering a median overall survival of 21.7 months. But what about a 32-year-old with moderate rheumatoid arthritis who is otherwise healthy and could live another 50 years? Or a patient with early lupus nephritis trying to preserve kidney function?

These patients need treatments that they can take for years, or decades, without accumulating prohibitive infection risk or malignancy risk. Current broadly immunosuppressive approaches force an uncomfortable tradeoff: accept ongoing disease activity and joint damage, or accept 5-10 serious infections per 100 patient-years of treatment plus uncertain long-term cancer risk. For young, otherwise healthy individuals, neither option is acceptable.

TCR-based approaches offer a path out of this dilemma through several mechanisms. First, by targeting only disease-relevant T-cell clones, they preserve the vast majority of immune function. Second, by using transient engagement or regulatable systems, they can potentially allow immune reconstitution between treatment cycles. Third, by avoiding the need for continuous broad immunosuppression, they reduce cumulative infection risk over decades of treatment. Fourth, the data from B cell depletion mediated by CAR-T or pan-B cell CD3 bispecifics has shown us that immunological reset is a possibility that is even suggestive of a cure opportunity for these patients.

Comparing Specificities: TCRs vs. Antibody-Based Biologics

The fundamental architectural difference between TCR-based and antibody-based therapies explains their divergent safety profiles. Antibodies, whether targeting cytokines like TNF-α or surface receptors like CD20, affect every cell expressing that target. Humira blocks TNF-α produced by all macrophages, dendritic cells, T cells, and other immune cells throughout the body - including TNF-α that’s doing important work protecting against tuberculosis or other intracellular pathogens. This explains why tuberculosis screening is mandatory before starting anti-TNF therapy and why the TB reactivation rate, while reduced with screening, remains a persistent concern.

In contrast, a properly designed TCR therapeutic for rheumatoid arthritis could, in principle, target only the T cell clones recognizing disease-relevant autoantigens (such as citrullinated proteins in RA or nephritogenic antigens in lupus) while sparing T cells specific for tuberculosis, influenza, or any of the thousands of other antigens the immune system monitors. Studies of TCR repertoires in autoimmune diseases consistently show oligoclonal expansion of particular TCR sequences, suggesting that a limited set of T cell clones drives pathology. Targeting these clones specifically, rather than their downstream cytokine products, offers the possibility of “surgical” immunomodulation.

The counterargument, of course, is that we don’t always know which T cell clones are pathogenic, or that multiple clones may be involved. This is valid and increasingly surmountable as mentioned above. Single-cell sequencing technologies now allow comprehensive mapping of T cell repertoires in diseased tissues. Machine learning approaches, such as that from Synteny, can predict TCR-peptide-MHC interactions with increasing accuracy. The technical barriers to identifying disease-relevant TCRs are falling rapidly.

Lessons from Oncology, Applied to Autoimmunity

The approval of afamitresgene autoleucel (Tecelra) in August 2024 as the first TCR-engineered T cell therapy for cancer provides valuable insights. This product engineers patient T cells to express a TCR recognizing MAGE-A4 peptides presented by HLA-A*02:01, achieving a 39% overall response rate in heavily pretreated synovial sarcoma patients. Notably, while manufacturing takes time, and bridging therapy is sometimes needed, the treatment itself can produce durable responses in responders, with two complete responses lasting the entire 3-year study period.

The TCR-T cell therapy experience in oncology has revealed both opportunities and challenges directly relevant to autoimmune applications. On the positive side, TCR-T cells can traffic to diverse tissue sites, persist long-term, and mediate sustained immune effects - all desirable properties for chronic autoimmune disease. On the challenge side, off-target toxicity from TCR cross-reactivity remains a concern. Fatal cardiotoxicity occurred in early trials when TCRs targeting MART-1 or MAGE-A3 cross-reacted with antigens expressed at low levels in heart tissue. This emphasizes the critical importance of comprehensive safety testing and target validation before clinical development.

For autoimmune diseases, the calculus is different. Rather than killing target cells (as in oncology), the goal is often to delete or suppress pathogenic T cell clones, pathogenic B cells or to redirect Tregs to sites of inflammation.

Looking Forward: What This Means for Patients

For patients with autoimmune diseases, particularly younger patients facing decades of treatment, the promise of TCR-based therapies is transformative. Imagine a rheumatoid arthritis patient who receives a one-time or periodic infusion of TCR T cell engagers targeting pathogenic T cells against citrullinated peptides, achieving years of drug-free remission while maintaining normal immune function. Or a lupus nephritis patient treated with a TCR T cell engager that eliminates pathogenic autoreactive T cell clones, preserving kidney function without the continuous immunosuppression and infection risk of current maintenance regimens.

These scenarios are not science fiction - they are logical extensions of what has already been accomplished in oncology and what analysis of TCR repertoires in autoimmune patients shows is possible. For example, it has been known that HLA-B*27 is a risk allele for ankylosing spondylitis (AS). TCR repertoire profiling revealed clonal expansion of TRBV9+ T cells. In a Phase 2 clinical study ELEFTA, targeted depletion of TRBV9+ T cells with an anti-TRBV9+ monoclonal antibody produced therapeutic benefits (ASAS40) in roughly 40% of HLA-B*27+ AS-diagnosed individuals at 24 weeks in contrast to 24% in the Placebo group. A more selective depletion of pathogenic T cells may result in greater efficacy. Similar approaches can be applied to AS and other autoimmune diseases depleting pathogenic T cells which have been clonally expanded and have common targetable features (e.g. unique CDR3 sequences).

The biggest barrier for establishing TCR therapies includes the complexity of identifying disease-relevant TCR targets across heterogeneous patient populations. The HLA restriction can be addressed with therapies designed to target multiple HLA, something that Synteny has shown is possible. Thus these barriers are not insurmountable. Companies like Immunocore are already developing TCR bispecific platforms for multiple indications. Companies like Adaptive, Gentibio and Repertoire Immune Medicines are mapping autoreactive T cell repertoires in major autoimmune diseases like RA and IBD. Regulatory agencies have shown willingness to engage constructively on innovative approaches, as evidenced by the approval pathways for CAR-T products in autoimmune indications like lupus.

The Ethical Imperative

There’s an ethical dimension to this discussion that is worthy of further discussion. When we prescribe current broadly immunosuppressive therapies to otherwise healthy young adults, we’re making a societal decision that the convenience and proven track record of existing drugs outweighs the burden of increased infection risk, malignancy risk, and need for lifelong treatment. As TCR-based approaches mature, continuing to default to broadly immunosuppressive regimens for new patients may become harder to justify.

A 25-year-old woman with lupus nephritis faces perhaps 60 years of treatment. Under current standard-of-care, that might mean 60 years of infection risk from mycophenolate and corticosteroids, with cumulative risks of osteoporosis, cataracts, diabetes, and other sequelae of chronic immunosuppression. If TCR-based therapies can offer an alternative path - perhaps periodic treatment allowing prolonged remission with intact immunity between treatments - then the standard-of-care calculation changes fundamentally.

This is particularly acute for pediatric autoimmune diseases. Children with juvenile idiopathic arthritis, childhood-onset lupus or Type I Diabetes may face 70 or 80 years of treatment. The cumulative toxicity burden of current approaches over such timespans is substantial. For these patients especially, the precision and potentially intermittent nature of TCR-based therapies could dramatically improve quality of life and long-term outcomes.

Conclusion: Precision Medicine Meets Immunology

The evolution from methotrexate to anti-TNF biologics to IL-23 inhibitors represents progressive refinement in our approach to autoimmune disease. Each generation of therapeutics has improved targeting and reduced certain toxicities. But all remain fundamentally limited by their lack of true cellular specificity: they affect all cells expressing their targets, not just those driving disease.

TCR-based therapeutics represent a qualitative leap beyond this paradigm. By harnessing the exquisite specificity evolution has built into the adaptive immune system, they offer the possibility of targeting only disease-relevant immune cells while preserving protective immunity. For healthy individuals whose only medical problem is an inappropriately activated immune response to self-antigens, this precision matters enormously. It’s the difference between a treatment they can envision taking for decades and one whose long-term risks may rival the disease itself.

The field of autoimmune therapeutics stands at an inflection point. Synteny and others are powering the technologies needed to realize TCR-based precision immunotherapy. What remains is the collective will - from biopharmaceutical companies, academic researchers, funding agencies, regulators, and patient advocacy groups - to prioritize this development path. For the millions of patients living with autoimmune diseases, particularly young patients facing lifelong treatment, that prioritization cannot come soon enough.

References

1. Dens C, Bittremieux W, Affaticati F, Laukens K, and Meysman P. Interpretable deep learning to uncover the molecular binding patterns determining TCR-epitope interaction predictions. ImmunoInformatics. 2023; Vol 11, Sept 100027.

2. Croce G, Bobisse S, Moreno DL, Schmidt J, Guillame P, Harari A and Gfeller D. Nature Communications. 2024; 15: 3211.

3. Thrift WJ, Lounsbury NW, Broadwell W, Heidersbach A, Freund, E, Abdolazaimi Y, Phung QT, Chen J, Capietto A-H, Tong A-J, Rose CM, Blanchette C, Lill JR, Haley B, Delamarre L, Bourgon R, Liu K, Jhunjhunwala S Towards designing improved cancer immunotherapy targets with a peptide-MHC-I presentation model, HLApollo. Nature Communications 2024; 15: 10752.

4. Pai JA and Satpathy AT. High-throughput and single-cell T cell receptor sequencing technologies. 2021 Nature Methods, 19:881-892.

5. Marrer-Berger E et al The physiological interactome of TCR-like antibody therapeutics in human tissues.2024 Nature Communications. 15: 3271.

6. Nathan P, Hassel JC, Rutkowski P, et al. Overall survival benefit with tebentafusp in metastatic uveal melanoma. N Engl J Med. 2021;385(13):1196-1206. doi:10.1056/NEJMoa2103485

7. Hassel et al. (2023). “Three-year overall survival with tebentafusp in metastatic uveal melanoma.” N Engl J Med, 389:2256–2266.

8. Immunocore Ltd. KIMMTRAK (tebentafusp-tebn) Prescribing Information. 2024.

9. Hua G, Carlson D, Starr JR. Tebentafusp-tebn: A novel bispecific T-cell engager for metastatic uveal melanoma. J Adv Pract Oncol. 2022;13(7):717-723.

10. Burmester GR, Panaccione R, Gordon KB, et al. Adalimumab: long-term safety in 23 458 patients from global clinical trials in rheumatoid arthritis, juvenile idiopathic arthritis, ankylosing spondylitis, psoriatic arthritis, psoriasis and Crohn’s disease. Ann Rheum Dis. 2013;72(4):517-524. doi:10.1136/annrheumdis-2011-201244

11. Weinblatt ME, Keystone EC, Furst DE, et al. Adalimumab, a fully human anti-tumor necrosis factor α monoclonal antibody, for the treatment of rheumatoid arthritis in patients taking concomitant methotrexate. Arthritis Rheum. 2003;48(1):35-45. doi:10.1002/art.10697

12. Winthrop KL. Risk and prevention of tuberculosis and other serious opportunistic infections associated with the inhibition of tumor necrosis factor. Nat Clin Pract Rheumatol. 2006;2(11):602-610.

13. Facheris P, Jeffery L, Del Duca E, et al. T Helper 2 IL-4/IL-13 Dual Blockade with Dupilumab Is Linked to Some Emergent T Helper 17-Type Diseases, Including Seronegative Arthritis and Enthesitis/Enthesopathy, but Not to Humoral Autoimmune Diseases. J Invest Dermatol. 2022;142(9):2660-2670. doi:10.1016/j.jid.2022.03.018

14. Rosenberg AS, Puig L, Launay O. Dupilumab-associated inflammatory arthritis: a systematic review. J Eur Acad Dermatol Venereol. 2022;36(Suppl 2):36-43.

15. Anders HJ, Chan TM, Sanchez-Guerrero J, et al. Efficacy and safety of guselkumab in patients with active lupus nephritis: results from a phase 2, randomized, placebo-controlled study. Rheumatology (Oxford). 2024 Dec 2:keae647. doi:10.1093/rheumatology/keae647

16. D’Angelo SP, Araujo DM, Razak ARA, et al. Afamitresgene autoleucel for advanced synovial sarcoma and myxoid round cell liposarcoma (SPEARHEAD-1): an international, open-label, phase 2 trial. Lancet. 2024;403(10434):1460-1471.

17. FDA approves first cell therapy to treat patients with unresectable or metastatic synovial sarcoma. FDA News Release. August 2, 2024.

18. Nasonov et al 2025 Doklady Biochem and Biophys. Jun; 522 (1): 387-403.

19. Mougiakakos D, Kronke G, Volkl S, et al. CD19-targeted CAR T cells in refractory systemic lupus erythematosus. N Engl J Med. 2021;385(6):567-569.

20. Mackensen A, Müller F, Mougiakakos D, et al. Anti-CD19 CAR T cell therapy for refractory systemic lupus erythematosus. Nat Med. 2022;28(10):2124-2132.

21. Tsui C, Maldonado P, Montaner B, et al. Checkpoint inhibition and T cell engagers in cancer immunotherapy. Immunity. 2022;55(12):2212-2233.

ARLO is our deep learning model for estimating the affinity of TCR-pMHC interactions without collecting expensive SPR measurements of affinity. This is possible because ARLO is trained on data from MYRIAD[2], our experimental data foundry, in which the frequency of stable interactions is measured indirectly through sequencing read counts.

There are two important subtleties involved in handling such data for the purpose of model training. First, measurements from MYRIAD, or indeed any other enrichment-based biological assay, are actually pairs of sequencing counts, and second, measurements come from not one but multiple experiments.

MYRIAD begins with a pooled, high-diversity, library of TCRs that vary in their initial abundance due to library construction and sampling effects. Sequencing the library prior to any binding-based selection via flow cytometry yields pre-sort counts, which reflect the background abundance of each TCR. The enriched population upon pMHC binding is sequenced again to produce post-sort counts. A large post-count does not imply high affinity per se; the pre- and post-sort counts must be interpreted together in order to disentangle affinity signals from background abundance.

A simple approach would be to normalise post-sort counts by the pre-sort counts, and use this enrichment ratio as input for ARLO [3]. However, the raw enrichment ratio ignores sampling noise in the low-count regime. A more principled alternative is to place Beta priors on the underlying count probabilities [4]. This yields a posterior estimate of equivalent to the naive ratio but with constant pseudocount factors added to the numerator and denominator:

In practice, we find that this corrected enrichment ratio does not model the data with sufficient accuracy. This motivates an approach that involves learning a direct calibration, rather than imposing explicit assumptions on the behaviour of the pre- and post-sort counts. The idea is to learn a calibration function that directly maps from the two counts to a binding affinity score.

To develop some intuition for how we might proceed, consider a subset of TCR pairs where one has a larger post-sort count and the other a larger pre-sort count. This relationship defines a partial ordering: for any pair (i, j) such that

interaction i is said to dominate j. We train an affinity model that maps from TCR sequences u to affinity estimates,

using a pairwise ranking loss;

where P is the set of ordered pairs. With this baseline established, we freeze the initial approximation and return to the full dataset to learn a calibration function

which is trained to map from from raw count pairs to a scalar affinity score using the regression objective

Finally, we freeze the calibration function and fine-tune the entire network on the complete dataset with the objective

This process can be made to work in practice, but involves training to convergence on three separate occasions (the initial approximation, the calibration, and the fine-tuning). This staged process quickly becomes brittle and difficult to scale with the magnitude of MYRIAD data generation. Naively, we could try to sum the three objectives and train all three stages at once. However, the attentive reader will notice a problem: this creates a circular dependency where the calibration function and the affinity model are simultaneously trying to ‘chase’ each other, and the joint objective implicitly enforces the mutual regressions

which admits trivial solutions.

To break this symmetry, we decouple the source of the targets from the model being trained. This idea is closely related to the class of student-teacher self-supervised learning methods that stabilise training by separating an online network from a slowly evolving target network, most notably BYOL [5]. In those settings, an exponential moving average (EMA) of the model parameters is used to generate training targets that change more slowly than the model itself. Here, we adopt the same principle in a different domain: rather than learning representations from augmented views, we use a target network to provide a stable reference for calibrating noisy, experiment-dependent count data.

We maintain two versions of the affinity model and calibration map: an online version that updates at every step, and a target model, defined by the EMA of the online weights [6]. The target parameters of are updated as

This creates a slow-moving objective that prevents the calibration function from collapsing into trivial solutions. The calibration function is trained to regress from raw counts to the target model’s affinity scores and vice versa:

When combined with the ranking objective, this architecture enables the model to self-calibrate, extracting a clean affinity signal in a single, end-to-end pass.

One advantage of learning the calibration function, rather than explicitly modelling the relationship between the pre- and post-sort counts, is that we can analyse the behaviour of the calibration. Indeed, the calibration function is plotted below for three different experiments, where red and blue indicate a high and low estimate of affinity, respectively.

Notice that the calibration functions vary significantly between experiments, demonstrating the drawback of imposing explicit modelling assumptions on the counts. Moreover, the empirically-derived calibration functions are complex and non-linear, explaining why a simple enrichment ratio is not sufficient to capture the relationship.

[1]

[2]

[3] Rubin, Alan F., et al. “A statistical framework for analyzing deep mutational scanning data.” Genome biology 18.1 (2017): 150.

[4] León-Novelo, Luis G., et al. “Semiparametric Bayesian inference for phage display data.” Biometrics 69.1 (2013): 174-183.

[5] Grill, Jean-Bastien, et al. “Bootstrap your own latent-a new approach to self-supervised learning.” Advances in neural information processing systems 33 (2020).

[6] Tarvainen, Antti, and Harri Valpola. “Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.” Advances in neural information processing systems 30 (2017).

The ability to engineer TCRs with precision hinges on the quality, diversity and accessibility of data; public data is scarce, biased and suffers from many biological and technical confounders. To achieve this, Synteny has developed a next-generation assay platform, MYRIAD, targeting the TCR-pMHC in diverse ways, generating hundreds of millions of data points. For us, developing a Data Platform to allow AI scientists and biologists to interact with a secure single access layer to interrogate these millions of data points in each cycle is at the foundation of everything we do. It allows us to integrate diverse datasets, track experimental results, set agents loose to explore large arrays of policies, and ensure that every model we train is grounded in reliable, comprehensive, well-labelled information. Without this robust infrastructure, insights become fragmented, experiments can not be reliably reproduced, and the potential of AI-driven design is severely limited.

We have designed our platform to be intentionally interpretation-agnostic. Traditionally, experimental pipelines in biology are built around a single interpretation of the data. Once an experiment is run, one canonical processing path produces the dataset that all downstream analyses and models consume. This creates hidden constraints on the data. When an assay is bound to a single “official” analysis, every downstream model inherits the assumptions, thresholds, and biases baked into that first interpretation. At Synteny, we have deliberately decoupled experimental execution from downstream data interpretation, treating all interpretations as first-class citizens.

For example, when evaluating the results of our sequencing pipeline, we can infer sequences from reads using BlastN (with many varying parameters), anchored alignment, Aho-Corasick, and so on. With this design, all interpretations of our sequencing pipeline are exposed to our agentic system - allowing it to make the decision of which interpretation to use for the given task. Moreover, we have enforced programmatic consistency through a typed, schema validation model, so that our agentic system can write new methods for data interpretation.

A powerful agentic layer relies on a data infrastructure designed for autonomy, not just storage. Agentic AI does more than consume datasets: it makes decisions, selects interpretations, forms hypotheses, and iterates over time. To support this, the data platform must capture not only experimental outputs, but also process: how data was interpreted, which assumptions were made, what alternatives were considered, and how those choices influenced downstream actions. This requires full lineage across data, models, and decisions, ordered in time, queryable, and reusable, so agents can reason over past behaviour, detect blind spots, and improve future strategies. Without this, agents operate myopically, repeating work or overfitting to narrow views of the data. With it, the platform becomes a learning substrate: enabling robust generalization, faster discovery of patterns (including rare or unexpected ones), and reliable autonomy across applications, from TCR design to protocol optimization and target discovery.

To provide as much data as possible to our system, Synteny has complemented our proprietary data with a substantial amount of disparate public data sources. These sources often have a wealth of data but in an inaccessible form: they have missing metadata, lack standardised identifiers and are curated to varying standards without use of ontologies. We harmonise and enrich these datasets into a common schema based on AIRR (Adaptive Immune Receptor Repertoire) standards. By bringing public data sources together in a unified representation with our proprietary data, we enhance the data available for target selection, model training and Agentic AI.

As AI-driven design continues to accelerate drug discovery, the organisations that will lead the field are those that can pair sophisticated models with equally sophisticated data infrastructure. At Synteny, our focus on building a unified, standards-driven data platform is not just a technical choice - it’s a strategic one. By harmonising public datasets, abstracting complex experimental outputs, and ensuring programmatic consistency across the entire data lifecycle, we’re creating the conditions under which Agentic AI can truly thrive.

In short, the future of TCR engineering will be defined by the quality of the data ecosystems that support them. By deeply investing in these foundations now, we’re positioning AI-designed TCR therapeutics to move from possibility to reality.

Whether it is Sakana’s AI Scientist, or Google’s research agents, they all tend to follow the same script: LLM agents dressed in digital lab coats, faithfully re-enacting human academic hierarchies. You have a “Literature Reviewer” which passes a PDF to a “Hypothesiser,” which prompts a “Coder,” which hands off to a “Writer.”

These are genuinely impressive engineering artefacts. They can execute end-to-end workflows that used to take humans weeks. But we must ask the uncomfortable question: Are we engineering the intelligence of discovery, or just the bureaucracy of science?

At Synteny, where we work on engineering therapeutics, we are finding that this role-playing architecture is not just suboptimal, it may be fundamentally misaligned with the ultimate goal of scientific machine learning. As a final abstraction for discovery, mimicking the org chart of a university lab seems like a local optimum.

We suspect the future is not a committee of chatbots, but a single, continually learning, connectionist fabric that experiences the entire lab loop end-to-end and gradually learns how to explore. This post is not a product announcement, nor is it a solved blueprint. It is an attempt to sharpen the questions we should be asking if we want AI to genuinely transform scientific discovery, especially in regimes like T cell receptor (TCR) engineering, where every experiment hurts.

The Emergence of the Standard Model

Over the past twenty-four months, the field of automated scientific discovery has largely converged on a unified architectural pattern: the Multi-Agent Research Framework. A survey of the current landscape reveals distinct tiers of this architecture, all sharing a common “agentic” DNA.

End-to-End Automation: The most prominent example is Sakana AI’s AI Scientist. Their v1 and v2 systems [1, 2] propose a fully autonomous pipeline that generates novel research ideas, writes the necessary code, executes experiments on a simulator, and compiles the results into a LaTeX-formatted paper. This approach has already yielded tangible results, including the system’s first peer-reviewed acceptance [8], effectively treating scientific discovery as a verifiable code-generation task.

Human-in-the-Loop Co-Pilots: Google’s AI Co-Scientist [3, 9] adopts a collaborative approach. It utilizes specialized agents for “Generation,” “Reflection,” and “Ranking” to assist researchers in biomedical hypothesis generation. Rather than replacing the scientist, it employs Elo-based tournament systems to filter candidate ideas before human review, aiming to accelerate breakthroughs in complex wet-lab domains.

Lab-Autonomous Agents: Systems like FutureHouse’s Robin and Kosmos [4, 10] focus on bridging the gap between reasoning and execution. These agents function as “lab partners” capable of planning experiment series and directing robotic cloud laboratories to verify hypotheses physically, moving beyond pure simulation.

General Frameworks: We also see the rise of “meta-frameworks” like CodeScientist [5] and broader surveys of AI4Research [6], which abstract these patterns into reusable libraries. These tools allow developers to instantiate custom research agents across various domains, further cementing the “agentic science” paradigm [7].

Despite different targets, these systems share a structural isomorphism: they cast science as a directed graph of personas, a “Literature Searcher,” “Planner,” and “Coder/Runner.” This mirrors the broader trend of verticalized agents in domains like software and law, where LLMs are wrapped in job-title prompts. The approach seems effective for exploitation in information-rich settings, but the most striking successes still cluster in simulation-heavy domains [2] or tasks with easily verifiable code-based ground truth. In short, it digitizes human bureaucracy, good for organizing known work, less clearly suited to high-entropy physical-world discovery.

The Anthropomorphic Dead End

From a machine learning perspective, there is something strange about this approach. These AI Scientist systems are engineering artefacts, not learned artefacts. The “scientific method” they implement is hard-coded as a sequence of roles and stage gates. The orchestration is prompt routing and if-else logic, not a policy learned from experience.

Contrast this with the major breakthroughs in modern ML. The field advanced by suppressing the instinct to hand-craft pipelines and instead building connectionist models trained end-to-end, letting gradients flow through the whole system. We learned to trust data and optimization more than our intuitions about how cognition “should” be structured [12]. AlphaGo, AlphaFold, diffusion models, and LLMs themselves are all victories of this pattern.

So it’s worth asking a natural question:

Why are we suddenly comfortable hard-coding the architecture of “how science should be done” into a cast of digital roles?

Rebuilding Bureaucracy in Silicon

There is a sociological risk here: we may end up recreating academic bureaucracy in silicon. We are building agents that imitate peer reviewers, enforce rigid stage gates (“you must have a hypothesis before you explore”), and reward conservative, citation-heavy reasoning as “rigor.” If our goal is to discover new ways of doing science, faithfully copying our existing institutional structure into software is an odd place to stop. We risk bottlenecking discovery by baking in our current incentives and blind spots, not to mention the risk that humans will simply set the wrong goals for these agents to chase.

Brute-Force Novelty

It is also unclear how much genuine novelty these systems produce. In a controlled head-to-head ideation study in NLP, LLM-generated ideas were judged more novel than human experts (p < 0.05) but slightly weaker on feasibility, with additional concerns about limited diversity and unreliable self-evaluation [11].

Most current pushes toward “originality” rely on relatively simple knobs: higher sampling temperatures, diversity penalties, or larger candidate pools, rather than an explicit, learned objective for discovery itself. We talk as if agentic wiring will magically yield better science, but in machine learning, real power usually appears only when we can define a learning signal and optimize for it.

The Hard Truth of the Physical World

Discovering the real world is significantly harder than solving digital benchmarks.

In domains like TCR engineering, the knowledge we need does not already exist in digital form, it must be created through slow, expensive, and noisy experiments. We cannot brute-force millions of trials or rely on cheap simulators the way digital lab-coat systems implicitly assume.

In our loop, the real bottleneck is learning how to acquire information efficiently under hard biological constraints. While this may sound like classic territory for active learning and experimental design, the “AI scientist” opportunity is to connect those principles to an end-to-end, multi-modal system that can leverage diverse sources of knowledge, tools, and resources. Such a system requires knowing what we don’t know, choosing experiments that actually reduce uncertainty (rather than just confirming biases) [14], and improving that skill over time. In the end, we need better discovery systems, not just the assumption that more scripts, more roles, or faster orchestration will magically produce truth.

From Extrinsic Tasks to Intrinsic Learning

What would an AI Scientist look like if we built it with the connectionist, end-to-end instinct that drove the rest of deep learning?

We believe it would look less like a committee and more like a learning organism, a single system that leverages all resources (humans, computers, wet labs) to “learn how to learn” the real world.

While this is speculative, here are a few key departures from the digital lab coat picture:

1. Emergent, Not Hard-Coded Roles: If “roles” exist, they should be emergent properties of the network, not hard-coded prompts. We shouldn’t tell the system “this part is the reviewer”; we should care about the behavior of the whole system and let internal specialization emerge if it is useful for the objective.

2. Exploration as the Primary Objective: The goal is not just “maximize reward on a single candidate quickly” (exploitation), but “maximize reduction in predictive uncertainty about biology that matters.” Curiosity and uncertainty become first-class rewards [13]. Intrinsic signals, prediction error, ensemble disagreement, or information gain must drive what the system chooses to test when extrinsic rewards (e.g., successful clinical outcomes) are sparse or years away.

3. Humans as Entropy Sources: In this setup, humans are not bosses handing out Jira tickets. Humans are entropy sources. We inject hunches, hints, constraints, and “this feels important” priors. The system incorporates those signals into its state, but the long-term experiment-selection strategy is something it is trained to improve.

Meta-Learning the Logic of Discovery

Ultimately, this points toward meta-learning. We want to learn the policy of exploration itself.

This is distinct from standard exploitation (finding the best molecule in a known distribution). We need a policy that knows how to use and build tools for discovery, how to leverage the right information, and how to adapt when moved to a new domain. Ideally, we are working toward a community effort to learn this exploration policy across domains, a model that “knows how to discover” regardless of whether it is looking at proteins or small molecules.

Exploitation, actually delivering good candidates, does not disappear in this view. Instead, it emerges as a behavior, shaped by the deeper drive to improve the system’s world model.

We must acknowledge that this is a massive research challenge. Defining the reward function for “good exploration” in the real world is incredibly difficult, especially when we lack large-scale datasets of “good discovery trajectories.” Most of our best human strategies are scattered across lab notebooks and intuition, not stored in machine-usable formats. But the difficulty of the problem shouldn’t deter us from recognizing that it is the problem.

Conclusion

Today’s AI Scientist systems are effectively cosplaying human scientists: they wear digital lab coats, follow our job descriptions, and automate our paperwork. This is fine, even great, for many efficiency tasks.

But if we stop there, we are missing the point of machine learning. We are getting distracted by the illusion of intelligence that comes from large-scale memorization and role-play. At Synteny, our bet is that the real frontier is not perfectly mimicking a human lab in software, but learning exploration policies over the actual lab loop.

That future system won’t wear a lab coat. It will, increasingly, be the lab.

References

[1] Lu, Chris, et al. “The ai scientist: Towards fully automated open-ended scientific discovery.” arXiv preprint arXiv:2408.06292 (2024).

[2] Yamada, Yutaro, et al. “The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search.” arXiv preprint arXiv:2504.08066 (2025).

[3] Gottweis, Juraj, et al. “Towards an AI co-scientist.” arXiv preprint arXiv:2502.18864 (2025).

[4] Mitchener, Ludovico, et al. “Kosmos: An AI Scientist for Autonomous Discovery.” arXiv preprint arXiv:2511.02824 (2025).

[5] Jansen, Peter, et al. “Codescientist: End-to-end semi-automated scientific discovery with code-based experimentation.” Findings of the Association for Computational Linguistics: ACL 2025. 2025.

[6] Chen, Qiguang, et al. “AI4Research: A Survey of Artificial Intelligence for Scientific Research.” arXiv preprint arXiv:2507.01903 (2025).

[7] Wei, Jiaqi, et al. “From ai for science to agentic science: A survey on autonomous scientific discovery.” arXiv preprint arXiv:2508.14111 (2025).

[8] The AI Scientist Generates its First Peer-Reviewed Scientific Publication https://sakana.ai/ai-scientist-first-publication/

[9] Accelerating scientific breakthroughs with an AI co-scientist https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

[10] Demonstrating end-to-end scientific discovery with Robin: a multi-agent system https://www.futurehouse.org/research-announcements/demonstrating-end-to-end-scientific-discovery-with-robin-a-multi-agent-system

[11] Si, Chenglei, Diyi Yang, and Tatsunori Hashimoto. “Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers.” arXiv preprint arXiv:2409.04109 (2024).

[12] Sutton, Richard. “The bitter lesson.” Incomplete Ideas (blog) 13.1 (2019): 38.

[13] Pathak, Deepak, et al. “Curiosity-driven exploration by self-supervised prediction.” International conference on machine learning. PMLR, 2017.

[14] Yang, Jason, et al. “Active learning-assisted directed evolution.” Nature Communications 16.1 (2025): 714.

If you’ve been watching the explosion of immune therapies over the past decade, you might think we’ve already mapped the immune system’s potential. CAR-Ts, checkpoint inhibitors, antibody drugs have all changed medicine. However, we have only been fishing in the shallow end: the real ocean of opportunity lies inside the cell, and TCRs are the key to unlocking it. Unlike antibodies that can only “see” the 10–20% of proteins sitting on the cell surface, TCRs can target the other 80–90%, the intracellular world where most disease biology actually happens. This isn’t just a new class of drugs; it’s an entirely new universe of targets across oncology, virology, and autoimmunity that have been completely off-limits until now.

The beauty of TCRs is their versatility. Cancers, autoimmune conditions and infectious diseases aren’t simple but rather messy, heterogeneous, polygenic diseases. TCRs let us fight complexity with precision. Instead of one blunt therapy that only works for a fraction of patients, imagine precise cocktails of TCRs tuned to a patient’s HLA type and mutational landscape. Suddenly, “precision medicine” isn’t a marketing slogan, but a scalable reality. This flexibility makes TCRs the ultimate adaptive platform for tackling the most challenging diseases we face.

Cell therapies like TCR-Ts have already proven that this biology works, but they’re expensive to manufacture and tough to deliver safely at scale. Enter TCR T cell engagers (TCR TCE): off-the-shelf molecules that harness the body’s own T cells to do the killing. Without the need for cell factories, multi-week manufacturing cycles, we can design safe, potent, exquisitely targeted molecules that direct your native immune system to clear disease. The next wave of immune therapeutics won’t be built on the backs of cell factories, it will be built on TCRs.

This panacea is no longer theoretical: the first two clinical proof points are already here. KIMMTRAK (Immunocore’s gp100 TCR bispecific) became the first approved TCR therapy in 2022, delivering unprecedented survival benefits in metastatic uveal melanoma. TECELRA (Adaptimmune’s autologous cell therapy against MAGE-A4) was approved in 2024 for treatment of advanced synovial sarcoma, an important weapon in the fight against solid tumours. Unlike many antibody modalities, which require broader in vivo assessment of Fc functions and complex pharmacokinetics, TCR programs often advance with a more streamlined in vivo package, supporting faster and more cost-efficient progression into early clinical studies.

The field is no longer a “what if”, it’s a “now”: the science is proven, the biology validated, and the commercial white space enormous. The TCR revolution has begun, and we at Synteny are looking to define its future.

TCRs target intracellular antigens: over 70% of the proteome that is largely unaddressable by other modalities

What makes TCRs such a commercially attractive platform is that they can - in principle - target any intracellular protein. The targetable landscape is therefore much larger than it is for small molecules and antibodies, which are for the most part limited to surface-expressed proteins only.

TCRs survey the intracellular proteome by exploiting what is effectively the waste-disposal system in cells. When proteins are degraded, they’re first broken down into different-length fragments, and eventually those fragments are broken down into amino acids. But along the way, some of those fragments are shuttled into the endoplasmic reticulum (ER), a compartment inside the cell that contains HLA class I proteins (see figure below). These HLA proteins bind to peptides and take them to the cell surface for TCRs to observe.

One of the fascinating aspects of T cell responses is that they are adaptive, as opposed to being innate, or germline-encoded. The receptors of T cells are generated randomly in the thymus, with each T cell containing one receptor. T cells are then discarded if they fail one of two checks: first, they have to be able to bind to the HLA complexes of the host which are presented by specialised cells within the thymus. This requirement is then caveated by a second requirement, that they do not bind to these thymic HLA complexes too strongly. As the thymus presents lots of HLA complexes that contain potentially millions of self peptides - peptides which derive from the host’s proteome - this ensures that any T cells that recognise a self peptide are removed. After these checks, you’re left with an army of potent immune cells that are able to bind to non-self (e.g. pathogen-derived) peptides, but shouldn’t be autoreactive.

The final piece of the puzzle, and the reason this system protects us so effectively, is the extraordinary diversity of HLA proteins. Different HLA alleles present very different sets of peptides. As a result, if a pathogen mutates its proteins to avoid binding your HLA molecules, those same mutations are unlikely to help it evade mine. At the population level, this leaves the pathogen with very little evolutionary advantage. Thus, adaptive immunity is generated somatically, giving each individual a uniquely diverse repertoire of immune receptors that are tuned to that person’s specific HLA alleles.

While HLA diversity provides population-level protection against rapidly evolving pathogens, it also introduces a major challenge for developing TCR-based therapies. Because disease states are presented differently to the immune system depending on a person’s HLA type, therapies that rely on engineered TCRs can only target a limited set of peptide–HLA complexes. The biggest commercial opportunities for TCR-based therapies are found when a disease-associated peptide is presented on a HLA allele that is commonly presented amongst a population.. The most common HLA allele in the US and EU markets is HLA-A*02, which is expressed in around 40% of individuals. Notably, both KIMMTRAK and TECELRA target the HLA-A*02-expressing subpopulation. However, HLA-A*02 is not the only allele worth considering.

At Synteny, we are focused on programming TCRs to recognise multiple HLA alleles that present the same peptides; providing an opportunity to expand the number of patients accessible with a single therapy. One example of this is combining HLA-A*11:01 and HLA-A*03:01, which have a significant overlap in the peptides they present. Owing to differences between these two HLAs, it is not usually the case that a TCR that binds to one will bind to the other. In fact, it is quite the opposite. Therefore, it becomes a protein design problem to engineer TCR sequences that are able to tolerate the differences between these HLAs, but maintain sufficient affinity and specificity to the presented peptides to be safe and effective therapies. By unifying multiple HLAs, we are able to address a larger fraction of patients.

Turning TCRs into potent therapeutics

Biologically-derived therapeutics (‘biologics’) have revolutionised drug treatment in the 40 years since their first approval, with biologics now accounting for over a third of all newly approved drugs. The drug development process for biologics offers both advantages and challenges compared to traditional small molecules. For example biologics retain many of the natural properties of the molecules they derive from, and so can remain biologically active within the body for a long time. However, they are also large and typically complex molecules, presenting challenges in how they are made, stored, and delivered into the body.

The majority of approved biologics are based on monoclonal antibodies. This technology exploits antibodies, soluble molecules produced by B cells, whose role is to identify anything ‘non-self’ and generate an immune response towards it. These naturally occurring molecules are one of the most abundant proteins in sera, providing a protective barrier against reinfection by rapidly neutralising pathogens before they have had time to establish an infection. Monoclonal antibody technology relies on identifying a B cell clone which produces an antibody that binds to a target, typically by immunising mice with a target protein of interest. Individual B cells from immunised mice are then fused with an immortalised cell derived from a myeloma, creating an immortalised cell line, known as a hybridoma, which acts as a protein production facility making large amounts of the target specific antibody.

As antibody-based biologics have grown in number, the platforms which support both identification and production have evolved in efficiency and robustness. There are now a variety of ways to identify antibody sequences that bind to a target of interest. In addition to traditional immunisation strategies using mice, a variety of in vitro and in vivo systems have been developed. For example, camelid antibodies, which have a single chain responsible for target binding instead of the 2 chains used in humans and mice, have been adapted to generate large synthetic libraries suitable for high throughput screening, increasing the speed and efficiency of target screening. Likewise the traditional process of hybridoma generation has evolved in response to technological advances in DNA sequencing and gene synthesis. Antibodies can now be rapidly sequenced, synthesised, and cloned into specialised mammalian cell systems highly optimised for producing consistent drug products at high yield.

In addition to producing high quality antibodies more quickly, modern antibody technology platforms are able to generate modular antibody-like structures, where specific antibody domains have been switched to impart multi-specificity or enhanced function. For example, a major growth area in antibody-based therapeutics is the bispecific protein, where two antibody chains with different target specificities are brought together into one therapeutic protein. A major group of antibody-based bispecifics are the T cell engagers (TCEs), where antibody-based target binding is coupled with bystander T cell engagement through binding to CD3.

Antibodies have dominated the growth in biologics due in part to their natural properties as soluble molecules. As TCRs are membrane proteins, we need to adapt their structure for use as soluble therapeutics. Conversion of TCRs into soluble proteins typically involves first removing the areas of protein which interact with the membrane, and introducing amino acids that help to stabilise the interactions between TCR chains. TCRs also do not have a direct capacity to signal to the immune system, and so solubilised TCRs require the addition of a further structure to generate a biologic effect, known as an effector function. Currently TCR-based therapeutics in the clinic function as T cell engagers (TCR-TCEs), which use a similar antibody-based effector domain to that of other TCEs, engaging bystander T cells through binding CD3.

The more than 200 antibody-based TCEs currently in clinical trials across the world offer an exciting opportunity to exploit the technological enhancements in this area in the development of TCR-based TCEs. TCRs and antibodies are structurally similar proteins, with constant and variable domains that can be interchanged with each other. This kind of modular approach has been applied to several TCR-TCEs within the clinic, that are able to be produced at scale, resolving many of the early challenges in TCR-TCE product yields, and also enable the imparting of more antibody-like properties on the molecules, such as longer biological half lifes.

TCR repertoires: a fossil record of disease

Designing individual TCRs shows the power of precise antigen recognition, but these receptors sit within a vastly larger and more complex ecosystem. Understanding that ecosystem requires looking at entire TCR repertoires. A TCR repertoire is essentially a census of the T cells present in a biological sample. In practice, this is usually measured by multiplex PCR amplification of rearranged TCR genes, most commonly the β chain, followed by high-throughput sequencing. The approach is inexpensive, scalable, and does not require any knowledge of antigen specificity: it simply returns the set of TCR sequences present in the blood or tissue at the time of sampling. Classical repertoire sequencing captures single chains, but new methods are beginning to link α–β pairs in bulk (for example, TIRTL-seq), allowing more complete characterisation of the receptors present in a sample without the cost or complexity of single-cell sequencing.

A typical repertoire survey generates 10,000–1,000,000 unique TCR sequences per person. Because TCR generation is a deeply stochastic process, most sequences are “private”: they are unique to an individual or present in very few people. Consequently, every new person sampled adds tens of thousands of previously unseen data points. This has created a rare situation in immunology where large datasets, tens of thousands of individuals, can be built cheaply, with each sample providing new and informative sequence diversity.

Why does this matter?

Even in their unlabelled form (i.e. without knowing what antigen they recognise), repertoires provide an extraordinarily rich corpus for learning the underlying “grammar” of TCRs. From an AI perspective, this is an ideal training domain: billions of real biological sequences shaped by evolution and tolerance, but without task-specific labels that might bias or limit the space of solutions. Models trained on these data can learn the latent structure of TCRs, motifs, constraints, allowed rearrangements, biophysical preferences, and use this foundation to predict antigen binding or generate new TCRs with desired properties.

Large case–control studies add another powerful layer. By comparing the repertoires of people with and without a disease, it is possible to identify specific TCR sequences associated with that condition, an approach exemplified by the landmark Emerson et al. (2017) study in Nature Genetics. Many autoimmune diseases have now been shown to harbour “public” pathogenic TCRs that are enriched across patients. The key limitation, until recently, has been that we typically do not know what these disease-associated receptors bind.

This is where a working TCR–antigen model and high-throughput mapping technologies become essential. If we can de-orphan pathogenic TCRs, linking sequence to antigen using in silico prediction or in vitro platforms such as large-scale signalling pMHC libraries, we can finally identify the causal antigens driving autoimmune pathology. For diseases like ankylosing spondylitis, type 1 diabetes, and others with strong immunogenetic signatures, this offers a route to discovering previously unrecognised therapeutic targets and, ultimately, new ways to intervene upstream of tissue damage.

Towards disease diagnosis from repertoires

Looking ahead, the ability to infer antigen specificity directly from sequence unlocks a much broader opportunity: using repertoire sequencing as a diagnostic test. Today, repertoire surveys are cheap, fast, and widely deployable, but limited by the fact that almost all TCRs are “orphans.” As models improve, we will be able to annotate increasing fractions of a repertoire with predicted antigen targets and immunological context. This creates a path to diagnosing hard-to-detect diseases, especially autoimmune conditions, where the pathogenic clones are often abundant in blood, potentially even to early cancer detection, where tumour-associated T cells circulate long before imaging or symptoms emerge.

In this future, a simple, low-cost repertoire assay could become a powerful screening tool: a snapshot of the adaptive immune system that reveals which diseases the immune system is currently responding to, or about to respond to. With the combination of large population-scale repertoires, accurate TCR–antigen models, and scalable wet-lab mapping, this is rapidly moving from aspiration to reality.

Learning the laws of TCR-antigen recognition: a problem perfectly suited to AI

The versatility that makes TCRs so attractive as a diagnostic tool and a therapeutic modality also makes them hard to engineer using lab-based or traditional computational methods. A functional repertoire spans ~10¹⁵ possibilities, and current discovery methods such as screening donor blood, panning for binders, or affinity-maturing individual clones, only ever sample an infinitesimal corner of that space. Unlike small molecules, there is no TCR counterpart to fragment-based heuristics: amino acid motifs in TCRs are almost completely uninformative when removed from their global structural and sequence context.

Combinatorial explosion used to be the standard argument against applying AI to biological design. That view has not survived the past decade. AlphaGo, image generation, and ultimately with language models operating across endless combinations of tokens have shown that scale and complexity are no longer blockers. In protein engineering, AI approaches have breached the domain of antibody design and TCRs sit squarely in the crosshairs of this space where generative AI and AI-led discovery excel. For example, we at Synteny have shown that, given appropriate data, TCR affinity programming is directly accessible as a statistical inference problem.

But TCR therapies demand much more than discovering single target binding: balancing self-tolerance with potency, engineering cross-HLA recognition, and tuning multi-antigen binding while avoiding off-targets. These are essentially multi-objective challenges and are not screenable at scale. They are, however, exactly the kinds of conditional optimisation problems that modern generative and probabilistic models are designed to solve.

Indeed one could even make the argument that every question in TCR drug discovery and diagnostics can be expressed as a conditional probability over biological sequences and relevant property-encoding random variables. What does this receptor bind? How do I change it while preserving binding to this antigen? What is the likelihood of cross-reactivity? These map naturally onto modern probabilistic and generative AI frameworks.

Taken together, frontier biology and frontier AI render TCR design tractable and open an entirely new window into programmable biology.

Effective TCR bispecific therapies depend on high-affinity interactions between a T cell receptor (TCR) and a target peptide–major histocompatibility complex (pMHC) [1]. A fundamental challenge in protein engineering is to generate a TCR sequence with a desired binding affinity (pKd) to a target [2]. Quantifying the affinity of an interaction is therefore crucial.

The gold standard of affinity measurement is surface plasmon resonance (SPR) [3]. This is slow and very expensive to perform (~£700 per interaction), which is impractical for large-scale TCR bispecific therapy discovery. In this post we introduce ARLO, a deep learning model which is able to estimate affinity without collecting SPR measurements. ARLO enables estimation of binding affinities at a fraction of the cost and time.

ARLO is powered by data from MYRIAD, our next-generation experimental assay that captures millions of TCR–pMHC interactions in their native mammalian context. In addition to in vitro binding readouts multiplexed across multiple pMHC targets, MYRIAD generates evidence of functional activation bridging the gap between high-throughput screening and real immune response. We defer further details of MYRIAD to the end of this post, but the germane points are

1. It is ~10⁵ times cheaper to generate data per interaction than SPR (<£0.01).

2. It produces sequencing counts data for each interaction.

3. The counts data are highly noisy.

Each individual datapoint generated by MYRIAD is too noisy to be directly useful for affinity quantification. One solution might be to measure each observation in replicate, but this is expensive and time-consuming. Instead, we adopt a deep learning approach, leveraging the ability of modern models to handle noisy data just like this. Our deep learning model, ARLO, is fine-tuned directly on the output of MYRIAD.

Figure 1 (left) shows that counts, in isolation, convey little about affinity. And yet, after we fine-tune ARLO on the counts data, a close correlation between ARLO scores and SPR affinity measurements materialises (right).

Figure 1. (left) Before fine-tuning ARLO, there is no significant correlation between MYRIAD counts and affinity. Counts per TCR are normalised. (right) After fine-tuning ARLO, scores are predictive of affinity (r = 0.78). Values of pKd are measured using SPR in both plots. All predictions and measurements involve the same pMHC target. (NB: the right plot has more points as it contains some TCRs that were not screened in MYRIAD, including some higher affinity TCRs).

How can ARLO be used to optimise TCRs for affinity?

We would like to use ARLO to guide perturbations to TCR sequences that steer the affinity to a desired level. A well-known limitation of protein foundation models is their lack of sensitivity to single amino acid mutations [4]. This is critical for making rational perturbations, so it is particularly important that ARLO is able to learn this effect. Each point in Figure 2 represents a pair of TCRs that differ by a single mutation. The predictions and measurements are identical to those in Figure 1, except that we are considering only single amino acid mutations. We see that the change in ARLO is able to estimate the changes in pKd.

Figure 2. ARLO is robust to the effects of single amino acid mutations on affinity (r = 0.80).

Having established that ARLO predicts the effects of even minimal changes, we took a candidate TCR and introduced a handful of mutations to increase the ARLO score. Through ARLO-guided sequence perturbations (up to 14 amino acid edits from the candidate sequence), we were able to increase the true affinity of the interaction by a factor of over 10⁴, moving the TCR sequence further away from the candidate TCR with each alteration. Figure 3 shows the results of this process.

Figure 3. ARLO is able to increase the binding affinity of a TCR to a target pMHC through an iterative process.

How much data is required to fine-tune ARLO around a target?

In order to fine-tune ARLO around a given target, we typically screen several million TCRs in MYRIAD. Figure 4 demonstrates that the predictive power of ARLO scales with the number of TCRs screened. We note a log-linear relationship between the scale of data and correlation. Of course, this trend cannot continue indefinitely, but we see that as we continue to scale up our data, we are yet to hit a ceiling in performance.

Figure 4. The predictive power of ARLO is closely correlated to the number of TCRs screened with MYRIAD (r = 0.96).

How is ARLO trained?

MYRIAD involves delivering high diversity TCR libraries into a model cell line, where they can be expressed on the cell surface in their native environment. Each cell expresses on average just one single receptor. pMHC-like reagents couple our targets of interest to a fluorescent dye and bind to TCRs on the cell surface. This enables us to specifically isolate the cells carrying TCRs that bind to those targets and identify their sequences using deep sequencing approaches, which generate sequencing read counts for each interaction. Those counts are used as input for ARLO.

In order to train ARLO, we made a key assumption that for any pair of counts within the same sample, a higher affinity interaction is likely to receive more counts. Based on this, we trained ARLO with a ranking loss objective [5] between pairs of TCRs within the same sample. For each pair of TCRs, if ARLO assigned a higher score to the TCR with a higher count, then there was no contribution to the loss. If it assigned a lower score, the loss was proportional to the score difference. ARLO does not have to draw arbitrary cutoffs to determine binding and non-binding interactions.

ARLO also benefited from large-scale pretraining on other sources of data, before successive rounds of fine-tuning on this objective on counts data for a specific target. The ARLO architecture consists of a modern transformer-based stack [6], similar to frontier protein language models [7].

What does ARLO mean for the future of TCR discovery?

ARLO demonstrates how deep learning can bridge the gap between noisy, high-throughput experimental data and precise biophysical quantities like affinity. By removing the need for costly assays, it enables scalable and rational optimisation of TCR sequences directly from inexpensive experimental screens. We expect ARLO to continue to play a crucial role in our internal drug-discovery programmes.

References

[1] Stone, Jennifer D., and David M. Kranz. “Role of T cell receptor affinity in the efficacy and specificity of adoptive T cell therapies.” Frontiers in immunology 4 (2013): 244.

[2] Riley, Timothy P., and Brian M. Baker. “The intersection of affinity and specificity in the development and optimization of T cell receptor based therapeutics.” Seminars in cell & developmental biology. Vol. 84. Academic Press, 2018.

[3] Schuck, Peter. “Use of surface plasmon resonance to probe the equilibrium and dynamic aspects of interactions between biological macromolecules.” Annual review of biophysics and biomolecular structure 26.1 (1997): 541-566.

[4] Pak, Marina A., et al. “Using AlphaFold to predict the impact of single mutations on protein stability and function.” Plos one 18.3 (2023): e0282689.

[5] Chen, Wei, et al. “Ranking measures and loss functions in learning to rank.” Advances in Neural Information Processing Systems 22 (2009).

[6] Vaswani, Ashish, et al. “Attention is all you need.” Advances in neural information processing systems 30 (2017).

[7] Hayes, Thomas, et al. “Simulating 500 million years of evolution with a language model.” Science 387.6736 (2025): 850-858.

The challenge of mapping genotype to phenotype defines much of modern biology. From the protein folding problem, where sequence determines structure, to emerging virtual cell models, these efforts reveal how sequence encodes function across molecular and cellular scales. Now, advances in synthetic biology and machine learning are allowing us to explore these maps in ways that seemed impossible even a decade ago.

At Synteny, we believe that by pointing these new technologies toward the problem of molecular recognition — how immune receptors recognise their targets — we can open up new classes of programmable therapeutics across oncology, autoimmunity and beyond.

From reading biology to writing it

Over the past twenty years, biology has undergone a revolution in scale. The arrival of next-generation sequencing gave us the ability to read DNA cheaply at astonishing depth and accuracy. More recently, advances in synthesis technologies have made it possible to write bespoke DNA sequences at comparable scale. Using these technologies it is possible to generate millions of precisely designed sequence variants, express and test them in living systems and measure their effects across a range of rich phenotypes. These advances have fuelled a new generation of experiments that bridge sequence and function—from deep mutational scanning for the design of novel enzymes, to Perturb-seq, which employs systematic genetic perturbations to unravel the regulatory logic of human cells. Collectively, these studies underscore a profound concept: when perturbations and measurements can be conducted systematically at scale, biology becomes predictable.

A new frontier: molecular recognition

At Synteny, our focus is to bring this same experimental and computational scale to one of biology’s most complex and consequential phenomena: T-cell receptor (TCR) recognition of peptide–MHC (pMHC) targets. This is the core of adaptive immunity — the molecular handshake that distinguishes “self” from “non-self,” infection from tolerance, cancer from health. Despite decades of study, this recognition landscape remains largely unmapped. We can measure individual interactions, but we cannot yet predict which TCR will bind which antigen, with what affinity, or what cellular outcome that interaction will drive. We believe that solving this molecular recognition problem will unlock a new era of programmable biologics and immune therapeutics.

The Synteny Data Foundry

To tackle this challenge, we’ve built the Synteny Data Foundry: a platform for generating large-scale, high-fidelity data on TCR–pMHC interactions. Like a foundry that forges raw materials for engineering, our data foundry produces the fundamental data from which AI-guided models of molecular recognition can be built. It rests on four key pillars:

1. TCR perturbations at scale

At the core of our platform lies the ability to design, synthesise and test millions of TCR variants. Using our generative design engine, MANIFOLD, we propose vast libraries of candidate TCRs predicted to engage a given target. These designs are synthesised at scale using modern DNA-writing technologies, enabling us to systematically explore sequence space across diverse complementarity-determining regions (CDRs) and frameworks. Each cycle begins imperfectly, with diverse hypotheses about what might bind, and iteratively improves as data flow back into our models.

2. High-throughput TCR screening for discovery, development, and safety

Molecular recognition cannot be fully understood outside its native context.

Rather than relying solely on reductionist in-vitro binding assays, we embrace the complexity of the human immune cell. We’ve developed two innovative high-throughput cell-based screening platforms to characterise up to a million TCRs against a panel of peptides or up to a million of pMHC against a panel of TCRs in a single assay. But why go via the cell-based route, and what makes this setup a game-changer? Let’s break it down.

2.1 Why a cell-based system?

Understanding TCR–pMHC recognition requires studying receptors in their native environment. Our high-throughput cell-based assays express TCRs on living T cells, preserving membrane architecture, co-receptor engagement, and downstream signalling cascades. This enables us to quantify antigen-driven activation under near-physiological conditions, capturing key determinants such as avidity, clustering, and signalling thresholds that are invisible to in-vitro binding assays. It’s like testing a car on the road instead of just in a wind tunnel; far more indicative of performance under real conditions.

2.2 Screening millions of TCRs for discovery

Size matters. Our discovery engine can assay up to one million TCR variants against panels of peptide–MHC targets in a single experiment, allowing systematic exploration of vast areas of sequence space. This scale reveals rare, high-affinity, and multi-HLA-reactive receptors that traditional low-throughput assays cannot detect. Each screen yields quantitative mappings between sequence and function, enabling us to refine our generative design models iteratively. The result is accelerated identification of optimised TCRs with the desired balance of potency, specificity, and manufacturability.

2.3 Screening millions of antigens for safety

Building on our cell-based TCR screening platform, we’ve pioneered a complementary system that takes screening off-target antigens to the next level. Libraries comprising millions of peptide–MHC complexes are expressed in engineered reporter cells that signal upon productive engagement, allowing comprehensive profiling of TCR specificity across the human peptidome. This approach identifies rare but clinically relevant cross-reactivities that may underlie toxicity or autoimmunity. By integrating these data into our design pipeline, we can eliminate problematic motifs and design safety into our molecules from the outset, producing TCRs that are both safe and potent.

3. Ultra-cheap paired-chain sequencing

A major barrier to studying endogenous TCRs at scale has been sequencing cost and complexity. Pairing endogenous TCRs requires expensive single-cell approaches. Our synthetic constructs used by our foundry mean that paired-chains can be read off in a single amplicon that reads both TCR chains together in a bulk sequencing approach, allowing us to track how α–β combinations influence cell surface pairing and function. This dramatically reduces cost per observation and ensures that sequence–function relationships remain intact. It also allows us to detect subtle patterns: the specific motifs, hydrogen-bond networks, or charge complementarities that make or break recognition.

4. Lab in the loop

Data alone are not enough, learning requires iteration. Our foundry operates as a closed experimental–computational loop. Each round begins with a generative design, proceeds through synthesis, expression, and phenotypic screening, including cross-reactivity testing, and then feeds back into our AI models. With every cycle, the models become sharper: learning the grammar of molecular recognition from empirical evidence. This loop enables us to not only predict binding but to control it with conditional generative design.

A programmable immune system

Molecular recognition defines what our immune system can and cannot see. By learning its rules we can begin to rewrite them, designing TCRs that target cancers, autoimmune antigens, or infectious diseases with unprecedented precision. Synteny’s data foundry is our engine for this future. By combining high-throughput synthetic biology, rich functional phenotyping, and AI-driven design, we are turning molecular recognition from a mystery into an engineering discipline.

Experimentally, measuring binding affinity – usually using surface plasmon resonance (SPR) or biolayer interferometry (BLI) – is precise but resource-intensive, often dominating the cost and timelines of lead optimisation. The prospect of accurately predicting affinity without needing to physically measure it is a clear motive when every increment in predictive accuracy promises shorter cycles, fewer wet-lab iterations, and ultimately faster drug programmes.

Predicting binding affinity: a combination of physics, data and machine learning

Previous approaches for computational affinity prediction span a spectrum from physics-derived first-principles to data-driven learning. In recent years, relative free energy perturbations (FEP) and absolute binding free energy (ABFE) calculations have seen renewed traction [1] and while these approaches can be highly informative, their accuracy is limited and they remain computationally intensive. Machine learning promises a complementary path where a direct mapping between input molecules and measured affinities is learnt [2]. However, the accuracy of ML methods has been limited and while the literature is full of bold claims of state-of-the-art performance, these have rarely translated to improved drug discovery efficiency. Of course, hybrid strategies that try to merge physical priors with ML are an obvious route to hack performance, but here we stop to ponder why purely data-driven affinity prediction – without any physics in the loop – has often struggled in practice?

Many of the complications lie not in complex architectures but in the available data landscape. Careful benchmarking exercises have shown that public affinity datasets exhibit substantial data leakage between training and test sets. When this leakage is controlled for, almost all performance on unseen examples is abolished. This lack of generalisability has motivated the development of applicability domains and other pessimistic control mechanisms and yet even with careful splitting and debiasing, shortcut learning persists in which models latch onto superficial cues rather than underlying interaction physics [3].

Representation in binding affinity models: structure versus sequence

While data quality issues apply to all data-driven affinity predictors, we must plow onward and consider a second differentiating axis for machine learning methods: representation. How should models “see” binding partners – by sequence alone, or through their 3D structure?

Sequence-based models have surged ahead in recent years, buoyed by the success of general language models such as ChatGPT. A sequence-only representation has practical appeal: it applies universally (structures are often unavailable), and it avoids conflating affinity with a single static conformation that may not capture the dynamic, entropic contributions to binding. However, researchers that embrace this route should ask the underlying question: are we betting that the physics of binding can be inferred directly from sequence? Or are we content with learning sequence-based distances in “affinity space” that capture useful correlations, without necessarily encoding the underlying forces?

Structure-based approaches, by contrast, seek to model physical interactions more directly. Researchers initially experimented with increasingly raw atomic representations of 3D coordinates: from grids and pharmacophore fingerprints to CNNs, graph networks, or point cloud–based architectures [2]. The hope was that the physical laws of binding would emerge from the data itself, arguably the most successful example of this is AlphaFold itself. However, the structural data only represents low energy states making it hard to learn a physically realistic world model, and the data is complex and the data space is continuous with many sources of bias and error such as incomplete modelling, artefacts of crystallisation, and favouritism towards certain druggable protein families. Embracing the issues presented by structural data is far from a free lunch when compared with representing the sequences alone.

While weighing the merits of sequence versus structure has been studied extensively, we argue that the extent to which each representation is susceptible to underlying data biases is an under-appreciated reason for differences in performance.

Early experiments: embracing structure by adapting Boltz-2 for protein-protein affinity prediction

We recently adapted Boltz-2 [4] – a state-of-the-art structure-based model for protein–ligand affinity – to the protein–protein setting. The adapted model, Boltz-2-PPI, was benchmarked against sequence-based alternatives on two datasets:

• TCR3d [5], structural data of TCRs in complex with their peptide-MHC targets curated from the Protein Data Bank (PDB). Each TCR-pMHC complex has been annotated with experimental binding affinity (251 complexes).

• PPB-affinity (filtered) [6, 7], a larger set of ~8,500 protein–protein affinity measurements representing diverse protein classes, many without structural data.

We trained for affinity prediction on both datasets (Tables 1 & 2) and the outcome was consistent: sequence-based models outperformed structure-based Boltz-2-PPI.

Table 1. Predictive performance on a sequence-dissimilar test set of TCR3d.

Table 2. Predictive performance on the PPB-affinity (filtered) test set.

These models are not strong predictors, partly because we use challenging splits with proteins unseen during training. Random splitting – still common in our field – artificially inflates performance; for example, on random splits of TCR3d, a model trained only on TCR and pMHC names (without sequences) reaches a Pearson r = 0.6 on the test set. The encouraging takeaway is that honest, challenging validation prevents overestimation and provides a reliable measure of model performance.

A key weakness of this early model is the limited data we have provided for training. TCR3d includes only 150 training structures – too few for such a complex task – and adapting Boltz-2-PPI with a smaller affinity head yields similar performance, indicating that data scarcity, not model size, is the bottleneck. The larger PPB-affinity dataset performs better, suggesting that robust affinity prediction will require aggregated, multi-fidelity data from diverse sources.

To probe the reasons for the performance gap between sequence and structure methods, we tried forgoing structure-prediction and training directly on the experimentally resolved structures in TCR3d, Boltz-2-PPI still lagged behind our sequence-based method (implemented using ESM2-650M). This suggests that structure prediction accuracy is not the bottleneck (yet) and the learned structural embeddings face more fundamental issues which render them currently unsuitable for affinity prediction.

Interestingly, combining structure and sequence embeddings offered modest gains, particularly for weaker sequence models. Concatenating Boltz-2-PPI affinity module embeddings with baseline ESM2 or ProtT5 embeddings trained on the PPB-affinity (filtered) nudged performance upward, indicating that the structural representation contains complementary signal, albeit weaker than those captured by sequences.

Lessons and outlook

• Sequence leads today. With massive pretraining, sequence models capture signatures of affinity robustly, even without explicitly modelling binding interactions.

• Structure-based embeddings need to be improved to smoothly express both the sequence- and structural- variations that determine the free energy of binding – especially under dataset shift.

• While our early experiences have been based on Boltz-2, we expect that the arguments made here apply to many available learned representations of protein structure.

• A fusing of both physics and machine learning, and also structure- and sequence is currently crucial for realising state-of-the-art predictive performance for affinity modelling.

At Synteny, we believe large-scale data generation is the key to decoding molecular interactions with ML. Sequence-only corpora were transformative because they were big and standardised and we believe affinity needs the same treatment. Alongside data generation we expect ML methods to mature to suit some of the challenges introduced by continuous, euclidean space data distributions.

Our early internal results mirror the external picture: sequence-first baselines are hard to beat, but hybrid models begin to close the gap when trained on richer, standardised datasets. We’ll share more as these datasets and fusion strategies mature.

References

[1] Siebenmorgen, Till, and Martin Zacharias. “Computational prediction of protein–protein binding affinities.” Wiley Interdisciplinary Reviews: Computational Molecular Science 10.3 (2020): e1448.

[2] Isert, Clemens, Kenneth Atz, and Gisbert Schneider. “Structure-based drug design with geometric deep learning.” Current Opinion in Structural Biology 79 (2023): 102548.

[3] Scantlebury, Jack, et al. “A small step toward generalizability: training a machine learning scoring function for structure-based virtual screening.” Journal of Chemical Information and Modeling 63.10 (2023): 2960-2974.

[4] Passaro, Saro, et al. “Boltz-2: Towards accurate and efficient binding affinity prediction.” BioRxiv (2025): 2025-06.

[5] Lin, Valerie, et al. “TCR3d 2.0: expanding the T cell receptor structure database with new structures, tools and interactions.” Nucleic Acids Research 53.D1 (2025): D604-D608.

[6] Liu, Huaqing, et al. “PPB-Affinity: Protein-Protein Binding Affinity dataset for AI-based protein drug discovery.” Scientific data 11.1 (2024): 1316.

[7] Alsamkary, Hazem, et al. “Beyond Simple Concatenation: Fairly Assessing PLM Architectures for Multi-Chain Protein-Protein Interactions Prediction.” arXiv preprint arXiv:2505.20036 (2025).

Drug discovery can be viewed as a constrained statistical sampling problem. Like composing a sentence, designing molecules and protein sequences demands precise control over both semantics and syntax. The semantics include properties such as target binding affinity, solubility, and chain stability, while the syntax refers to the identities and arrangement of atoms or amino acid residues.

In late 2016, amid rising excitement around deep learning in vision and speech, a preprint by Gómez-Bombarelli et al [1] quietly made a breakthrough that marked a turning point for drug discovery. The authors introduced a variational autoencoder (VAE) capable of generating molecular structures – not pixelated reconstructions of faces, but chemically valid graphs of atoms and bonds. This was more than a technical novelty. It opened the door to continuous latent spaces that could be sampled and decoded into tangible drug candidates.

While the VAE (and also concurrent developments with recurrent neural networks [2]) marked the genesis of deep generative chemistry, a more recent major inflection point is score-based generative modelling. First applied in drug discovery contexts in a brace of papers just two years ago [3, 4], these diffusion models offered a rather different approach to generation by learning to reverse a modeller-scheduled forward stochastic process that gradually adds noise to data, eventually learning to reconstruct structured objects like molecules or proteins from pure noise. Their enduring popularity (over, say, language modelling approaches) stems from the ability to steer the generation towards many different constraints.

Why syntactic constraints matter

While de novo generation in drug design may be the approach du jour, many practical applications benefit more from completing or modifying existing sequences in reliable, property-preserving ways. Grafting, or scaffolding, motifs, redesigning CDR loops in antibodies and T cell receptors (TCR), or mutating just a single residue at a late design stage — these are all special cases of inpainting, a familiar task in the image processing domain. One of the earliest approaches to inpainting, which is often called replacement sampling, involves evaluating the diffusion gradient function at each denoising timestep with the values of the fixed data dimensions replaced with the corresponding noised target values of the forward process. We illustrate this process in Figure 1 below.

The pitfalls of naive sampling

Despite its simplicity, replacement sampling suffers from the foundational flaw that the sampled distributions are not actually the reverse of any valid forward diffusion process. This is not merely a technicality. For instance, the replacement sampling distribution introduces an irreducible approximation error [5, 6] that persists even with larger models and ever greater sampling steps.

For many applications, particularly in vision, this limitation is sometimes tolerated. Methods like RePaint [7] have built upon replacement sampling with iterative noise-cut-paste schemes, delivering striking improvements in visual fidelity. But drug discovery demands more than perceptual realism.

To illustrate the problem, consider the following toy experiment.

Say we have the task of sampling points on a map of the world according to human population density. Here we mock up such a toy population map by sampling from an imbalanced mixture of three 2D Gaussians to produce a synthetic dataset as shown in Figure 2 below.

We then train a (tiny) diffusion model to recover this joint distribution. If we were to then limit our sampling to the longitude by fixing the latitude to the lower value, this is essentially the task of inpainting the x-coordinate around a fixed y-coordinate (y=−1). Under correct conditional sampling, the sampled data should lie predominantly within the bottom left area where the probability mass is concentrated. However with replacement sampling, the model is driven toward the overall data mode, as we see above. It fills in data from the bottom right region, largely ignoring the conditional, because that is where the density is highest under the marginal. No increase in the number of diffusion sampling steps can correct this flaw.

This simple failure mode serves up a caution for drug design. When the conditional diverges significantly from the marginal, as it often does in real-world biology, inexact sampling tends to snap back to dataset priors. In TCR sequence design, for instance, publicly available datasets are heavily skewed toward viral antigens. Without proper treatment, generated sequences meant to bind cancer neoantigens, say, will drift toward the viral-antigen associated sequences that dominate the training data. The consequences go deeper than isolated residue mis-sampling as global effects come into play, such as altered epitope specificity, or poly-specificity [8], thereby exposing one to latent off-target binding risks.

Moving beyond heuristics to exact conditional sampling

These pitfalls have been met with a large number of innovative solutions in recent years. Here we highlight one class of methods that eschew approximations — such as the pairing of ancestral sampling with heuristic substitutions, or variational inference methods — for asymptotically exact sampling [5, 9, 10]. This is arguably an attractive strategy in domains that value fidelity over inference speed.

Because each intermediate distribution gently interpolates between the pure noise and the sharp conditional (posterior), we should be able to guide the sampling process using well-established Monte Carlo sampling techniques defined for sequences of annealed distributions that end on the posterior. These approaches are now at the heart of many modern protein generative models [11].

Training-free generalisations

But the powerful consequences of this framework is not restricted to its correctness but also its generality and efficiency.

As we show if one defines the intermediate distributions as products of the form

where the original label points to the prior reverse process distributions whose gradients are learnt during (unconstrained) diffusion model training and the second term a conditional forcing term, then provided they anneal to the posterior at the end of the diffusion process, there is no restriction on the form of the conditional distribution itself. This freedom unlocks a far richer class of other applications. For instance, it can be a delta function fixing a motif, like in inpainting, a soft constraint like sequence similarity, individual residue scores for biophysical heuristics such as solubility, residue-pair scores for protein stability, or indeed a theoretically infinite variety of conditions through combinations of distributions, all from a single trained model without the need for individual task-specific conditional training.

Example: Scaffolding Without Positional Fixing

Structure scaffolding is a common design pattern in protein engineering where one fixes the atomic coordinates of a desired functional motif (e.g. a set of helices) or an interacting region (e.g. an antibody loop or enzymatic site), and seeks to scaffold a full protein backbone around it. Traditional methods require the user to predefine the number of residues between fixed motifs, introducing brittleness and human bias. Prior work such as Floating Anchor Diffusion Models [12] introduced motif flexibility into the generative process but requires specialised training procedures tailored to that objective. By contrast, our method introduces no additional retraining burden [6]. Under the training-free guidance framework, we define the conditional forcing term as a mixture distribution across a range of inter-motif lengths. The model learns to scaffold both motif placement and the surrounding structure during sampling.

We demonstrate this using RFdiffusion [3] as the base model, taking care to exclude task-specific components. Unlike replacement sampling, which requires the user to specify the number of intervening residues, our scaffolds not only achieve higher fold success rates, but also recover the correct motif spacing automatically in a single shot, reducing the risk of poor folding through misspecified positioning. We show this below for a simple scaffolding task around a discontinuous motif where the model correctly picks the correct spacing midway through the set of intermediate distributions, leading to high-confidence resulting structures.

References

[1] Gómez-Bombarelli, Rafael, et al. “Automatic chemical design using a data-driven continuous representation of molecules.” ACS central science 4.2 (2018): 268-276.

[2] Segler, Marwin HS, et al. “Generating focused molecule libraries for drug discovery with recurrent neural networks.” ACS central science 4.1 (2018): 120-131.

[3] Watson, Joseph L., et al. “De novo design of protein structure and function with RFdiffusion.” Nature 620.7976 (2023): 1089-1100.

[4] Ingraham, John B., et al. “Illuminating protein space with a programmable generative model.” Nature 623.7989 (2023): 1070-1078.

[5] Trippe, Brian L., et al. “Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem.” arXiv preprint arXiv:2206.04119 (2022).

[6] Cornwall, Lewis, et al. “Training-Free Guidance with Applications to Protein Engineering.” NeurIPS 2024 Workshop on AI for New Drug Modalities.

[7] Lugmayr, Andreas, et al. “Repaint: Inpainting using denoising diffusion probabilistic models.” Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.

[8] Karthikeyan, Dhuvarakesh, et al. “Conditional generation of real antigen-specific T cell receptor sequences.” Nature Machine Intelligence (2025): 1-16.

[9] Cardoso, Gabriel, et al. “Monte Carlo guided diffusion for Bayesian linear inverse problems.” arXiv preprint arXiv:2308.07983 (2023).

[10] Wu, Luhuan, et al. “Practical and asymptotically exact conditional sampling in diffusion models.” Advances in Neural Information Processing Systems 36 (2023): 31372-31403.

[11] Wohlwend, Jeremy, et al. “Boltz-1 democratizing biomolecular interaction modeling.” BioRxiv (2025): 2024-11.

[12] Liu, Ke, et al. “Floating anchor diffusion model for multi-motif scaffolding.” arXiv preprint arXiv:2406.03141 (2024).