SEPIA is now untethered from the PDB: any antibody with a predicted structure can seed a new SEPIA experiment. This capability allows SEPIA to grow with every new antibody discovery or de novo design campaign and sample previously unexplored regions of Ab-Ag sequence and structure space. Here we highlight several new directions for SEPIA data generation and discuss what untethering SEPIA from the PDB means for de novo antibody design.

SEPIA at PDB Scale

The Protein Data Bank (PDB) contains ~10,000 experimentally resolved Ab-Ag complexes [1]. This captures only a small fraction of possible antibody, epitope, and paratope diversity, creating the structural data bottleneck we highlighted in our SEPIA research paper [2] and previous Substack post [3].

In the four months since releasing the SEPIA paper, we have rapidly expanded both the scale and diversity of SEPIA, and the pace of data generation continues to accelerate. The paper described three campaigns spanning ~26 million affinity measurements, 190 parental antibodies, 45,000 tested SEPs, and ~1,200 positive SEP-Ab pseudo-structures. Today, SEPIA includes 3.2X more SEPs tested against 2.4X more parental antibodies, yielding 5.3X more positive pseudo-structures and 3.1X more hard negatives (Figure 2). In total, SEPIA now contains ~145,000 experimentally measured SEP-Ab pseudo-structures, including ~6,200 positive pseudo-structures and ~139,000 high-confidence hard negatives along with ~45,000 antibody mutant measurements that quantify local affinity landscapes for validated pseudo-structures. Collectively, this represents 10X more pseudo-structures with affinity labels than there are antibody entries in the PDB.

Data volume alone is not enough; we also need data diversity. Structural clustering of all positive pseudo-structures using Foldseek [4] shows that diversity has increased alongside volume. The ~1,200 positive pseudo-structures (on-target hits) reported in the paper formed 617 fold-level clusters (TMScore ≥ 0.5), while today’s ~6,200 positive pseudo-structures produce 5,025 clusters—an eightfold increase.

Importantly, these positive pseudo-structures are almost entirely distinct from known structures in the PDB: only 2 of 6,168 SEPIA positive pseudo-structures have a close PDB match (Foldseek complex TMScore > 0.7). SEPIA is therefore exploring regions of Ab-Ag structural space that the PDB has never covered.

Three Ways We Have Extended SEPIA

Since releasing the paper, we have developed SEPIA along three complementary axes:

1. Expanding scale: increasing SEPIA volume and diversity by designing against both a targeted library of high-hit-rate VHHs and a much broader parental antibody pool.

2. Targeting antibodies without solved structures: designing SEPs to bind VHHs from public antibody sequence databases (primarily sourced from INDI [5]) and against de novo VHH hits drawn from internal de novo antibody design campaigns.

3. Extending antibody modalities: designing SEPs against single-chain Fvs (scFvs).

Campaign Snapshot

Table 1 compares our previous SEPIA paper dataset with the four new SEPIA cohorts by parental antibody count, SEPs designed, on-target hits, hard negatives, and mutant hits.

Table 1. Per-cohort breakdown of SEPIA data added since the paper.

The four new cohorts serve complementary purposes. The targeted scaling cohort increases data depth for a fixed panel of antibodies, creating a resource for understanding information saturation. The remaining cohorts expand SEPIA’s scope to VHHs lacking solved structures, de novo designed VHHs, and scFvs.

Expanding Scale

To increase SEPIA data depth, we returned to 25 VHHs known to yield high SEP hit rates and substantially increased design volume against them. This cohort includes ~40,000 tested SEPs, ~3,000 on-target binders, and ~36,000 hard negatives—nearly three times as many positive pseudo-structures as the dataset from the original paper.

Unlike the other new cohorts, this campaign was focused on depth rather than antibody diversity. Generating many additional positive and negative pseudo-structures against the same parental antibodies creates a depth-focused dataset that tests how much added coverage of one paratope improves machine learning models. It also helps determine whether future data collection should prioritize broader VHH and SEP coverage or deeper coverage of existing parental antibodies.

Targeting Antibodies Without Solved Structures

SEPs binding sequence-only VHHs

In the original SEPIA study, every parental VHH had a solved PDB structure. We have now extended SEPIA to ~180 VHHs lacking known structures by first predicting their structures and then designing SEPs against them, yielding hundreds of novel validated positive pseudo-structures. Unlocking this capability greatly increases our ability to scale SEPIA across a much broader range of antibodies. A resolved parental antibody structure is no longer required, which expands the addressable input from the ~15,000 antibody-containing PDB entries to the full universe of available antibody sequences.

SEPs binding de novo VHHs

Generating SEPs against de novo antibodies provides the strongest demonstration that SEPIA is now untethered from common constraints on biological data scaling. We designed SEPs against ~50 A-Alpha Bio-designed VHHs from a prior de novo antibody design campaign that, by definition, do not exist in nature and lack experimentally solved structures. We observed our highest SEPIA hit rate to date (8.4%), with nearly every parental VHH yielding at least one binder. These campaigns generate fully de novo Ab-Ag complexes in which neither the parental VHH nor the designed SEP exists in nature or the PDB. Every successful de novo antibody campaign can therefore seed additional SEPIA campaigns, producing a flywheel where de novo antibody design continuously expands SEPIA.

SEPs binding VHH mutants

We measured binding affinities for tens of thousands of antibody mutants against SEPs (Figure 2). For nearly every SEP-VHH pseudo-structure, at least one VHH point mutant improved SEP binding affinity, pairing pseudo-structural context with quantitative mutant binding data. The results show that SEPIA point-mutant measurements can supply binding-affinity labels for zero-shot optimization models that rank VHH variants for improved SEP binding. We will discuss this work further in future blog posts.

Extending Antibody Modalities

A natural test of SEPIA’s generality is whether it extends beyond single-domain VHHs to larger, more flexible antibody formats. Single-chain variable fragments (scFvs), which connect heavy- and light-chain variable domains with a flexible linker, are an obvious next step, and we have successfully designed SEPs against this antibody modality. Hit rates for scFvs are lower than for VHH panels. The highest-affinity binder in the scFv cohort is a 1.2 nM SEP design with a β-sheet-rich fold (Figure 3). As we iterate on SEPIA campaigns to target scFvs, we anticipate that approaches such as increased computational sampling and improved filters will both increase scFv hit rates and expand the scale and diversity of scFv-derived SEPIA data.

The Data Flywheel for Antibody Engineering

Data scale and diversity are critical for improving ML-driven antibody engineering models for tasks such as de novo antibody design and zero/few-shot optimization [6–8]. Using SEPIA, we continuously collect antibody-antigen (Ab-Ag) data at scale and test how data depth, antibody and antigen breadth, and sequence and structural diversity affect model performance. For training binding models, the growing SEPIA dataset provides paired positive pseudo-structures, hard negative examples, and local affinity landscapes, allowing models to learn both binding and nonbinding relationships across related sequences. We are currently leveraging these datasets to measure how training data composition affects model performance and will share more on these results soon.

To support these efforts, we have made SEPIA data generation faster and less constrained. In the four months since our paper, SEPIA has grown ~3X, positive pseudo-structures ~5X, and structural diversity ~8X. By showing that predicted antibody structures are sufficient to seed SEPIA campaigns, we have expanded the addressable input from antibodies with experimentally solved structures to essentially any antibody sequence, including de novo designs. New antibody discovery and design campaigns can therefore directly seed new SEPIA experiments ad infinitum.

We believe that the scale, diversity, and speed of data generation represented in SEPIA will enable us to create a data flywheel for antibody discovery, engineering, and optimization: improved antibody design models and discovery campaigns generate new antibodies, those antibodies seed new SEPIA campaigns, and the resulting positive, negative, and affinity landscape data are used to improve the next generation of models. As this cycle accelerates, SEPIA will continue to grow into regions of Ab-Ag sequence and structural space that are largely absent from existing databases. This expanding dataset can support AI models that evaluate antibody design hypotheses across Ab-Ag sequence and structural space.

If your team is working on similar problems and wants to collaborate on dataset generation, please reach out to contact@aalphabio.com.

To explore our expanding datasets, visit atlas.aalphabio.com

Acknowledgements

Alex Eng, Kerry McGowan, Davis Goodnight, Lucian DiPeso, Colleen Shikany, Emily Engelhart, Leah Homad, Miranda Lahman, Juliana Barrett, Shyam Gandhi, Mackenzie Goodwin, Kenny Herbst, Charles Lin, Mimi McMurray, Randolph Lopez, David Younger, Ryan Emerson, Mark White, and Drew Duglan contributed to this work and the preparation of this text.

References

1. Schneider C, Raybould MIJ, Deane CM. SAbDab in the age of biotherapeutics: updates including SAbDab-nano, the nanobody structure tracker. Nucleic Acids Res. 2022 Jan 7;50(D1):D1368–72. doi:10.1093/nar/gkab1050

2. Altieri N, Harman JL, Noble D, Murakowska N, Eng A, McGowan KL, et al. The Synthetic Epitope Atlas: High-Throughput Design and Validation of De Novo Antibody-Antigen Complexes [Internet]. Synthetic Biology; 2026 [cited 2026 Jul 31]. Available from: http://biorxiv.org/lookup/doi/10.64898/2026.04.17.719295 doi:10.64898/2026.04.17.719295

3. Harman J, Murakowska N, Noble D, Altieri N, Lange AW. The Synthetic Epitope Atlas: Scaling Structural Training Data for De Novo Antibody Design [Internet]. 2026 Apr 22. Available from: https://aalphabio.substack.com/p/the-synthetic-epitope-atlas-scaling

4. Van Kempen M, Kim SS, Tumescheit C, Mirdita M, Lee J, Gilchrist CLM, et al. Fast and accurate protein structure search with Foldseek. Nat Biotechnol. 2024 Feb;42(2):243–6. doi:10.1038/s41587-023-01773-0

5. Deszyński P, Młokosiewicz J, Volanakis A, Jaszczyszyn I, Castellana N, Bonissone S, et al. INDI—integrated nanobody database for immunoinformatics. Nucleic Acids Res. 2022 Jan 7;50(D1):D1273–81. doi:10.1093/nar/gkab1021

6. Hummer AM, Schneider C, Chinery L, Deane CM. Investigating the volume and diversity of data needed for generalizable antibody–antigen ΔΔG prediction. Nat Comput Sci. 2025 Jul 8;5(8):635–47. doi:10.1038/s43588-025-00823-8

7. How to build a scaling law for biological AI. Jura Bio, Inc. [Internet]. Available from: https://www.jurabio.com/blog/scalinglaw

Here, we discuss how AlphaSeq affinity data is uniquely suited to train the next generation of protein design models. Our analysis focuses on three key properties:

• Reproducibility: do repeated or related measurements preserve binding relationships?

• Comparability: do different platforms (e.g., AlphaSeq and biophysical methods) support the same practical decisions for protein engineering?

• Volume: does the data have the scale and diversity needed for training generalizable models?

What does it mean to measure affinity?

Protein affinity measures the interaction strength between two proteins. For protein engineers, affinity is used to answer critical questions that inform downstream decisions: Does an antibody bind its target and how strongly? Does a mutation improve or weaken binding? Does a de novo designed binder form the intended complex? Does a binder engage its target specifically, or does it interact broadly with other proteins?

Affinity is typically reported as a dissociation constant (K_D), which relates the equilibrium concentrations of the free binding partners to the concentration of the bound complex. Stronger binding corresponds to a lower K_D. Established biophysical methods for measuring protein affinity, such as surface plasmon resonance (SPR) and biolayer interferometry (BLI), analyze purified proteins under defined experimental conditions, one interaction at a time. In these assays, one binding partner is immobilized on a surface while the other is introduced in solution, and the instrument records changes in signal as binding and unbinding occur over time.

AlphaSeq measures affinity in a different way: as a library-on-library, cell-based assay. A typical AlphaSeq experiment measures a large matrix of protein-protein interactions using DNA sequencing as a high-throughput readout of interaction strength [1]. Briefly, thousands of different proteins are displayed on the surfaces of yeast haploid cells, with one yeast library expressed in MATa cells and another in MATα cells. When the two libraries are mixed, interacting proteins between two haploid cells cause cellular fusion to form diploid cells. DNA barcodes originating from each haploid cell are sequenced to identify the interacting proteins, and relative sequence counts across the population of diploid cells reveal the abundance, and therefore the strength, of each protein-protein interaction. To place AlphaSeq’s relative binding strength measurements on an absolute affinity scale, each experiment includes reference interactions with known BLI-measured K_D values. Across the reference interactions, the measured AlphaSeq signal is log-linear with affinity, providing an experiment-specific calibration curve for reporting the strength of novel interactions with apparent K_D values, or K_Dapp. Throughout the following sections, K_Dapp refers to AlphaSeq-derived affinity measurements and K_D refers to measurements by SPR or BLI.

Affinity measurements depend on experimental context. Measurements from biophysical methods, such as SPR and BLI, can vary with active protein concentration, immobilization conditions, analyte concentration, measurement times, buffer conditions, and fitting assumptions [2, 3]. AlphaSeq likewise has a distinct assay context defined by yeast surface expression, protein-presentation geometry, and culture conditions [1, 4, 5]. A reported affinity value therefore reflects both the underlying molecular interaction and the conditions under which it was measured. For model training, reproducibility within that context is critical because consistent measurements provide a stable signal from which models can learn [6].

Reproducibility: Do repeated measurements preserve the same binding relationships?

Calibration to biophysical protein affinity measurements places AlphaSeq measurements on a BLI-based affinity scale (K_Dapp). Reproducibility asks whether those measurements behave consistently within that scale: when the same interaction is measured again, does it return a similar affinity, and when the same set of interactions is measured in a related display format, does AlphaSeq preserve the same ranking? For model training, reproducibility is essential. Affinity labels need to preserve the topology of the underlying interaction landscape, so that quantitative differences correspond to differences in binding rather than uncontrolled variation from assay runs, library construction, display format, or data processing.

Two case studies illustrate AlphaSeq reproducibility, drawn from years of assay development and internal benchmarking:

The first example looks at reproducibility across separate experiments for a diverse set of antibody-antigen interactions. We selected 38 antibody-target interactions spanning different antibody parents and target antigens, then measured the same interaction set in a second independent AlphaSeq experiment. In the replicate experiment, the DNA libraries were independently constructed, the yeast libraries were newly transformed, and the full AlphaSeq workflow was repeated. The resulting affinities were highly correlated across experiments (Figure 1; Pearson r = 0.93), showing that AlphaSeq recovers similar interaction strengths across independently generated libraries and assay runs.

The second example looks at reproducibility of relative affinity when the same interactions are measured in different display contexts. We measured 19,307 antibody variants against one protein target in both single-chain Fv and Fab formats, then compared the resulting K_Dapp values (Figure 2). Despite the change in antibody presentation, the two formats were strongly correlated across nearly twenty thousand measurements (Pearson r = 0.84). The systematic offset and different dynamic range between formats are expected; moving from Fab to scFv changes presentation geometry, expression, and stability. What matters for reproducibility is that the ranking is largely preserved.

Together, these examples demonstrate reproducibility at two levels: across independent AlphaSeq experiments and across closely related display formats. In practice, this means AlphaSeq can compare strong binders, weak binders, non-binders, cross-reactive binders, specific binders, and closely related variants within a consistent assay context. Consistency matters for model training: when affinity labels are stitched together from many disparate contexts (e.g., public databases), models learn dataset-specific artifacts instead of meaningful binding relationships. By measuring millions of diverse interactions in an identical experimental context, AlphaSeq affinity labels can be compared, aggregated, and learned from at scale.

Comparability: Can we distinguish strong binders from weak binders from non-binders?

Comparability asks whether AlphaSeq K_Dapp values preserve binding relationships measured across orthogonal biophysical methods. We look for decision-level agreement in two primary settings: (1) do the methods agree in their classification of binders and non-binders, and (2) do the methods agree in their ranking of stronger and weaker binders against a given target, including the correct identification of strengthening or weakening mutations? We demonstrate comparability in both settings with representative datasets.

Classification

To explore the classification question, we analyzed an AlphaSeq experiment designed to test whether literature-reported interactions with BLI or SPR affinities below 1 uM could be recovered in an AlphaSeq assay. The validation set included 99 reported protein interactions spanning receptor–ligand, antibody–ligand, and antibody–receptor interactions. Because AlphaSeq operates in a library-on-library format, the same experiment also measured 2,764 protein interactions without reported affinity measurements, which were used to calculate a false-positive rate. Each protein interaction was represented by multiple construct designs, truncations, or display orientations, yielding 619 measurements of reported interactions and 17,547 measurements of unreported interactions. For the classification analysis below, we reduced these construct-level measurements to one value per unique protein-protein interaction by using the strongest above-background AlphaSeq affinity measurement across the relevant constructs. Interactions with no measurements above assay background were treated as non-binders.

At a 1 uM affinity threshold, AlphaSeq recovered 39 of the 99 reported protein interactions, corresponding to 39% recall (Figure 3). This recovery rate is consistent with the constraints of a display-based method: some proteins do not express, fold, or present functionally on the yeast surface, and some antibodies lose binding activity when reformatted from full-length IgG into display-compatible fragments such as scFvs. The key practical point is that once a target construct has been shown to display functionally in AlphaSeq, recovery is much stronger for novel binders evaluated against that same format. To date, we have validated thousands of literature-reported protein interactions across hundreds of extracellular and intracellular proteins, and those functionally validated constructs provide the basis for our protein binding data generation efforts. When a high-value target is difficult to display, we use in silico stabilization approaches, such as structure-guided redesign, then verify functional display in AlphaSeq by confirming specific binding to a positive control binder.

The same 1 uM affinity threshold was highly selective across interactions without reported affinity measurements. Among 2,764 unreported interactions, only one appeared as a specific hit, corresponding to an apparent false-positive rate of 0.04%. A key advantage of AlphaSeq for classification tasks is that each protein is tested for binding against a library of other proteins. This library-on-library perspective provides a direct readout of specificity and allows us to distinguish specific protein interactions from nonspecific or polyreactive binding.

Although AlphaSeq separated reported from unreported interactions, the absolute K_Dapp values were not well correlated with literature affinities. This is not surprising for a comparison that combines measurements from different publications, instruments, protein formats, and assay conditions. We performed a controlled benchmark using internally generated BLI data to assess whether AlphaSeq preserves ranking within defined target systems.

Ranking

To assess whether AlphaSeq preserves the ranking of binders measured by biophysical methods, we generated an internal BLI benchmark across five anonymized target proteins previously validated for functional expression in AlphaSeq. The dataset contained 44 total BLI measurements: 37 with quantitative affinity measurements used for correlation and ranking analysis, plus 7 measurements with AlphaSeq affinity labels, but no detectable binding by BLI. Binders were drawn from internal de novo VHH design campaigns and antibody mutational series, with selections made to cover a range of K_Dapp values within each target system. This comparison was intentionally focused on binder variants expected to directly impact affinity, such as CDR changes. This matters because broader mutational scans can mix interaction effects with protein-quality effects: framework mutations or antigen mutations can destabilize the displayed protein and reduce functional presentation, causing a loss of K_Dapp that is not solely due to a weakened binding interface [7].

When comparing affinity values between AlphaSeq and BLI, AlphaSeq preserved the main binding relationships across the five target systems (Figure 4). Binders that appeared weakest by AlphaSeq tended to be weakest by BLI, including several that fell outside the measurable BLI range. Among binders with quantitative BLI measurements, ranking was generally preserved within each target system, with Spearman ρ ranging from 0.55 to 0.90 across the five systems.

As expected, each target system showed a systematic offset between AlphaSeq and BLI. Absolute affinity measurements depend on assay context: AlphaSeq is influenced by target truncation, display orientation, folding, and active protein presentation, while BLI is influenced by immobilization strategy, tag placement, and active protein fraction. We observe this context dependence even within each platform, where changing target format or immobilization strategy can shift the apparent affinity measured for the same binder.

Since absolute affinity measurements are context dependent in both AlphaSeq and BLI, target-specific calibration is necessary when comparing binders across different targets. For projects requiring interoperability between AlphaSeq and BLI across targets, we calibrate AlphaSeq measurements by measuring a small number of interactions in both AlphaSeq and BLI, then applying a target-specific y-intercept offset to the AlphaSeq K_Dapp measurements. This preserves the within-target ranking while aligning the larger AlphaSeq dataset across multiple targets. We demonstrate this approach in Figure 5, where one interaction from each target is used to set the offset before pooling the calibrated measurements from multiple targets. After calibration, the pooled AlphaSeq and BLI affinity values align closely, with 28 of 32 held-out quantitative measurements falling within one log₁₀ unit in K_D space.

Taken together, these analyses define a practical view of comparability. When targets are functionally displayed in AlphaSeq, the assay supports strong binder/non-binder classification and recovers useful affinity rankings within a defined target context. Among binders, AlphaSeq reports a quantitative K_Dapp signal across at least four orders of magnitude, enabling comparison of strong, weak, and intermediate interactions within a target, library, or variant series. When affinity values need to be compared across target systems or made interoperable with BLI or SPR, target-specific calibration is used to correct for context dependencies while preserving the within-target binding relationships. For model training, this provides a consistent, high-throughput affinity signal for learning binder/non-binder boundaries, affinity gradients, specificity patterns, and mutational effects, while BLI and SPR remain powerful orthogonal methods for generating kinetic measurements and conducting detailed validation under defined biophysical assay conditions.

Volume: Can we generate enough diverse data to learn binding behavior?

Reproducibility and comparability both relate to data quality. Volume asks whether we can generate the scale and diversity of affinity data needed to train models that learn generalizable binding relationships. This requires measuring weak and strong binders, negatives, cross-reactivity and off-target interactions, as well as mutational neighborhoods across diverse target and binder sequences.

The major advantage of AlphaSeq over biophysical methods is volume: AlphaSeq operates four orders of magnitude beyond the scale of traditional affinity measurements. That scale is possible because of a workflow that is based on pooled DNA from end to end: from a pooled DNA library input to a pooled next-generation sequencing output. This contrasts with other approaches for quantitatively measuring binding affinity, like SPR and BLI, which require isolated protein expression, purification, quality control, and measurement. With the pooled approach, AlphaSeq captures up to one million protein-protein interaction affinity measurements in parallel, while biophysical methods typically scale to a maximum of hundreds of protein interactions at a time.

One dataset that exemplifies the information volume contained in an AlphaSeq experiment is a collection of structure-guided VHH affinities. We assembled the majority of structurally characterized VHH-target interactions from the Protein Data Bank (PDB) and measured affinities for all protein interactions between 763 VHHs and 261 antigens together in a single AlphaSeq experiment (Figure 6) [8]. This corresponds to 199,143 VHH-antigen interaction measurements. This dataset serves as a convenient case study for assay volume because of the interpretability of the aggregate affinity data. When the VHHs are sorted on the x-axis by their on-target antigen from the PDB, the on-target interactions appear as a clear stripe along the diagonal. The same pooled assay also measures the surrounding interaction space. We can identify highly specific VHHs, cross-reactive VHHs (those showing strong binding for multiple homologous antigens), nonspecific VHHs (those showing strong binding for multiple unrelated targets), and polyreactive VHHs (those creating a dark vertical stripe). An important note is that on-target binding tends to be strong (~1 nM K_Dapp) while off-target binding is often far weaker, but still critical for understanding the properties of the binder and certainly for developing therapeutics where even weak nonspecific binding is a major liability.

In subsequent experiments, we explored the mutational space of a subset of these VHH–antigen interactions. For each of 100 selected VHHs, we generated approximately 100 variants containing mutations around the VHH–target interface. We then measured how those mutations strengthened or weakened binding to the respective targets, producing dense affinity landscapes. Together, these datasets comprise millions of protein affinity measurements that connect VHH-target sequence, structure, affinity, and specificity across many parental antibody-target systems and mutational neighborhoods, creating a dataset for training affinity models that would be impractical to build with biophysical assays. This is the kind of structured experimental data we believe is needed to train generalizable affinity models, and in a follow-up post we will describe how affinity landscapes can be used for that purpose.

From affinity values to affinity landscapes

AlphaSeq affinity measurements satisfy the reproducibility and comparability requirements for model training and deliver that data at a volume at least four orders of magnitude beyond what biophysical methods can practically reach. That volume changes what affinity data can do. Rather than characterizing a small number of carefully selected protein interactions, AlphaSeq maps binding across entire libraries, covering sequence diversity and mutational depth simultaneously. Affinity becomes a training signal not just for individual interactions but for understanding which sequences bind, which fail, where specificity breaks down, and how mutations reshape recognition.

The long-term goal of protein design is to move beyond finding binders by screening and toward learning the rules that make protein recognition programmable: designing proteins with defined affinity, specificity, cross-reactivity, and function. Achieving that goal requires models trained on experimental data that capture the structure of binding behavior across diverse sequences and contexts, not just isolated successful binders. Biophysical methods remain essential for detailed characterization of selected purified proteins, measuring binding on and off rates under defined assay conditions. AlphaSeq addresses a different problem: generating calibrated affinity measurements across large libraries, where binders, non-binders, off-target interactions, specificity patterns, and mutational effects are measured together. These landscapes become core infrastructure for model development: training models on binding relationships, fine-tuning them toward specific targets or scaffolds, benchmarking whether they improve, and validating their predictions experimentally.

Acknowledgements

We thank Natasha Murakowska, Ryan Emerson, and David Younger for substantial editorial guidance and critical feedback on the manuscript. We also thank Lucian DiPeso and Emily Engelhart for generating key experimental datasets that contributed to the analyses described here.

References

[1] Younger et al., 2017. High-throughput characterization of protein–protein interactions by reprogramming yeast mating. PNAS. https://doi.org/10.1073/pnas.1705867114

[2] Rich et al., 2009. A global benchmark study using affinity-based biosensors. Analytical Biochemistry. https://doi.org/10.1016/j.ab.2008.11.021

[3] Jarmoskaite et al., 2020. How to measure and evaluate binding affinities. eLife. https://doi.org/10.7554/eLife.57264

[4] Engelhart et al., 2022. A dataset comprised of binding interactions for 104,972 antibodies against a SARS-CoV-2 peptide. Scientific Data. https://doi.org/10.1038/s41597-022-01779-4

[5] Lopez-Morales et al., 2023. Titrating avidity of yeast-displayed proteins using a transcriptional regulator. ACS Synthetic Biology. https://doi.org/10.1021/acssynbio.2c00351

[6] Kramer et al., 2012. The experimental uncertainty of heterogeneous public Ki data. Journal of Medicinal Chemistry. https://doi.org/10.1021/jm300131x

[7] de Kanter et al., 2026 preprint. Effects of protein interface mutations on protein quality and affinity. https://doi.org/10.64898/2026.03.24.713863

[8] Noble, 2025. Pairing large-scale binding affinity measurements with antibody–antigen structures. To Affinity And Beyond. https://aalphabio.substack.com/p/pairing-large-scale-binding-affinity

Antibody engineering has a data problem

De novo antibody design is beginning to deliver tangible results. Multiple groups have reported design campaigns that produce validated binders for ≥50% of attempted targets and occasional double-digit target hit rates [1–8]. However, success varies widely across targets, epitopes, and studies, and most designs still fail. Current models capture meaningful signal related to antibody-antigen (Ab-Ag) binding, but their mixed performance reveals gaps in understanding key molecular interactions.

A major constraint in model development is the scale and diversity of training data. Nearly all design models rely on the PDB, which includes ~10,000 antibody-containing structures and only ~2000 VHH complexes against ~800 unique antigens. Growth is slow, with only tens to hundreds of new Ab-Ag structures added per year (Figure 2). At this pace, the scale of available training data will remain orders of magnitude below what is widely considered necessary for generalizable Ab-Ag binding affinity prediction [9]; this gap cannot be closed on any practical timeline using traditional structural biology alone.

The PDB is also inherently biased toward “positive” examples; it contains only complexes of antibodies and their intended targets that bind tightly enough to be experimentally resolved, with no information about off-target binding or non-binding. For machine learning, this creates a fundamental blind spot: models only see what works. Highly predictive models require “hard” negatives: plausible in silico structures that fail experimentally but are essential for calibrating models and reducing false positives.

We believe that overcoming the data bottleneck in antibody design requires fundamentally different training data: diverse positive and hard negative examples with experimentally validated structures and quantitative binding affinities.

Solving the data problem: designing de novo epitopes with SEPIA

SEPIA (the Synthetic EPItope Atlas) inverts the traditional de novo antibody design problem. Instead of designing antibodies to bind target antigens, we design small proteins (minibinders) to bind known antibodies (Figure 3A, top). We call these minibinders Synthetic Epitope Proteins (SEPs).

This inversion improves scaling for data generation. Minibinder design has higher hit rates than de novo antibody design [10–12], and the small size of minibinders enables rapid, low-cost DNA synthesis (an advantage that becomes especially significant compared to larger formats like scFvs). By fixing antibody paratopes and generating de novo SEPs against them, we increase experimental throughput while efficiently exploring a rich repertoire of diverse binding interfaces.

When a de novo designed protein binds strongly and specifically to its target, prior work shows that the computationally predicted complex structure typically resembles the experimentally determined one [13–15]. This means that strong, specific binding serves as a reliable proxy for the structural accuracy of a de novo designed complex. Building on this principle, we leverage AlphaSeq [16] to validate SEP binding and specificity experimentally in a high-throughput system, bypassing the need for expensive, laborious structure determination. We call these in vitro-validated structures “pseudo-structures” (Figure 3B).

Pseudo-structures come in two types: positive pseudo-structures (high-confidence designs that bind strongly and specifically to their target) and negative pseudo-structures (high-confidence designs that fail to bind). With this classification, every tested design contributes meaningful training signal and together address the needs for scale, diversity, and negative data in training the next generation of de novo antibody design models.

Experimental validation at scale

We validated the pseudo-structure approach across three experimental rounds with AlphaSeq. We observed strong, specific binding for 1,161 SEP-VHH positive pseudo-structures targeting >80 diverse VHHs and confirmed non-binding for ~45,000 SEP-VHH negative pseudo-structures.

Expanding to VHH and SEP point mutants, we also measured tens of thousands of variant interactions for a total of >75,000 strong, specific VHH-SEP mutant interactions across more than 26 million binding measurements. In doing so, we validated SEP-VHH interaction interfaces and created a rich dataset of sequences, structures, and quantitative binding labels across diverse sequence space and within local mutational landscapes.

We confirmed that SEPs were not simply recapitulating PDB structures. Only ~2% of SEP hits match any known PDB structure (TM-score >0.7), and they cluster into >250 unique fold families. This demonstrates that SEPIA explores genuinely novel structural space rather than simply rearranging known folds. Notably, the SEP design process can also be directed towards generating SEPs that mimic specific therapeutic targets, linking SEPIA structural diversity to real-world pharmaceutical design applications.

Training models on SEPIA: better ranking, better designs

We evaluated SEPIA’s impact on ML model performance across two tasks: decoy detection and de novo antibody design ranking.

Task 1: Decoy detection (Figure 4A)

We assessed whether models could distinguish between known binding VHH-antigen complexes from the PDB and non-binding “decoys” generated from swapped versions of the same VHH-antigen complexes. We trained binding classifiers on VHH-antigen complexes from SAbDab-nano (SDN) [17], SEPIA pseudo-structures, or both datasets combined. Training on the two combined datasets outperformed training on either dataset alone.

Beyond raw performance, the key insight from this result is the opportunity for continued performance improvement through data scaling. Model performance (top rank accuracy) on SDN improves with additional data but quickly plateaus at its effective limit (~400 non-redundant binding structures in this cross-validation experiment). In contrast, adding SEPIA data continues to improve model performance beyond this ceiling. Since pseudo-structures can be generated at far greater scale and speed than experimentally solved co-structures, this result points to a clear path for using SEPIA to build stronger de novo design models.

Task 2: Ranking de novo designs (Figure 4B)

Since only a small fraction of designs can be realistically tested experimentally, ranking designs to prioritize what is tested in validation is a critical and highly consequential step in de novo design workflows. We compared ipSAE [18], a current state-of-the-art confidence metric from structure prediction models, against a custom Boltz-2-based classifier that we trained on SEPIA and the decoy dataset. We call this classifier ABACUS – AntiBody Affinity Classifier Using pSeudo-structures. The task was to rank ~5,000 de novo VHH designs targeting 25 antigens (37 hits against 8 antigens) by likelihood of binding, using AlphaSeq measurements as ground truth.

Our results show that ABACUS substantially outperforms ipSAE at ranking de novo antibody designs. While ipSAE provides modest enrichment at loose thresholds (top ~30-60% of designs), its performance degrades at more stringent cutoffs, becoming worse than random selection. As noted in prior work, this failure reflects overconfidence in structure-based interface metrics, and ipSAE primarily functions as a negative filter rather than a ranking metric [19–21]. In contrast, ABACUS shows steadily increasing enrichment at more stringent cutoffs, achieving 5–10X enrichment at single-digit top-percent thresholds.

What this means for de novo design: building the SEPIA flywheel

Training on SEPIA reframes what structural data can do for de novo design. Instead of relying on structures from traditional structural biology techniques (which are scarce, biased, lack binding affinity and specificity measurements, and are limited to successful binders), we show that experimentally validated pseudo-structures provide a scalable training signal that directly links sequence, structure, and quantitative binding outcomes.

This shift in data enables a shift in capability. Models trained on SEPIA move beyond filtering obvious failures toward meaningfully ranking candidate designs, outperforming confidence metrics that were never intended to capture binding affinity. At the same time, SEPIA expands the structural landscape available for learning, sampling interaction geometries and epitope spaces largely absent from the PDB, and providing the diversity required for models to generalize beyond familiar targets.

SEPIA currently focuses on VHH–SEP interactions, but the framework is inherently extensible. We are expanding to target antibodies that lack solved structures, larger formats such as scFvs, and new design problems. One emerging direction is native epitope discovery, where SEPIA can serve as a searchable resource for identifying binding-compatible sites on natural proteins, providing new starting points for de novo antibody design. More broadly, SEPIA enables a virtuous data cycle: each round of design and experimental validation generates new data for improving model performance. Improved models generate higher-quality designs, which in turn produce more informative data. As this loop compounds, both data scale and hit rates increase together at a pace that is not achievable through traditional methods.

Ultimately, the central bottleneck in de novo antibody design is not model architecture, but data: datasets that are limited in scale, biased toward crystallizable targets, and missing hard negatives. SEPIA demonstrates that this bottleneck can be addressed by coupling large-scale computational design with high-throughput experimental validation. Rather than waiting for structural data to slowly accumulate in the PDB, we can generate the required data at scale to train the next generation of antibody design models.

For the full story, including experimental protocols, model architectures, and detailed analyses, see our pre-print.

To generate abundant, high-quality datasets like those in SEPIA for training your models, reach out to us at contact@aalphabio.com.

Acknowledgements

Alex Eng, Kerry McGowan, Davis Goodnight, Lucian DiPeso, Colleen Shikany, Emily Engelhart, Leah Homad, Miranda Lahman, Juliana Barrett, Shyam Gandhi, Mackenzie Goodwin, Kenny Herbst, Charles Lin, Mimi McMurray, Aditya Agarwal, James Harrang, Randolph Lopez, David Younger, Ryan Emerson, and Drew Duglan contributed to this work and text.

References

[1] Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion 2024. https://doi.org/10.1101/2024.03.14.585103

[2] Mille-Fragoso LS, Wang JN, Driscoll CL, et al. Efficient generation of epitope-targeted de novo antibodies with Germinal 2025. https://doi.org/10.1101/2025.09.19.677421

[3] Swanson E, Nichols M, Ravichandran S, et al. mBER: Controllable de novo antibody design with million-scale experimental screening 2025. https://doi.org/10.1101/2025.09.26.678877

[4] Stark H, Faltings F, Choi M, et al. BoltzGen: Toward Universal Binder Design 2025. https://doi.org/10.1101/2025.11.20.689494

[5] Chai Discovery Team, Boitreaud J, Dent J, et al. Zero-shot antibody design in a 24-well plate. 2025. https://doi.org/10.1101/2025.07.05.663018

[6] JAM-2: Fully computational design of drug-like antibodies with high success rates. 2025. https://nabla-public.s3.us-east-1.amazonaws.com/2025_Nabla_JAM2.pdf

[7] Didi K, Reidenbach D, Penner M, et al. Latent Generative Search unlocks de novo Design of Untapped Biomolecular Interactions at Scale. 2026. https://research.nvidia.com/labs/genair/proteina-complexa/assets/proteina_complexa_validation.pdf

[8] ByteDance Seed. PXDesign: Fast, Modular, and Accurate De Novo Design of Protein Binders 2025. https://seed.bytedance.com/en/public_papers/pxdesign-fast-modular-and-accurate-de-novo-design-of-protein-binders

[9] Hummer AM, Schneider C, Chinery L, et al. Investigating the volume and diversity of data needed for generalizable antibody–antigen ΔΔG prediction. Nat Comput Sci 2025;5:635–47. https://doi.org/10.1038/s43588-025-00823-8

[10] Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature 2023;620:1089–100. https://doi.org/10.1038/s41586-023-06415-8

[11] Pacesa M, Nickel L, Schellhaas C, et al. One-shot design of functional protein binders with BindCraft. Nature 2025;646:483–92. https://doi.org/10.1038/s41586-025-09429-6

[12] Cao L, Coventry B, Goreshnik I, et al. Design of protein-binding proteins from the target structure alone. Nature 2022;605:551–60. https://doi.org/10.1038/s41586-022-04654-9

[13] Bennett NR, Coventry B, Goreshnik I, et al. Improving de novo protein binder design with deep learning. Nat Commun 2023;14:2625. https://doi.org/10.1038/s41467-023-38328-5

[14] Yeh AH-W, Norn C, Kipnis Y, et al. De novo design of luciferases using deep learning. Nature 2023;614:774–80. https://doi.org/10.1038/s41586-023-05696-3

[15] Krishna R, Wang J, Ahern W, et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science 2024;384:eadl2528. https://doi.org/10.1126/science.adl2528

[16] Younger D, Berger S, Baker D, et al. High-throughput characterization of protein–protein interactions by reprogramming yeast mating. Proc Natl Acad Sci 2017;114:12166–71. https://doi.org/10.1073/pnas.1705867114

[17] Schneider C, Raybould MIJ, Deane CM. SAbDab in the age of biotherapeutics: updates including SAbDab-nano, the nanobody structure tracker. Nucleic Acids Res 2022;50:D1368–72. https://doi.org/10.1093/nar/gkab1050

[18] Dunbrack RL. Rēs ipSAE loquuntur : What’s wrong with AlphaFold’s ipTM score and how to fix it 2025. https://doi.org/10.1101/2025.02.10.637595

[19] Overath MD, Rygaard ASH, Jacobsen CP, et al. Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders 2025. https://doi.org/10.1101/2025.08.14.670059

[20] Smorodina E, Ali M, Kropivšek Brumat K, et al. Structural Plausibility Without Binding Specificity: Limits of AI-Based Antibody-Antigen Structure Prediction Confidence Scores 2026. https://doi.org/10.64898/2026.03.02.709004

[21] Harman J, Murakowska N. Overconfident: What Structure Prediction Confidence Scores Tell Us About Binding 2026. https://aalphabio.substack.com/p/overconfident-what-structure-prediction

A new study from researchers at the University of Oslo, A-Alpha Bio, and others demonstrates that conventional binding affinity measurements do not cleanly separate true binding strength from protein stability. A weak apparent affinity can arise for at least two different reasons: a variant may fail to adopt the correct binding geometry, or it may be unstable or poorly folded and therefore unable to present that geometry consistently. In standard binding assays, those failure modes can look similar, even though they imply very different underlying biology. This also helps explain what we observed in a previous post, namely that structure prediction confidence metrics may enrich for binders not because they directly rank intrinsic affinity, but because they filter out variants unlikely to maintain a binding-competent structure. The core challenge is posed by the following question: how do we decouple stability from binding when interpreting variant-level affinity measurements?

The Hidden Confounder in Mutation Data

Disentangling affinity from protein quality becomes especially relevant when generating large-scale binding datasets. Deep mutational scanning (DMS) allows protein engineers to study a protein binding interface and its local affinity landscape by measuring the impact of many thousands of mutational variants in parallel. In these experiments, mutations that reduce binding are typically interpreted as disrupting amino acid contacts at the binding interface. However, there is an alternative explanation: a mutation may instead impair protein folding, stability, or expression, reducing the number of functional molecules available to bind.

Observed changes in binding signal reflect a mixture of two effects:

• Protein-protein interaction changes (affinity)

• Protein quality changes (folding, stability, and expression)

Most DMS datasets used to train and benchmark binding prediction models lack controls for protein quality, creating a fundamental ambiguity in the signal these models capture.

Disentangling Binding from Protein Quality

To address the confounding effects of affinity and protein quality, the study authors developed an experimental framework to separate these effects at scale.

The experiment included four parental VHH-antigen complexes selected from SabDab [2], focusing on antigens with multiple VHHs known to bind distinct epitopes. For each antigen, a non-competitive control VHH was included that recognized a separate, non-overlapping epitope from a primary VHH. Because the two VHHs contact different epitope residues, mutations that similarly reduce binding of both are more likely to reflect changes in antigen protein quality than specific disruption of the primary VHH’s binding interface. More than 7,000 VHH and antigen variants were then generated by mutating residues at the corresponding interfaces.

Leveraging AlphaSeq to quantify protein binding in a library-on-library format, the authors measured the binding strength of all antigen variant combinations with primary and control VHHs. By comparing how each mutation affected the primary VHH versus the control VHH, the observed binding signal was separated into a “protein quality” component and a “protein interaction” component.

Failing to Learn the Binding Interface

The results reveal a striking pattern. Across the four VHH-antigen systems, 37-54% of mutations that significantly reduced binding did so primarily through protein quality effects, rather than by specifically disrupting the binding interface (Figure 1). 73-93% of epitope positions did not demonstrate clear effects on protein-protein interaction strength (i.e., via differential impact on primary VHH versus control VHH when mutated), where an epitope is defined here as being within 4.5 Å of any primary VHH atom.

From a structural perspective, positions where mutations have a clear isolated effect on binding affinity are rare compared to mutations that impact protein quality. Affinity-driving positions are most commonly residues that form direct interface contacts, are involved in hydrogen bonding, or exhibit sensitivity to charge-altering substitutions—features consistent with specific interface energetics. In contrast, most mutations that reduce binding signal act through global effects on protein stability. This distinction is reflected in model performance. Evaluation of thermostability predictors and inverse folding models revealed that both classes consistently captured protein quality effects but struggled to identify mutations that alter binding through interface-specific mechanisms.

What this Means for Model Training and Protein Design

The practical implications are clear. A model’s high performance when predicting binding signal in DMS-derived affinity datasets does not necessarily indicate an understanding of protein-protein interaction energetics. Instead, it may primarily reflect the ability to predict protein quality.

For protein designers, this reframes how common scoring functions should be interpreted. Inverse folding likelihoods, ΔΔG predictions, and related metrics are effective at filtering unstable designs, serving as valuable negative filters to shrink the viable search space. But they provide limited signal for ranking improvements in binding affinity and contain minimal predictive power that is specific to a particular protein interaction.

For certain situations such as antibody optimization, improving a conflated stability-and-affinity signal may be acceptable or even desirable when gains in both properties are beneficial to overall antibody performance. However, if models have primarily learned global properties about protein quality, then the field’s ability to predict affinity-enhancing mutations at specific protein-protein interfaces remains largely untested against properly controlled data. Training the next generation of interaction models and designing de novo antibodies will require datasets that explicitly separate protein quality from protein interaction—either through binding controls, as demonstrated here, or through orthogonal measurements of stability and expression. To this end, we have taken first steps to isolate the protein interaction signal during training from protein quality effects, and initial results suggest that the separation is informative.

Without separating these variables, models will continue to learn what is easiest to observe most broadly: not how specific proteins bind, but whether they express and fold.

Acknowledgements

Emily Engelhart, Colleen Shikany, Mackenzie Goodwin, Shyam Gandhi, Kenny Herbst, Juliana Barrett, Charles Lin and Mimi McMurray contributed to the lab experiments. Emily Engelhart, Kerry McGowan, Natasha Murakowska, Victor Greiff, Eva Smorodina, Ryan Emerson, and Drew Duglan contributed to this work and text.

References

[1] de Kanter JK, Smorodina E, Minnegalieva A, et al. Effects of protein interface mutations on protein quality and affinity. bioRxiv 2026. https://doi.org/10.64898/2026.03.24.713863

[2] Dunbar J, Krawczyk K, Leem J, et al. SAbDab: the structural antibody database. Nucleic Acids Res 2014;42(D1):D1140–6. https://doi.org/10.1093/nar/gkt1043

Here, we examine the relationship between Boltz-2 confidence scores and experimentally measured binding affinity for AlphaSeq datasets spanning tens of thousands of VHHs, scFvs, native and engineered antigens, de novo minibinders, and mutational scans totaling >7 million quantitative binding affinity measurements. We evaluate the impact of templating during prediction and analyze how confidence scores perform across distinct design scenarios. Our findings reveal that confidence scores work best as negative filters: they effectively remove many non-binders. However, they fail to reliably rank affinity, produce high false positive rates, and especially struggle to capture the effects of point mutations. Moving beyond filtering will require explicitly training models on large, quantitative datasets that span both broad sequence diversity and fine-grained mutational landscapes.

Confidence Scores are Central to Modern Protein Design Workflows. But How Well do they Work?

Structure prediction models such as AlphaFold, RosettaFold, and Boltz-2 are key elements of de novo protein design workflows, both for generating candidates and ranking them for experimental testing. A common strategy for these design approaches is to optimize and filter on interface-focused confidence scores with the goal of enriching true binders [1–4].

Current de novo design methods [5–10] extend this approach through integrated docking and sequence proposal strategies, upgraded structure prediction models, and expanded filtering criteria. Many methods explicitly optimize interface confidence scores, such as interface predicted TMscore (ipTM) and interface predicted score from aligned errors (ipSAE) [11]. Retrospective analyses of minibinder [12] and antibody [9] datasets consistently rank these metrics among the strongest computational predictors of binding, and ipSAE remains the primary screening metric for wet-lab validation in recent protein design competitions [13].

Despite their widespread use, the predictive value of confidence scores is more limited than their prevalence suggests. Recent work has begun to quantify this: in the Adaptyv EGFR competition, ipTM achieved an AUROC of 0.64 for binding classification, leading the authors to conclude that these metrics were “insufficiently predictive of true experimental binding probability or affinity” [13]. We observed a high false positive rate for ipSAE in an AlphaSeq dataset of inverse-folded VHH variants. Point mutation campaigns are a particular blind spot: AlphaFold is known to poorly capture the energetic effects of point mutations [14], and metrics like ipSAE have not been systematically evaluated on point mutation data. Quantitative datasets at the scale of AlphaSeq—spanning tens of thousands of measurements across diverse binder formats, engineered antigens, and mutation-rich campaigns—offer an opportunity to characterize more precisely where confidence metrics succeed and where they fail.

We ask two main questions in this analysis:

• How robust are interface confidence scores across native, engineered, and de novo PPIs, particularly for antibody-antigen systems?

• How do structure prediction variations, such as providing a template structure during inference, impact the performance of confidence scores?

To answer these questions, we generated Boltz-2 complex predictions for tens of thousands of known and designed PPIs with experimentally measured AlphaSeq affinities. These datasets included de novo minibinders and antibodies, wildtype and engineered antigens, and mutation-rich optimization campaigns, each representing a distinct protein design use case (Table 1). Our previous blog posts describe our inverse folding and SabDab-nano (SDN) point mutant datasets. The SEPIA datasets are part of a novel “synthetic antigen” data generation approach that we will discuss in greater detail in a forthcoming preprint; briefly, SEPIA datasets consist of de novo minibinders designed to bind to VHH paratopes.

Structure Prediction and Evaluation Framework

To produce confidence scores for each measured PPI, we used Boltz-2 to predict three complex structures (diffusion samples) without generating multiple sequence alignments. We also investigated the effect of providing different structural inputs (templates) during Boltz-2 complex prediction, as previous work shows that providing different structural priors during complex prediction affects agreement between confidence scores and experimental success [1,3]. We examined three different templating strategies:

No templates – Provide only the sequences of the binder and antigen.

Soft templating – Provide the full complex structure.

Hard templating – Provide the complex and constrain predictions within a distance threshold (3Å) of the supplied template.

Binder-antigen co-structures were supplied as templates for all datasets except for the scFv point mutant dataset, which lacked a solved antibody-antigen co-structure. For the scFv dataset, the antigen monomer structure was provided for both soft and hard templating.

For each dataset, we evaluated how well ipSAE classified designs as binders (AlphaSeq affinity <1μM) or non-binders by computing the area under receiver-operator and precision-recall curves (AUROC and AUPRC). We also reported AUPRC fold enrichment, which is defined as AUPRC divided by the known positive binder fraction (Table 2). For each dataset, we defined an optimal ipSAE cutoff using Youden’s J-index (i.e. the threshold that best balances discrimination of binding/non-binding) [15].

ipSAE’s Performance

VHH Binder Classification

Across the four VHH-containing datasets, which consist of ~97,000 measured VHH-antigen interactions (Table 1), soft templating consistently outperformed both hard templating and no template strategies (Figure 2a, Tables 2 and 3). Soft templating achieved the highest true positive rate (17.2%) but also admitted the most false positives (19.6%). More broadly, ipSAE exhibited a persistently high false positive rate (>10%) across templating strategies. This finding reveals a consistent trade-off when using ipSAE as a binding classifier: a higher true positive rate also permits more false positives. In practice, ipSAE behaves primarily as a negative filter: it reliably removes many non-binders but struggles to cleanly separate strong binders from geometrically plausible failures.

VHH Point Mutation Classification

Mutation-level prediction remains challenging, as we and others have noted [14], with performance varying markedly across affinity predictions of point mutants. We compared two VHH datasets that introduce single mutations into VHH paratopes: SDN point mutants (VHH point mutants binding natural antigens) and SEPIA point mutants (VHH point mutants binding de novo minibinders). For SDN point mutants, ipSAE strikingly performed no better than random selection. In contrast, SEPIA point mutants showed strong binder enrichment (5.2X) for functional paratope mutants, comparable to the binder enrichment observed for the more sequence-diverse base SEPIA dataset (5.5X). One possible explanation is dataset context: SDN variants are derived from VHH–antigen structures present in the PDB that are present in Boltz-2 training data, whereas SEPIA interactions involve de novo designed minibinders that are absent from public structural databases. Together, these results suggest that ipSAE can detect mutation-driven functional changes in some settings, but with substantially lower reliability when compared to broader design tasks.

scFv Point Mutant Binding Classification

For the scFv mutant dataset, templating was critical for improving binder classification. Hard templating performed best globally, while soft templating performed best when averaging across antigens. However, soft and hard templating both showed variability, possibly due to the antigen monomer structure being used as the template for this dataset, as opposed to a full complex. These results reveal a practical lesson: when co-structures are unavailable, locking in antigen geometry (with a known or predicted monomer structure) can recover useful confidence signal. Without structural priors, ipSAE approaches noise.

Remodeled Antigen Binding Classification

Addressing challenging antigens is often critical in protein engineering, where truncated or stabilized variants are required to enable antigen purification, yeast display, or screening. This was the motivation behind our “Remodeled Antigens” dataset. It contained computationally remodeled antigens binding to VHHs, scFvs, and native binding partners with known co-crystal structures. Soft templating again performed markedly better than other templating strategies (Tables 2 and 3). Strikingly, soft-templated predictions yielded a 3.6x binder fold enrichment (Table 2). These results reveal that for substantially modified antigens, soft-templated complex predictions provide the most reliable discrimination between binders and non-binders and are more robust than either enforcing a strict structural prior or providing no structural prior at all.

Increased Confidence Requires Affinity Data at Scale

Across tens of thousands of PPI measurements spanning diverse systems of engineered binders and antigens, several patterns emerge.

First, structure prediction confidence scores are useful, but they do not perfectly differentiate between binders and non-binders and are ineffective at predicting relative binding affinities. Metrics like ipSAE enrich for binders, but many high-confidence binders fail when tested experimentally. In practice, confidence scores function best as negative filters: effective at removing poor designs, unreliable for ranking the best ones.

Second, structural context strongly influences performance. Providing templates during Boltz-2 prediction markedly improves binder classification, with soft templating consistently outperforming both hard constraints and no templates. Predictions made without structural priors produce confidence scores that poorly classify binders, while hard templating helps primarily in cases where geometry must be tightly enforced.

Lastly, mutation-heavy campaigns expose the limitations of confidence scores. Confidence scores often struggle to detect the subtle energetic changes that determine whether a single mutation improves or weakens binding, particularly for larger or structurally complex antigens. Performance varies by antigen, likely reflecting biases in training data, with familiar targets possibly yielding more reliable confidence scores than unfamiliar ones.

Large, quantitative datasets make these patterns visible. At scale, we observe overconfident non-binders, hallucinated interfaces, and antigen-specific blind spots. These results suggest that ipSAE and related metrics primarily capture bulk geometric plausibility rather than the energetic differences that determine binding affinity. The variability in confidence score performance for classifying binders is not so much a failure of structure prediction, but rather a mismatch between what confidence scores measure and how they are often used.

For protein designers, there are several practical takeaways:

• Use confidence scores to filter out poor designs, not to rank binding strength.

• Provide structural templates whenever possible. Even antigen monomer structures help.

• Exercise caution when applying confidence metrics to mutation-heavy optimization campaigns.

• Expect performance to vary by antigen. Better performance may be expected for antigens that are better represented in training data.

Looking forward, the path to more reliable confidence scores likely depends on the training data. Current confidence models were not trained on large, quantitative affinity measurements – they were trained on structural databases that capture only a relatively small, biased sampling of experimentally known co-complexes, with no information on how tightly they bind. Closing this gap will require explicitly training confidence models on datasets that span both highly diverse sequence and structure space, as well as local sequence perturbations. Large-scale affinity datasets of the kind described here represent exactly this opportunity: not just a benchmark for evaluating today’s tools, but a resource for building better ones.

Acknowledgements

Joseph Harman, Aditya Agarwal, Nick Altieri, Adrian Lange, David Noble, Kerry McGowan, and Natasha Murakowska performed the experimental design for these datasets. Davis Goodnight, Emily Engelhart, Mackenzie Goodwin, Shyam Gandhi, Kenny Herbst, Juliana Barrett, Charles Lin and Mimi McMurray contributed to the lab experiments. Joseph Harman, Natasha Murakowska, Nick Altieri, David Noble, Kerry McGowan, Adrian Lange, and Drew Duglan contributed to this work and text.

References

[1] Bennett NR, Coventry B, Goreshnik I, et al. Improving de novo protein binder design with deep learning. Nat Commun 2023;14:2625. https://doi.org/10.1038/s41467-023-38328-5.

[2] Wicky BIM, Milles LF, Courbet A, et al. Hallucinating symmetric protein assemblies. Science 2022;378:56–61. https://doi.org/10.1126/science.add1964.

[3] Bennett NR, Watson JL, Ragotte RJ, et al. Atomically accurate de novo design of antibodies with RFdiffusion 2024. https://doi.org/10.1101/2024.03.14.585103.

[4] Watson JL, Juergens D, Bennett NR, et al. De novo design of protein structure and function with RFdiffusion. Nature 2023;620:1089–100. https://doi.org/10.1038/s41586-023-06415-8.

[5] Pacesa M, Nickel L, Schellhaas C, et al. One-shot design of functional protein binders with BindCraft. Nature 2025;646:483–92. https://doi.org/10.1038/s41586-025-09429-6.

[6] Nabla Bio, Biswas S. De novo design of epitope-specific antibodies against soluble and multipass membrane proteins with high specificity, developability, and function 2025. https://doi.org/10.1101/2025.01.21.633066.

[7] Chai Discovery Team, Boitreaud J, Dent J, et al. Zero-shot antibody design in a 24-well plate 2025. https://doi.org/10.1101/2025.07.05.663018.

[8] Mille-Fragoso LS, Wang JN, Driscoll CL, et al. Efficient generation of epitope-targeted de novo antibodies with Germinal 2025. https://doi.org/10.1101/2025.09.19.677421.

[9] Swanson E, Nichols M, Ravichandran S, et al. mBER: Controllable de novo antibody design with million-scale experimental screening 2025. https://doi.org/10.1101/2025.09.26.678877.

[10] Stark H, Faltings F, Choi M, et al. BoltzGen: Toward Universal Binder Design 2025. https://doi.org/10.1101/2025.11.20.689494.

[11] Dunbrack RL. Rēs ipSAE loquuntur : What’s wrong with AlphaFold’s ipTM score and how to fix it 2025. https://doi.org/10.1101/2025.02.10.637595.

[12] Overath MD, Rygaard ASH, Jacobsen CP, et al. Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders 2025. https://doi.org/10.1101/2025.08.14.670059.

[13] Cotet T-S, Krawczuk I, Stocco F, et al. Crowdsourced Protein Design: Lessons From the Adaptyv EGFR Binder Competition 2025. https://doi.org/10.1101/2025.04.17.648362.

[14] Pak MA, Markhieva KA, Novikova MS, et al. Using AlphaFold to predict the impact of single mutations on protein stability and function. PLOS ONE 2023;18:e0282689. https://doi.org/10.1371/journal.pone.0282689.

[15] Ruopp MD, Perkins NJ, Whitcomb BW, et al. Youden Index and Optimal Cut‐Point Estimated from Observations Affected by a Lower Limit of Detection. Biom J 2008;50:419–30. https://doi.org/10.1002/bimj.200710415.

In this article, we break down what de novo actually means, why performance claims vary so widely, and how to understand the true difficulty of different design problems. We introduce a simple tier system that defines a spectrum of de novo design difficulty, and we discuss where today’s models excel and fail. By the end, you’ll have a clearer lens for evaluating any DND claim—ours included.

Navigating the complexity of antibody engineering

De novo protein design (DND) of antibodies, including IgGs, scFvs, and VHHs, offers a powerful strategy for exploring new therapeutic solutions to complex diseases. Traditional approaches to antibody generation leverage animal immunization or phage panning to identify “seed” candidates that are later optimized for affinity, specificity, and developability. However, these approaches require a sufficiently immunogenic epitope and do not provide any control over the epitope or paratope, limiting their ability to selectively engage a precisely defined binding interface [7,8]. In contrast, in silico designs can be engineered to intentionally engage specific epitopes on a target, bypass many of the pitfalls of antibody optimization, and enable drug development against previously inaccessible targets [1-6, 18–20].

As of 2026, leading generative models for protein engineering employ both structure and sequence-based approaches to design and screen in silico candidates that are validated in vitro [2–6, 18–20]. These methods report both impressive hit rates across multiple targets and a striking concordance between the designed and experimentally confirmed structures. This indicates that modern DND approaches can explicitly dictate binding location and orientation, and that in silico structural representations capture biologically meaningful features of the epitope-paratope interface [1–6, 19, 20].

Structure-based design methods must ensure predicted sequences capture appropriate interactions across the local antibody-antigen (Ab-Ag) interface, since these local interactions dictate binding strength and specificity. Although global fold accuracy matters, recent state-of-the-art (SOTA) approaches increasingly incorporate interface-focused filters that are empirically associated with experimentally higher success rates [2, 4]. This trend underscores a broader principle: models that learn representations directly from functional protein-protein interfaces should provide better designs.

Yet, a key challenge remains: how do we determine the generalizability of antibody DND models? Reported hit rates in the literature vary widely, and each SOTA method has only partial overlap in the targets tested in vitro, leading to challenges in benchmarking and direct model comparisons [1–7, 18–20]. Moreover, not all epitopes present the same level of difficulty for de novo design, which makes performance hard to interpret. Some epitopes appear directly or indirectly in model training data, yet systematic analyses of interface-level data leakage are rare [23]. Because experimentally solved Ab–Ag structures are limited and biased toward a narrow set of drug targets that can be co-crystallized with an antibody, it’s often unclear where models excel versus struggle in epitope design.

In that case, what design tasks are truly “de novo”? To clarify what counts as de novo in the context of epitope targeting, we outline a tiered framework that categorizes design tasks by their underlying structural information and leakage risk, providing a more consistent view of model performance and generalizability.

Tier 1: An explicit antibody-antigen binding complex is known, and a model can leverage the bound structure as a template, regardless of whether the parental sequence appears in the training distribution. The model is given the “correct” interface and needs to modify or adapt to either preserve or enhance the interaction to the known epitope. Novel antibody designs arise from the model’s ability to transfer the known interaction geometry onto an alternative framework or generate a distinct paratope that nonetheless engages the same binding site. Because the binding mode and epitope-paratope geometry are fixed, tier 1 is template-guided re-design of a known interaction rather than truly novel creation of a new interaction. For this reason, we do not consider tier 1 design problems de novo design.

Tier 2: An antibody-antigen binding complex is not known, but some binding complex is available for the epitope of interest (e.g. to a native, non-antibody binding partner). The model’s task is to design an antibody against a known interface, thus probing its ability to generate a complementary surface, potentially inspired by a native binding partner. Novelty reflects the model’s ability to translate a native protein-protein interface into an antibody-compatible binding mode for which no known antibody-bound structure is currently known.

Tier 3: No binders or binding geometries are known for the epitope of interest. Absent of this data, Tier 3 measures whether a model infers not only where to bind, but how to generate a compatible paratope without interface supervision. Model generalizability likely depends on how similar the putative epitope is to epitopes seen in the training distribution of the model, but the success rate of designing binders also depends on biophysical complexity.

The tiered framework is most useful when combined with epitope-specific analysis. While we expect success rates to decrease on average as structural supervision is removed, antibody DND performance depends strongly on whether a target epitope resembles interfaces previously observed during training. Inherent epitope properties (i.e. level of disorder, hydrophobicity, solvent accessibility) must also be considered as part of assessing target and epitope difficulty; tier 3 does not guarantee novelty, nor does tier 1 guarantee tractability. Instead, the framework highlights where models may be relying on learned interface priors versus true inference of new binding modes. Future benchmarks should therefore pair tier-based categorization with quantitative measures of epitope similarity and biophysical accessibility, enabling clearer distinctions between interpolation, extrapolation, and genuine de novo design.

Recipe for success: key ingredients in a de novo pipeline

In broad strokes, de novo pipelines have 2 key steps: proposal generation and scoring. Pipeline efficacy depends on the intrinsic difficulty of the task (tiers), but also the extent to which models accurately perform these steps. Earlier pipelines often viewed the proposal step sequentially from complex generation; structure models were explicitly fine-tuned on antibody datasets to enable complex generation, while an independent inverse-folding model would generate sequence proposals, as with RFAntibody [1].

In this last year, a multitude of co-design methods have combined structure prediction and design into one concurrent step. Some, including Germinal and mBER, combine a strategy like ColabDesign, where gradients of Alphafold-Multimer and a protein language model (pLM) of choice enable a structure-informed sequence proposal [2, 3]. Separately, BoltzGen enables the proposed design at the structural level directly [4]. After designs are proposed, all methods undergo a filtering and ranking step. Universally, all strategies will leverage some structural confidence metric upon refolding (such as iPTM, pAE) to select their candidates [1 – 4].

We summarize the SOTA methods for the top open-source models and select closed-source models specifically in the context of antibody–antigen design:

Several models above report impressive hit rates. For example, Chai-2 produced successful de novo antibodies for 26/52 antigens while only testing ≤20 designs per antigen, and JAM-2 successfully produced binders against 16/16 antigens tested (each with highly variable hit rates per antigen) [6, 20]. However, because each of these studies evaluated designs against different epitopes, with varying levels of difficulty and different design objectives and targets, it’s nearly impossible to compare model performance on a common basis. A definitive answer requires a controlled, head-to-head benchmark in which all models are tasked with the same problem across different targets.

To address this issue, we are currently benchmarking several of the open-source antibody DND models above against a standardized panel of the same antigens and epitopes—which span the tiers of DND difficulty—in a comprehensive AlphaSeq experiment that will enable direct model comparison. We will report the results of this “bake-off” in a follow-up blog post. Below, we describe earlier benchmarking studies of inverse folding models for antibody DND that were performed in late 2024 and early 2025.

Inverse folding’s open-book exam

In a pilot Tier 1 study, we benchmarked a panel of inverse folding models on the task of re-designing VHHs present in known antibody-antigen co-crystal structures. We used 2 test systems: SARS-CoV RBD and 41BB. We generated ~34,000 VHH designs using proteinMPNN, AbMPNN, ESM-IF, AntiFold, and SaProt [13–17]. VHH design windows included both CDR and CDR + framework regions.

A critical feature of this pipeline is that no re-folding or filtering step was used for design selection. This was intentional, as we sought to isolate how each inverse folding model performs on an “easy” DND task without any additional filtering. Further, the scale of AlphaSeq enables us to quantitatively measure binding affinities for each design tested and then ask: “how much better would we have done?” were we to have applied any given set of filters—including complex prediction metrics, the results of which we discuss in the next section.

The resulting designs were synthesized and experimentally tested for binding affinity. We compare model-generated designs, against the strength of affinity of these designs against the parent sequence—the original VHH sequence with an antigen-bound crystal structure.

The results are striking; despite being given an explicit crystal of the binding complex, the bulk of designs across most models are significantly worse than the original sequence (Fig. 2). Generally, AbMPNN and proteinMPNN (pMPNN) style models performed best, consistent with findings published earlier this year [18].

We explored the results further, looking at the relationship between the designed region and the affinity scores. We find that enabling the design of the CDRs and frameworks contiguously may “rescue” some methods, but still rarely competes with the original sequence (Fig 3.). Unsurprisingly, as models extrapolate further from the original parent sequence, performance degrades (Fig 4).

The above experiment indicates that design models can yield a wide variance in plausible sequences, highlighting how critical the screening/ranking step is before validating designs.

All metrics are right in silico, some are right in vitro

Structure prediction methods like AlphaFold and Boltz-2 assign per-residue and pairwise residue confidence values for predicted protein complexes. Interface-focused variants of these metrics, such as interface-predicted TMScore (ipTM) and interaction prediction Score from Aligned Errors (ipSAE) [12], have been shown to correlate with binding affinity for de novo designed minibinders [9, 10]. Indeed, recent SOTA DND methods filter and rank their designs according to carefully calibrated combinations of confidence metrics from structure prediction models [1–4].

We examined the relationship between AlphaSeq binding affinity and ipSAE for designs generated in the inverse folding experiment (Fig.5). We predicted complex structures of the designed VHH sequences bound to the two target proteins using Boltz-2 without MSAs, providing the known crystal structure as a template.

We observe striking concordance between high ipSAE scores and strong binders (low affinity), as well as low ipSAE scores and poor binders (high affinity). Notably, there are a non-negligible number of high ipSAE scores with poor binding (false positives).

At first glance, this relationship is very encouraging: modern structure prediction methods capture coarse binding signal and distinguish between binders and non-binders with surprising fidelity (additionally observed by other groups [18]). This also suggests that interface confidence metrics encode some biophysically meaningful information about the compatibility of two putative binder candidates.

The density plot reveals a critical inflection point for de novo design: false positives. A striking fraction of designs fall within this regime, where model confidence (as indicated by ipSAE) is high, yet binding affinity is weak. This disconnect highlights a key limitation of current pipelines: without explicit screening, many designs are statistically plausible in silico but doomed to fail in reality.

Learning what not to do

Most design frameworks explicitly enrich for high-confidence predictions, prioritizing “what looks right” according to internal model priors. However, this strategy can be self-reinforcing: by filtering for structural plausibility rather than functional truth, we discard informative counterexamples that expose the limits of a model’s learned biases. While filtering and ranking are indispensable for practical success, they also come at a considerable computational cost, requiring re-folding to assess the quality of each design.

The most informative signal may be in the examples where the model confidence diverges from experimental truth. Crystallographic structures capture ground-truth data that contain only positive examples. True generalization depends on learning from negative examples as well, i.e. sequences that may fold, yet fail to function. To learn what matters, models must learn what not to do. Such examples reveal a model’s inherent biases and point to opportunities to deliberately integrate counterfactual supervision to improve proposal diversity and discriminative performance.

We imagine several strategies to help solve the problem. These include:

1. Synthetic augmentation: Increasingly, teams are incorporating synthetic data into their workflows to greatly expand the available training data for models. Compared to other domains, annotated data in protein-protein complexes remains scarce (for example the PDB is 6-8 orders of magnitude smaller than the training corpus of most conventional LLMs) [11, 21, 22].

2. Affinity supervision: Incorporating quantitative affinity data can enable models to disentangle relationships between thermodynamic stability and binding energetics. Some high-confidence designs may indicate sterically plausible but unfavorable interactions; with annotated data of both positive (strong binding) and negative (weak binding/non-binding) examples, models can contextualize their designs for function.

3. Local vs global examples: Generalizability often trades off with local sensitivity. Training datasets need to incorporate both “global” examples of structural diversity balanced with local perturbations of the sequences that can influence binding. Including point mutants of known binders helps models learn how subtle geometric or physicochemical changes can drastically alter affinity.

Ultimately, advancing generalization in de novo design requires datasets and benchmarks that go beyond what succeeds in silico or is available in the PDB. We must capture not only the manifold of functional proteins, but also the boundaries where current models fail.

Acknowledgements

Joseph Harman, Aditya Agarwal, Nick Altieri, and Adrian Lange performed the experimental design for these datasets. Davis Goodnight, Mackenzie Goodwin, Shyam Gandhi, Kenny Herbst, Juliana Barrett, Charles Lin and Mimi McMurray contributed to the lab experiments. Natasha Murakowska, Joseph Harman, Nick Altieri, David Noble, Kerry McGowan, Adrian Lange, and Drew Duglan contributed to this work and text.

References

1. Bennett, N. R., et al. “Atomically Accurate De Novo Design of Antibodies with RFdiffusion.” Nature, 2025, https://doi.org/10.1038/s41586-025-09721-5.

2. Swanson, Erik, Michael Nichols, Supriya Ravichandran, Pierce Ogden, et al. “mBER: Controllable de novo Antibody Design with Million-Scale Experimental Screening.” bioRxiv, 26 Sept. 2025, doi:10.1101/2025.09.26.678877.

3. Mille-Fragoso, L. S., Wang, J. N., Driscoll, C. L., Dai, H., Widatalla, T., Zhang, X., Hie, B. L., & Gao, X. J. (2025). Efficient generation of epitope-targeted de novo antibodies with Germinal. bioRxiv. https://doi.org/10.1101/2025.09.19.677421

4. Stärk, Hannes, et al. “BoltzGen: Toward Universal Binder Design.” arXiv Preprint, 2025, https://hannes-stark.com/assets/boltzgen.pdf

5. Biswas, Surojit, et al. “De novo design of epitope-specific antibodies against soluble and multipass membrane proteins with high specificity, developability, and function.” bioRxiv, 22 Jan. 2025, doi:10.1101/2025.01.21.633066.

6. Chai Discovery Team. “Zero-Shot Antibody Design in a 24-Well Plate.” bioRxiv, 5 July 2025, doi:10.1101/2025.07.05.663018.

7. Caradonna, Timothy M., and Aaron G. Schmidt. “Protein engineering strategies for rational immunogen design.” npj Vaccines, vol. 6, 2021, article 154, https://doi.org/10.1038/s41541-021-00417-1.

8. Laustsen, Andreas H., Victor Greiff, Aneesh Karatt-Vellatt, Serge Muyldermans, and Timothy P. Jenkins. “Animal Immunization, in Vitro Display Technologies, and Machine Learning for Antibody Discovery.” Trends in Biotechnology, vol. 39, no. 12, 2021, pp. 1263-1273, https://doi.org/10.1016/j.tibtech.2021.03.003

9. Overath, Max D., Andreas Rygaard, Christian P. Jacobsen, Valentas Brasas, Oliver Morell, Pietro Sormanni, and Timothy P. Jenkins. “Predicting Experimental Success in De Novo Binder Design: A Meta-Analysis of 3,766 Experimentally Characterised Binders.” bioRxiv, 14 Aug. 2025, doi:10.1101/2025.08.14.670059.

10. Bennett, N.R., Coventry, B., Goreshnik, I. et al. Improving de novo protein binder design with deep learning. Nat Commun 14, 2625 (2023). https://doi.org/10.1038/s41467-023-38328-5

11. Abramson, J., Adler, J., Dunger, J., et al. “Accurate structure prediction of biomolecular interactions with AlphaFold 3.” Nature, vol. 630, 2024, pp. 493–500. https://doi.org/10.1038/s41586-024-07487-w

12. Dunbrack, Roland L., Jr. “ipSAE: What’s Wrong with AlphaFold’s ipTM Score and How to Fix It.” bioRxiv, 10 Feb. 2025, doi:10.1101/2025.02.10.637595.

13. Dauparas, Justin, I. Anishchenko, N. Bennett, H. Bai, R. J. Ragotte, L. F. Milles, et al. “Robust Deep Learning-Based Protein Sequence Design Using ProteinMPNN.” Science, vol. 378, no. 6615, 2022, pp. 49-56. doi:10.1126/science.add2187

14. Dreyer, C., et al. “Inverse Folding for Antibody Sequence Design Using Deep Learning.” arXiv, 31 Oct. 2023, arXiv:2310.19513.

15. Alexander Rives, Joshua Meier, et al. “Biological Structure and Function Emerge from Scaling Unsupervised Learning to 250 Million Protein Sequences.” PNAS, vol. 118, no. 15, 2021. https://doi.org/10.1073/pnas.2016239118

16. Høie, Magnus Haraldson, Alissa Hummer, Tobias H. Olsen, Broncio Aguilar-Sanjuán, Morten Nielsen, and Charlotte M. Deane. “AntiFold: Improved Antibody Structure-Based Design Using Inverse Folding.” Bioinformatics Advances, vol. 5, no. 1, 2024, article vbae202. doi:10.1093/bioadv/vbae202

17. JinSu Chenchen Han, Yuyang Zhou, Junjie Shan, Xibin Zhou, and Fajie Yuan. “SaProt: Protein Language Modeling with Structure-aware Vocabulary.” bioRxiv, 1 Oct. 2023, doi:10.1101/2023.10.01.560349.

18. Janusz, Bartosz, et al. “Benchmarking Antigen-Aware Inverse Folding Methods for Antibody Design.” bioRxiv, 5 Aug. 2025, doi:10.1101/2025.08.05.668698.

19. Chai Discovery Team, et al. “Drug-like antibody design against challenging targets with atomic precision.” bioRxiv, 29 Nov. 2025, doi:10.1101/2025.11.29.691346.

20. Nabla Bio, Inc. JAM-2: Fully computational design of drug-like antibodies with high success rates. 2025.; https://nabla-public.s3.us-east-1.amazonaws.com/2025_Nabla_JAM2.pdf

21. RCSB Protein Data Bank. Protein Data Bank, Research Collaboratory for Structural Bioinformatics, 2025, https://www.rcsb.org

22. Brown, Tom B., et al. “Language Models are Few-Shot Learners.” arXiv.org, 22 July 2020. arXiv:2005.14165

23. Kovtun, Daniel, et al. PINDER: The Protein Interaction Dataset and Evaluation Resource. bioRxiv, 17 July 2024, doi:10.1101/2024.07.17.603980.

The Challenge of Antibody Engineering

Engineering clinically relevant antibodies is, at its core, an exercise in traversing high-dimensional fitness landscapes. Training models that can learn relevant features in this complex space will benefit from extensive, high-quality datasets, ideally with both structural and affinity labels. However, most antibody-antigen datasets illuminate only one or two aspects of the problem (sequence, or structure, or function), are often small, lack antibody or antigen diversity, or aggregate multiple assay outputs with inconsistent conditions [1–7].

Why Current Data Falls Short

Recent modeling advances have shown that multi-modal approaches can design functional antibodies. However, algorithmic innovation is outpacing the data being generated in antibody design; the bottleneck now is the scarcity of high-quality training data [2, 7, 8]. Antibodies exacerbate this challenge: unlike enzymes or natively evolving interacting proteins, they lack strong co-evolutionary signals in sequence databases, so traditional approaches are less informative for paratope–epitope recognition.

As of October 2025, structural coverage of VHH–antigen complexes is limited (<1,000 non-r-redundant entries, <100 with quantitative affinity) [1]. Sequence repositories such as OAS contain orders of magnitude more repertoires but are decoupled from antigen identity and lack functional ground truth [9]. Variant-level datasets such as deep mutational scans are uneven: biased toward a handful of parental antibodies and antigens, heterogeneous in assay readouts, and rarely providing matched negatives or cross-reactivity profiles. Synthetic data can extend but not replace real, diverse measurements.

Developing generalizable antibody ML models will require diverse, comprehensive data [2]. Robust design depends on these models not only learning features of strong binders, but also learning factors that impair binding – yet rich datasets that capture both positive and negative examples under consistent conditions remain scarce.

Our Dataset: Systematically Measuring VHH–Antigen Space

To address this gap, we used AlphaSeq to measure binding affinities for ~2 million antibody–antigen interactions across two assays:

1. Comprehensive coverage of VHH–antigen interactions from SAbDab-nano, including both on-target complexes and off-target mismatches (about 500K interactions).

2. Targeted mutational scans across CDRs for a subset of VHHs, introducing interface diversity with quantitative affinity labels (an additional 1.5M interactions).

AlphaSeq’s core strength is multiplexed, library-on-library, quantitative readout of pairwise protein–protein interactions under consistent, reproducible conditions [10]. This produced a dense matrix of VHHs × targets, where each interaction is a quantitative binding measurement. Figures 1 and 2 illustrate the global binding landscape and local mutational fitness landscapes measured in these assays respectively.

Measuring on- and off-target binding in SabDab-nano

Binding affinities for 261 antigens against 763 VHHs – for a total of 200,000 interactions, are shown in the heatmap below (Figure 1). This experiment reveals both expected on-target specificity between known interacting VHH:antigen pairs, as well as off-target interactions. Along the diagonal, we see strong, high-affinity signals where VHHs engage their intended antigen targets. A few off-diagonal signals also emerge: horizontal streaks reflect antigens bound by multiple VHHs, while vertical streaks indicate promiscuous VHH binding. Some off-target interactions are expected, including VHHs known to bind species homologs of their target. Others are novel or unexpected. These off-target effects underscore the intricate complexity of antibody-antigen recognition: even single-domain antibodies may exhibit cross-reactivity depending on structural similarity of the epitope or target of interest. The library-on-library nature of AlphaSeq enables us to quantify these interactions, providing extensive off-diagonal data for us to feed into antibody engineering models.

Figure 1 illustrates binding specificity and, in some cases, promiscuity. In a follow-up AlphaSeq experiment, we generate VHH point mutants against their intended targets to examine the sensitivity of these interactions to small perturbations. By introducing point mutations in each VHH, we observed substantial differences in binding relative to the parent VHH:antigen interaction. We probe this difference by measuring the delta-affinity (ΔK_d), or change in binding affinity of a VHH variant compared to its parent VHH against the intended target. A negative ΔK_d, indicates improved binding, while a positive ΔK_d reflects weakening binding. The results are striking; while many variants bind weaker than their parents, a surprising subset achieve stronger binding than expected (best ΔK_d <= -2.0, worst = 4.54, average = 0; AlphaSeq affinity scores are in log10 space). From a design perspective, this highlights many untapped opportunities to improve affinity.

Taken together, these two experiments emphasize the dual nature of antibody engineering. On one hand, VHHs can exhibit broad cross-reactivity, binding to multiple related antigens. However, they can simultaneously be remarkably sensitive to subtle sequence changes – where a few mutations may either abolish or unexpectedly enhance binding. This paradox underscores the need for comprehensive characterization: the antibody fitness landscape is highly nuanced, and we can capture these complexities through systematically profiling the space with AlphaSeq.

What Makes Our Dataset Unique

This dataset is distinct in several ways:

1. Multimodal: Each paratope–epitope pair is annotated with sequence, parent structure (if on-target), and quantitative binding affinity.

2. Captures negative space: Off-target interactions are explicitly measured, revealing specificity as a distribution rather than a binary.

3. Reduced noise: Thousands of interactions measured in a robust assay with consistent internal standards eliminates much of the hidden heterogeneity present in public datasets.

4. Rigorous evaluation splits: With both structural and functional labels, we can define leakage-free splits by binder, by target, by complex, and by epitope class.

Shifting the Paradigm to “Complete” Multimodal Data

“Complete” multimodal data requires broader diversity across targets, epitopes, and scaffolds. Ground truth crystal structures provide a detailed view into the interaction between a paratope-epitope relationship, but only define “positive” strong binding interactions. To develop robust models, we will need these models to accurately pinpoint the difference between on-target versus off-target interactions, thus enable our models to learn what constitutes to a poor PPI interaction. We are expanding along these axes, including generating synthetic complexes validated in vitro. The goal is not just volume but coverage of biologically meaningful variation.

By providing models with comprehensive multimodal antibody-antigen data, we shift the learning problem from detecting weak correlations to learning causal constraints implied by the joint distribution of sequence, structure, and function. We believe this shift will yield models that are more sample-efficient, interpretable, and actionable.

Acknowledgements

Aditya Agarwal performed the bulk of the experimental design for these datasets. David Noble, Natasha Murakowska, Joseph Harman, Kerry McGowan, Nick Altieri, Adrian Lange, and Drew Duglan contributed to this work and text. Mackenzie Goodwin, Shyam Gandhi, Kenny Herbst, Juliana Barrett, Charles Lin and Mimi McMurray contributed to the lab experiments.

References

1. Schneider, Constantin, Matthew I J Raybould, and Charlotte M Deane. SAbDab in the Age of Biotherapeutics: Updates Including SAbDab-Nano, the Nanobody Structure Tracker. Nucleic Acids Research 50, no. D1 (2022): D1368–72. https://doi.org/10.1093/nar/gkab1050.

2. Matsunaga, R. & Tsumoto, K. Accelerating antibody discovery and optimization with high-throughput experimentation and machine learning. J. Biomed. Sci. 32, 46 (2025). https://doi.org/10.1186/s12929-025-01141-x

3. NaturalAntibody. AbDesign Database — Database of point mutants of antibodies with associated structures reveals poor generalization of binding predictions from machine learning models. (2025). Available at: https://www.biorxiv.org/content/10.1101/2025.06.09.658639v1

4. Jain, T. et al. Biophysical properties of the clinical-stage antibody landscape. Proc. Natl Acad. Sci. USA 114, 944–949 (2017). https://doi.org/10.1073/pnas.1616408114

5. Zhao, X. et al. Benchmark for Antibody Binding Affinity Maturation and Design. arXiv 2506.04235 (2025). https://doi.org/10.48550/arXiv.2506.04235

6. Chungyoun, M., Ruffolo, J. A. & Gray, J. J. FLAb: Benchmarking deep learning methods for antibody fitness prediction. bioRxiv (2024). https://doi.org/10.1101/2024.01.13.575504

7. Hummer, Alissa M., Constantin Schneider, Lewis Chinery, and Charlotte M. Deane. Investigating the Volume and Diversity of Data Needed for Generalizable Antibody–Antigen ΔΔG Prediction. Nature Computational Science, Nature Publishing Group, July 8, 2025, 1–13. https://doi.org/10.1038/s43588-025-00823-8.

8. ProteinBase by Adaptyv: https://www.adaptyvbio.com/blog/proteinbase/

9. Tobias H. Olsen, Fergus Boyles, Charlotte M Deane. Observed Antibody Space: A Diverse Database of Cleaned, Annotated, and Translated Unpaired and Paired Antibody Sequences. Protein Science. October 29, 2021. https://pmc.ncbi.nlm.nih.gov/articles/PMC8740823/

10. Younger, David, Stephanie Berger, David Baker, and Eric Klavins. High-Throughput Characterization of Protein–Protein Interactions by Reprogramming Yeast Mating. Proceedings of the National Academy of Sciences 114, no. 46 (2017): 12166–71. https://doi.org/10.1073/pnas.1705867114

Subscribe now