AI is showing up in the part of a clinical trial that draws the most regulatory scrutiny: the analysis itself, where a model’s prediction becomes part of the treatment effect estimate. It’s a shift Unlearn helped set in motion. For the teams designing those trials, it’s critical to understand which use cases will hold up when a regulator asks them to defend the design.

That question has a clearer answer than most sponsors expect. And the place to start is a framework the FDA has already published.

A common framework for an uncommon problem

In the FDA’s guidance document, Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products, the agency lays out a seven-step process for assessing whether a model is credible enough for the job it’s being asked to do. The guidance was written for safety, efficacy, and quality decisions, but the logic travels well beyond that scope. It asks you to define the specific question the model is answering, the exact context in which it’s used, and then to weigh risk along two axes: how much the model influences the answer, and how serious the consequences are if that answer is wrong.

That second pairing is what makes the framework useful. A model can exert significant influence on a result and still be low risk if being wrong costs little. Another can play a smaller role and still demand rigorous validation, because the cost of error hurts the credibility of a registrational trial. Risk isn’t a property of “using AI.” It’s a property of a specific use, in a specific trial, answering a specific question.

Three uses, three different risk profiles

In a recent whitepaper, our team applies this framework to three uses of AI in trial analyses that we’ve worked through in real trials. Each one puts a model’s predictions directly into the treatment effect estimate, and each lands in a very different place on the risk spectrum.

Some sit on firm ground. Adding power to an RCT through prognostic covariate adjustment produces a valid treatment effect estimate whether or not the model is prognostic; a useful model tightens the estimate, and a weak one costs you almost nothing. It’s a method the EMA has qualified and that the FDA has supported. Using that same approach to prospectively reduce sample size raises the stakes a step, because now the design depends on the power the model is expected to deliver, which makes validation against relevant data essential.

Others carry risk that needs to be characterized carefully. Using model-based comparators to support a single-arm study, in which predicted control outcomes serve as a control group, offers real benefit to patients but no guarantee of unbiased estimation. That doesn’t rule it out. It raises the bar for transparency and validation, and it’s exactly the kind of use the credibility framework was built to interrogate before a trial begins, not after.

What you can tell before the trial starts

Our whitepaper aims to give clinical and biostatistics teams a way to assess its use against what regulators will expect to see, and to flag where the path is still being written. Model-based synthetic controls, for instance, hold a real advantage over data-based comparators, since a model can be validated prospectively, but their regulatory path is still forming.

If your team is weighing where AI fits in an upcoming trial, this is a practical place to start: which applications are defensible today, which need more evidence, and how to tell the difference before you commit to a design.

[Download the whitepaper]

The FDA recently issued a request for information on its AI-Enabled Optimization of Early-Phase Clinical Trials Pilot Program, part of a broader push toward real-time clinical trials. We filed comments on the public docket, authored by our Chief Scientific Officer, Jonathan Walsh. Here are our main takeaways:

Two use cases for how AI can sharpen early-phase trial decisions

The first use case begins with digital twins as efficacy benchmarks. A digital twin is an AI-generated prediction of a trial participant’s comprehensive clinical outcomes on control or standard of care. Digital twins are generated for each enrolled participant from their unique baseline data; thus, the resulting benchmark reflects the actual study population rather than a historical average. It is available in near real time, and because it sits alongside the pre-specified analysis rather than replacing it, it gives reviewers and sponsors a sharper reference for interpreting an early signal. It does not stand in for a randomized control, or for the confirmatory analysis when the science calls for one. Its role is narrower, and still valuable: it helps teams make more informed go/no-go decisions earlier.

The same predictions deliver a second benefit, one that bears on trial quality rather than efficacy. Once every participant has a predicted trajectory (their digital twin), the observed data can be compared against it, and when deviations cluster at a particular site or vendor, they often provide the first visible sign of an operational or data problem. Issues like these typically surface only at database lock. Measured against the model, they can appear weeks or months earlier, serving as an advisory flag that prompts a closer look.

The second use case scales the same idea up, from a twin of each patient to a twin of the entire trial. As data arrive from Phase 1 or early Phase 2, the model recalibrates against them and becomes a simulation engine for designing the next phase before the current one has closed. Questions that normally wait for a final readout can be addressed in silico: which population to carry forward, which endpoint to use, how long to follow patients, and how large the next study needs to be. Because the model draws on both the historical data and the trial as it stands, those answers reflect the population and treatment behavior actually observed so far. The result is a shorter gap between phases, which is the central aim.

AI Adoption is here. Doing it well is a choice.

AI is already being adopted across clinical development, and that isn’t likely to change. The real question is no longer whether to use it, but how to use it well, and answering that requires the FDA at the table from the beginning, not as a reviewer at the end. That is what makes the pilot’s design so important. If it moves too fast, it risks the trust the whole effort depends on. If it waits for the field to adopt these methods on its own, the delay could stretch into decades, with patients paying the cost. A good pilot avoids both traps: it sets clear rules up front for how trial data will be interpreted, and it meets sponsors at their level of technical experience instead of shutting out those still building it.

We have standing to say this because we have done it. Since 2017, we have taken our method for using AI-generated digital twins in clinical trials the full distance from research results to regulatory adoption, with EMA qualification in 2022 and supportive guidance from the FDA in 2023. Our position has held throughout: these tools give expert teams better evidence, and the teams still make the call. A focused pilot on digital-twin benchmarks and live trial simulation can deliver near-term gains in early-phase decisions while laying groundwork for the wider use of AI that outlasts the pilot itself. Our full comments are on the public docket, and if these questions are live in your own early-phase programs, we would welcome the conversation.

The economics of clinical development all point the same way. More than $100 billion spent a year, 9 in 10 programs failing, $535,000 for a single Phase III amendment, most trials running late. Our conversations with sponsors keep landing on the same source, and it sits upstream, in how high-consequence decisions get made and carried across a trial. Each call is made with care, but it passes through clinical, biostatistics, operations, regulatory, data science, CRO partners, and investigators, and somewhere in those handoffs the decision trail comes apart.

We call what accumulates in its absence Decision Debt: the unresolved trade-offs, lost rationale, scattered assumptions, and delayed signals that pile up every time a decision leaves its context behind. It comes due later, as protocol amendments, delayed timelines, and readouts where the reasoning has to be reconstructed after the fact Our new whitepaper, The Scientific Intelligence Layer for Connected Clinical Trial Decisions, traces where Decision Debt accrues and what changes when the work stays connected. A few of those places:

Planning: Stress-Test the Design Before It Hardens

Population, endpoint, sample size, control, feasibility, and power are interdependent choices, yet they get settled in separate tools by separate teams. In an early-phase ALS study, applying the sponsor’s exact eligibility criteria to our historical dataset enabled the team to weigh competing designs against the same patients before enrollment opened. One path was materially leaner than the protocol on the table at equal power, and adding digital twins widened the gap further. At roughly $250,000 per patient, the difference was not academic.

Monitoring: Catch Drift Before It Reaches the Dataset

Every trial runs on assumptions: that patients progress as the protocol expects, that the control arm tracks its historical reference, that raters score the same in month nine as in month one. Any of these can drift without tripping a data-quality check, because the values are still valid; they just no longer describe the trial you designed. Retrospectively tested on the ADCS DHA Alzheimer’s trial, our approach detected anomalies on a core cognitive subtest early in enrollment, well before the trial’s own safety review surfaced them.

Analysis: Match the Method to the Design

Analysis is where the earlier choices come due, and for AI-supported methods the credibility has to be built into the data, model, and documentation well before database lock. This is where Unlearn’s regulatory footing with the EMA and FDA matters. What earning that footing actually requires is the paper’s through-line.

When planning, monitoring, and analysis stay connected, the result is Decision Compounding: A sharper planning decision sets better expectations for monitoring; what monitoring surfaces makes the analysis easier to defend; what the analysis shows shapes the next protocol. Regulatory reasoning stops being buried history and becomes context the next team can reuse. And because the digital twins, harmonized data, and rationale carry over across programs and stages, the next trial starts from what the last one learned. It is the structural answer to Decision Debt, and it is how sponsors get rigorous trials to patients sooner.

Read the full whitepaper

In a single-arm early-phase ALS study, the sponsor gets one read. There’s no control group to anchor against, and no way to re-run the study if the data come back hard to interpret. Disease heterogeneity and steady functional decline blur the signal. When every participant is declining at a different rate, separating a treatment effect from natural variation is hard. How a sponsor designs the study from the start to handle this uncertainty determines what its data can ultimately say.

That’s why SOLA Biosciences chose to structure its collaboration with Unlearn as a staged engagement for SOL-257, its long-term Phase 1/2 study of an investigational gene therapy designed to address a core pathological driver of ALS. The engagement spans three stages: data-driven trial planning, regulatory support across the Pre-IND and IND process, and digital twins as participant-level external comparators through the Phase 1/2 study and long-term follow-up.

In the partnership announcement, Keizo Koya, Founder and CEO of SOLA, framed the value this way:

“SOL-257 is designed to address a core pathological driver of ALS. In this Phase 1/2 study, our objective is to generate data that meaningfully informs downstream development decisions. Incorporating AI-generated digital twins strengthens the scientific rigor of our study design and supports disciplined, data-driven development decisions as we advance SOL-257.”

What this buys SOLA is interpretability it cannot add later. Each enrolled participant will be paired with their individual digital twin – an AI-generated prediction of their own control outcomes across multiple endpoints, generated from their baseline data using Unlearn’s ALS Digital Twin Generator. The study compares each trial participant’s actual trajectory against their predicted one (their digital twin), not against a population average. In a heterogeneous, progressive disease, that individual-level anchor is the difference between a result the team can defend and one it can’t

Trial planning. Ahead of protocol finalization, SOLA’s clinical team worked through precedent across ALS trials, design assumptions, and how the comparator strategy fits with inclusion criteria, endpoint selection, and the statistical analysis plan. Those decision propagate to every subsequent analysis.

Regulatory support. The digital twin comparator strategy is built into the regulatory plan from the Pre-IND stage onward. The digital twins work within the framework of, and do not replace, established statistical and clinical analyses; they run alongside standard methodologies, consistent with the FDA’s draft guidance on the use of artificial intelligence to support regulatory decision-making for drug and biological products.

Trial execution and long-term follow-up. The ALS Digital Twin Generator is trained on more than 13,600 ALS participants from RCTs and observational studies, including NEALS, PRO-ACT, and PRO-ACE, and is independently validated. Once the Phase 1/2 study is underway, the same paired-comparator approach carries through long-term follow-up, where natural-history data is typically sparse and individual-level comparators are hardest to source and most valuable.

Our research team has recently outlined the statistical framework for using digital twins as synthetic control arms in single-arm trials, with worked examples in ALS and Huntington’s disease. The methodology underlying SOL-257 is part of that broader work.

SOL-257 has not enrolled its first patient. But the decisions that will determine what its data can tell us have already been made, in the design.

When ProJenX and our team presented at the ALS Drug Development Summit last year, we focused on a problem sponsors across early-stage ALS development keep running into: how to generate interpretable evidence from small, often open-label single-arm trials where no concurrent placebo arm is available. Digital twins are designed to help with this directly, by serving as a patient-level comparator for each participant: an individualized prediction of how their control trajectory would unfold without treatment, generated from their baseline data alone.

That broader challenge intersects with three more specific questions sponsors are weighing in early-stage ALS trial design today: how to stratify within the heterogeneous sporadic ALS population, how to extract reliable readouts from trials sized and timed to match patient urgency, and how to corroborate existing and novel late-phase endpoints in trials that are often too small to power them. Each comes back to the same underlying problem: how to read meaningful signal from constrained, often open-label data.

These questions aren’t unique to any one program. They’re part of how the field is thinking about early-stage ALS development right now.

One of those programs is PRO-101, ProJenX’s Phase 1 study of prosetin, a brain-penetrant MAP4 kinase inhibitor in development for the treatment of ALS. ProJenX’s team needs findings from this small, open-label Phase 1 that are rigorous enough to inform Phase 2 decisions.

ProJenX’s team addressed this using Unlearn’s digital twins, generated by a disease-specific ML model trained on more than 13,600 ALS participants from RCTs and observational studies, including NEALS, PRO-ACT, PRO-ACE, and APST Research. The analyses are intended to support the clinical team’s interpretation of the data, not to replace the statistical and clinical judgment they bring to it. They run alongside standard methodologies and are consistent with regulatory expectations for external comparators.

What this gives ProJenX’s team is a way to answer specific questions they couldn’t answer otherwise.

Is there a subgroup of participants who respond differently to prosetin? Because each participant is paired with their own digital twin, the team can compute an individual treatment effect for that participant: how their actual outcomes diverge from the trajectory predicted by their twin. Patterns across those individual effects surface candidate subgroups, even in a small single-arm study.

Which endpoints carry the strongest signal for Phase 2? Digital twins predict multiple endpoints for each participant in the same dataset. Comparing observed outcomes to predicted outcomes across endpoints lets the team see which endpoints show the strongest divergence relative to their noise, indicating which endpoints are most likely to power a Phase 2 study.

How do you read treatment effects in an open-label setting? In an open-label trial, comparing a treated cohort against an external expected outcome leaves the team with treatment effects that are difficult to interpret. Pairing each participant with their own digital twin replaces that with a per-participant comparison: each participant’s actual outcome against their own predicted untreated trajectory.

In an illustrative analysis we presented with ProJenX at the ALS Drug Development Summit, paired digital-twin comparators made treatment effects readable at both the cohort and subject level in a 35-participant open-label dataset, with no concurrent placebo arm and no external matching dataset required.

Erin Fleming, COO of ProJenX, framed the value of this approach in her own words at the Summit:

“Digital twins help us make the most of the data we’re getting from here. In this open-label study, digital twins provide built-in placebo controls for each participant. They have a lot of key advantages over propensity score matching or other natural history controls that allow us to have more confidence in the data we’re taking out of that person-intraperson comparison.”

Our research team has recently outlined a framework for using digital twins as synthetic control arms in single-arm trials, with worked examples in ALS and Huntington’s disease. The methodology underlying PRO-101 is part of that broader work.

A year or two ago, digital twins as patient-level comparators in single-arm rare-disease trials would have been an exception. They’re increasingly part of the toolkit for generating interpretable evidence in early-stage single-arm trials. PRO-101 is one of the trials helping make that real.

In the United States, someone is either diagnosed with or dies from amyotrophic lateral sclerosis (ALS) every 90 minutes. The disease is universally fatal, with a median survival of two to three years after diagnosis. That urgency complicates trial design: in rapidly progressing, fatal conditions, ethical concerns around placebo use and prolonged enrollment timelines often make placebo-controlled randomized trials difficult to execute in early-phase development. Sponsors instead turn to single-arm study designs in which all patients receive the investigational therapy.

Case in point, in February of this year, VectorY Therapeutics announced that the first participant had been dosed in its single-arm ALS study. Phase 1/2 PIONEER-ALS is a trial evaluating VTx-002, a first-in-class vectorized antibody targeting TDP-43 pathology in people with ALS. PIONEER-ALS includes only 12 patients, so every signal needs to carry its own weight. That’s precisely why VectorY’s team chose to partner with Unlearn.

As VectorY’s Chief Medical Officer, Olga Uspenskaya-Cadoz, M.D., Ph.D., put it:

“In the PIONEER-ALS study, we are focused on generating high-quality safety and biomarker signal data for VTx-002…integrating Unlearn’s patient-level digital twin technology into our prespecified exploratory analyses will help strengthen evidence generation from a single-arm design, with the aim to support more confident development decisions, disease progression modeling, and reduce timelines and patient burden.”

Each of the 12 enrolled participants in PIONEER-ALS is paired with their digital twin, an individualized prediction of that patient’s control outcomes generated from their baseline data. Digital twins are generated by an advanced disease-specific ML model trained on more than 13,600 ALS participants from RCTs and observational studies, including NEALS, PRO-ACT, PRO-ACE, and APST Research.

Digital twin predictions are generated across composite scales (ALSFRS-R, ALSSQOL, ALS-CBS), labs (including plasma neurofilament light), and vitals over continuous time. These cover many of the trajectories that PIONEER-ALS is tracking, including slow vital capacity and survival. The model is independently validated, and the analyses run alongside standard statistical and clinical methodologies, consistent with regulatory expectations for external comparators.

Each participant in PIONEER-ALS serves as their own comparator. Because the digital twin is generated from that participant’s baseline data alone, the analysis doesn’t need to reuse records from other participants, define a matching cohort in advance, or rely on whether an external dataset is sufficiently large for direct matching to VectorY’s specific protocol. The analyses support the clinical team’s interpretation of the data; they don’t replace the statistical and clinical judgment the team is bringing to it.

Unlearn’s research team has recently described the statistical framework for using digital twins as synthetic control arms in single-arm trials, with worked case studies in ALS and Huntington’s disease. PIONEER-ALS applies that approach in practice.

We’re proud to partner with VectorY’s clinical team on this work, now underway across sites in the U.S., Europe, and the U.K.

The first three posts in this four-part series examined a problem that runs through oncology trial design: the evidence available to predict outcomes pertains to broader populations and older standards of care, rather than to the specific biomarker-defined cohorts and rapidly evolving treatment contexts that today’s trials require. Published trial results provide reliable estimates, but only as population averages for the cohorts studied. Real-world data provide patient-level granularity but are observational, subject to confounding, and expensive to obtain. As described in Part 1, reconciling these two data sources is tricky and increasingly so as cohorts narrow and standards of care shift.

In Part 2, we showed how treating outcome prediction as a modeling problem rather than a data-matching problem can address this for trial design, and how precision trial simulation gives teams a way to stress-test assumptions before committing to a protocol.

But there is a related problem that may be even harder, and it affects decisions that happen both before and after a trial is designed. What do you do when the comparison you need has never been made?

Two regimens, no direct trial comparison

FOLFIRINOX and gemcitabine plus nab-paclitaxel are both established first-line treatments for advanced pancreatic cancer. Oncologists regularly choose between them, and the decision affects patients every day.

Both regimens entered practice through trials against gemcitabine monotherapy. The PRODIGE4/ACCORD11 trial showed FOLFIRINOX improved median overall survival to 11.1 months versus 6.8 months for gemcitabine. The MPACT trial showed that gemcitabine plus nab-paclitaxel improved the median survival to 8.5 months versus 6.7 months. But no randomized trial has ever directly compared FOLFIRINOX and gemcitabine plus nab-paclitaxel, and given the cost and commercial dynamics involved, one is unlikely to be conducted.

That leaves clinicians and sponsors relying on indirect comparisons —notably network meta-analysis (NMA) or RWD-based target trial emulation (TTE). The PRODIGE4 and MPACT trials enrolled different patient populations (PRODIGE4: younger, healthier), used different eligibility criteria, were conducted in different geographies (PRODIGE4 in France; MPACT internationally across North America, Eastern Europe, Russia, and Australia), and read out in different years. How much of the apparent difference between the two regimens is real, and how much reflects differences in the patients who were studied?

This isn’t solely an academic question either. The answer matters for setting treatment guidelines, making formulary decisions, and for sponsors positioning new agents against an existing standard of care, as well as for development teams choosing a comparator arm for a new trial.

Attempts at indirect comparison:

To assess the comparative effectiveness of these trials, researchers have applied Bayesian Network Meta-Analysis (NMA) to this pair of trials. Crucially, this technique relies on an assumption of transitivity —that the patients in each trial’s control arms are exchangeable with respect to any effect-modifying factors. The fact that the PRODIGE4 trial restricted patients to a younger age range and a healthier performance status suggests that this assumption is problematic. It would be better to incorporate some means of adjusting for the baseline distributional imbalances into the comparison.

Along these lines, researchers have applied the target trial emulation (TTE) framework to make head-to-head comparisons using real-world data. A recent TTE using population-level data from Alberta, Canada (Boyne et al., Annals of Epidemiology, 2023), specified a hypothetical target trial protocol and emulated it using linked administrative health records. The study found that initiation of FOLFIRINOX was associated with a median overall survival of 8.3 months, compared with 5.1 months for gemcitabine plus nab-paclitaxel, with a mortality hazard ratio of 0.78. By design, TTE adjusts for measured confounders such as age and performance status, addressing the transitivity concern that limits the use of NMA.

The limitation lies elsewhere. The Boyne study was designed to compare outcomes as observed in routine clinical practice rather than to emulate the eligibility criteria of the original trials, and the cohort was drawn from a single regional healthcare system in Alberta. Of the 1,192 patients identified, 590 were excluded for missing laboratory data alone, and only 407 patients ultimately met the study’s eligibility criteria. That is a reasonable design choice for its stated purpose, but it means the resulting estimates are not directly substitutable for a trial-versus-trial comparison. More broadly, even a rigorously executed TTE is time- and resource-intensive and does not incorporate the summary results of the randomized trials whose comparisons the sponsor actually cares about.

A different way to make the comparison

The two attempts at comparison mentioned above are problematic for complementary reasons. We suggest an alternative approach that harnesses the strengths of each. Our trial-calibrated modeling approach enables this approach.

In the previous posts in this series, we described how a generative model that is calibrated to published trial results can generate patient-level outcome predictions that are sensitive to patient-level prognostic/effect-modifying factors while still respecting gold-standard trial results. We validated this in non-small cell lung cancer and metastatic colorectal cancer, predicting control arm outcomes in settings where direct historical matches were sparse or unavailable in the patient-level training data.

The same mechanism can be applied to provide a simulation of a head-to-head clinical trial between two therapies. The model provides a patient-level simulation for each trial’s treatment arm; this means that one has a patient-level synthetic dataset for each line of therapy that recapitulates the aggregate tables in the trial publications, from baseline distribution tables to overall and subgroup outcomes. From here one can adjust one simulated arm’s baseline distribution to the other to obtain a hypothetical head to head comparison.

When this procedure is applied to the PRODIGE4 vs MPACT scenario, we obtain a correction to the naive comparison between FOLFIRINOX and gem + nab-pac in first-line metastatic pancreatic cancer. The correction ultimately results from the key baseline mismatches between the two cohorts: an older cohort in MPACT that has a prognostically worse performance status range. The effect is an evident softening of the 1-year and 2-year RMST difference between the two therapies, and a more conservative hazard ratio of .853 with 95% ci (0.736, 1.073). This compares to a hazard ratio of 0.79 (0.59, 1.05) derived from a network meta-analysis conducted previously.

Fig: Simulated 2-year OS for the MPACT trial, if it were run on a population matching the PRODIGE4 trial.

The result is not a replacement for a randomized trial, but it is a rigorous comparison between competing therapies that accounts for prognostic/effect-modifying factors and privileges the highest-quality summary evidence available. It is also more cost-effective and efficient than running a new trial or performing a large-scale RWD-based analysis; it can deliver answers efficiently by leveraging the broad benefits of transfer learning and using RWD in a manner constrained by RCT results. That speed matters in an indication where standards of care are shifting constantly. In some cases, a regimen’s window of clinical relevance is short enough that the comparison only matters now, and a modeling approach is the only realistic way to inform the decision in time.

Where this matters for development teams

The pancreatic cancer example is a clear illustration, but the underlying problem is common across oncology. Wherever a sponsor needs to understand how two treatments might compare in a trial setting, and no head-to-head trial exists, the same gap appears.

Consider a development team designing a trial for a new therapeutic agent. They need to choose a comparator arm design from among several options for SoC therapies and baseline cohort features. Ultimately, they need to compare expected outcomes under those choices and choose the design that maximizes the success parameters for the trial. That is a comparative effectiveness question, and it must be answered before the protocol is finalized, not after.

Or consider a medical affairs team preparing a health economics submission for a new therapy in a biomarker-defined subgroup. The submission requires a comparison against the current standard of care. But published trial data for that standard of care reflect an unselected population, not the subgroup the new therapy targets. The comparison the team needs has never been made at the resolution they require.

A calibrated modeling approach does not eliminate the need for judgment in either case. But it does provide a structured, reproducible framework for generating those comparisons, rooted in the best available randomized evidence and tailored to the specific patient population in question.

If you are facing a comparative effectiveness question in your program, reach out to our team.

Before an oncology protocol is finalized, most important decisions have already been made based on untested assumptions.

What happens if the eligibility criteria narrow from all KRAS mutations to KRAS G12C only? Does that improve the chance of seeing a signal, or just make the study harder to enroll? How much confidence does a team really have in the event-rate and power assumptions built into the design?

These are the questions that determine whether a trial enrolls smoothly or turns into an expensive amendment exercise later. They’re also the questions governance will likely push back on. A team that has run six scenarios will have more defensible results.

In our first and second blog posts in this series, we described why these assumptions are harder to get right and how a trial-calibrated modeling approach can help. The next question is the practical one: what does that actually change in the study-design workflow?

It means a team can ask “what happens if we narrow to KRAS G12C?” or “what does the control arm look like under the current standard of care?” and get a calibrated answer before the protocol is finalized.

What trial simulations for oncology mean in practice

Here’s how it works. The inputs are the same ones a real protocol depends on: inclusion and exclusion criteria, baseline characteristics, biomarker definitions, line of therapy, and treatment regimen. Once those are specified, the model can generate a cohort that reflects the trial a team is actually trying to run and estimate how outcomes may change as those design choices change.

This is the practical output of the FRESH modeling approach described in part two of this series. Because the model is calibrated to published trial results rather than built from a dedicated data study, the simulations are ready to go quickly. Teams do not need to commission a large real-world dataset to test different cohort definitions or comparators. The model generates calibrated, patient-level predictions for each scenario.

That gives teams more room to explore during the design phase.

Questions you can answer before protocol finalization

1. What should we expect from the control arm?

This is often the hardest question to answer cleanly. Standards of care evolve quickly, and published trial results remain the most credible source of evidence, but they are reported for specific historical populations and treatment contexts, not for the exact cohort a new study hopes to enroll.

Because the FRESH modeling approach calibrates predictions to published trial results and operates at the patient level, teams can estimate expected control arm outcomes for the specific cohort they are designing around, including narrow biomarker-defined populations where comparable historical datasets may not exist or be accessible.

2. How much does the design depend on the cohort definition?

Eligibility criteria often look like a clinical choice, a regulatory choice, or a feasibility choice. In practice, they are all three at once.

Narrowing a cohort may strengthen the biological rationale for the study. It may also change the expected event rate, alter the standard-of-care benchmark, affect the sample size, and make enrollment harder.

Take a familiar kind of decision: should a study include all KRAS-mutant patients, or restrict enrollment to KRAS G12C? Or in first-line NSCLC, should a design use one PD-L1 threshold or another?

Those choices shape the trial in ways that are easy to underestimate when only one or two scenarios get examined. Trial simulations make it possible to compare those scenarios directly while design flexibility still exists. Because calibrated predictions can be generated for each scenario without restarting the evidence assembly process, a team can examine several cohort definitions in the time it would previously have taken to commission one.

3. Which subgroup is worth prioritizing?

Oncology teams rarely choose between a good option and a bad one. More often, they are choosing among several plausible paths, each supported by some evidence and some degree of judgment. One subgroup may look biologically cleaner. Another may be easier to recruit. A third may better match commercial or program-level priorities. Those tradeoffs are real, and the supporting evidence is often uneven.

Simulations give teams a way to compare those subgroup strategies using a shared analytical framework. Instead of relying on scattered subgroup readouts, broad historical averages, etc. they can ask how outcomes may differ across several candidate populations before they commit to one. In validation against the POSEIDON trial, the model accurately reproduced differential survival across PD-L1 expression, histology, KRAS, STK11, and KEAP1 status: the same subgroup dynamics that drive these prioritization decisions.

A different relationship with the design process

The operational shift is fewer one-off analyses and more structured exploration.

When a team has access to a large, representative real-world dataset, the preferred quantitative path is clear: acquire the data, shape it into a cohort, and analyze design scenarios directly. In practice, though, most teams face real constraints. The available data may be too small, too expensive, riddled with gaps, or simply not accessible for the population in question, or some combination thereof. The result is that the standard quantitative process often cannot be done well, and teams end up making critical design decisions with less evidence than they’d like.

Calibrated simulations change this. The model generates patient-level cohorts on its own, so a team does not need to commission a separate data study every time they want to test something. That opens up questions that would otherwise be tabled: What if we narrowed eligibility? What happens under a different standard-of-care assumption? What does the effect estimate look like in an older population? Most programs never run those scenarios because the data work alone would take months.

Before the protocol is finalized, a team can now work through far more of those questions inside the model. If a comparator shifts, they can rerun the analysis in hours, (not weeks.) The design conversation improves because the team has actual evidence behind their choices instead of educated guesses about what the data would have shown.

More confidence, earlier

It’s unlikely we’ll ever completely eliminate uncertainty from oncology trial design. But we can discover the weak spots in assumptions during the design phase instead of after it’s already running.

The validation results described in part two of this series bear this out: the model predicted the BREAKWATER control arm without ever training on BREAKWATER data, reproduced subgroup-level survival in the POSEIDON trial, and interpolated accurately to the LEAP-006 regimen using only the KEYNOTE chemotherapy arms as calibration inputs. This approach can be applied to the specific populations, regimens, and cohort definitions a team is actually considering.

If you’d like to explore how Unlearn’s approach could apply to a specific program or if you have a design question, reach out to our team.

The decision-making gap in oncology trial design

If outcomes could be predicted for a specific population during the design phase for a trial, many of the hardest decisions in oncology development would become more tractable.

• What outcomes should be expected for the population defined by the protocol?

• How do those expectations change as eligibility criteria shift?

• Which comparator is most appropriate for a given setting?

• What is the likelihood that a trial will meet its endpoint?

These are the decisions that help sponsors determine if a study is feasible and likely to succeed

In Part 1, we described why answering these questions has become increasingly difficult. Patient populations are fragmenting into narrower biomarker-defined subgroups. Standards of care are evolving faster than trials can read out. Together, these two forces create a third problem: the evidence required to make design assumptions is distributed across sources that were not built to be used together.

For sponsors, this is not an abstract challenge. These decisions must still be made, often with incomplete and conflicting evidence.

When data matching breaks down

The BREAKWATER Phase III trial illustrates what this looks like in practice. The standard-of-care arm enrolled metastatic CRC patients with BRAF V600E mutations, a subgroup representing only 8 to 12 percent of the mCRC population. When we searched our data for available real-world and trial datasets of patients who fully matched BREAKWATER’s eligibility criteria, we found five patients. A data-matching approach would not provide a reliable basis for estimating outcomes in this setting. Yet decisions about trial design still need to be made.

This is not an isolated case. As oncology trials target increasingly specific populations, the number of directly comparable patients declines sharply. Real-world data remains an important source of patient-level detail, but assembling usable cohorts can cost millions of dollars and requires significant time to collect, and the resulting data is observational and subject to confounding. Published clinical trials provide reliable estimates of outcomes, but only for the populations that were studied. Applying those results to a new trial with different eligibility criteria requires assumptions that are difficult to validate.

The result is a gap between the existing evidence and the decisions that must be made. Study teams are asked to commit to trial designs, sample sizes, and comparators without a clear, data-driven estimate of the expected outcomes in the population they intend to study.

The solution is to treat outcome prediction as a modeling problem rather than a data-matching problem.

From data matching to modeling

To address this, we developed a pan-cancer foundation model trained on detailed clinical and genomic data from approximately 300,000 tumor biopsies. The model learns the joint distribution of clinical variables, genomic alterations, treatment histories, and outcomes across indications, enabling it to generate patient-level cohorts matching specified eligibility criteria and predict outcomes under a given treatment regimen.

Patient-level data alone is not sufficient for clinical decision-making; predictions must be consistent with what has been observed in randomized trials. We therefore apply a calibration procedure that anchors the model’s predictions to population-level results from published clinical trials, while preserving the patient-level relationships learned from the data. We refer to this combined approach as Fusion of Recent Evidence and Subject Histories (FRESH) modeling. Full technical details, including validation against published trial results and out-of-sample predictions across oncology indications, are described in our latest whitepaper.

Trial-calibrated digital twins

The output of this process is a set of trial-calibrated digital twins: simulated individuals whose predicted disease courses under a specified standard-of-care regimen reflect what has been observed in real trials while retaining the heterogeneity of real-world data. This means the model retains the granularity needed to ask patient-level questions about narrow subgroups, but its answers are anchored to gold-standard randomized evidence at the cohort level.

The BREAKWATER result illustrates this directly. We predicted the BREAKWATER control arm’s overall survival curve, matching the observed results at the median and at 6-, 12-, and 18-month time points, without using BREAKWATER data. The foundation model generalized from a broader pool of nearly 800 patients with BRAF V600E mutations across indications and regimens, and calibrated to prior unselected mCRC trials. A perfect historical match was not required. What was required was a model.

Answer hard trial design decisions during the design phase

Once you have calibrated, patient-level predictions, key aspects of trial design can be treated as a series of concrete “what if” questions before committing to a protocol.

In precision trial simulation, teams can define a trial cohort based on inclusion and exclusion criteria, biomarker profile, line of therapy, and standard-of-care regimen, and then ask: what outcomes should we expect? How do event rates change if we tighten from all KRAS mutants to KRAS G12C only, or enrich for higher PD-L1 expression? What happens to sample size requirements if we switch comparators? Instead of requiring months of bespoke evidence synthesis for each question, these become on-demand queries against a calibrated model.

In single-arm studies, which are common in early-phase oncology and in narrow biomarker-defined populations, sponsors often lack a reliable comparator for go/no-go decisions. Trial-calibrated digital twins can provide a rigorous, calibrated benchmark to evaluate single-arm results against expected standard-of-care outcomes. In the near term, this supports internal decision-making by giving development teams a structured way to contextualize efficacy signals before committing to larger, more expensive randomized trials.

For comparative effectiveness, sponsors often need to understand how different standard-of-care regimens perform in a specific patient population to inform portfolio positioning and commercial strategy. The model can compare on-market therapies across clinically defined subgroups, including narrow biomarker-defined cohorts, without requiring a head-to-head trial.

Validated across indications, out of sample

We have validated this approach in non-small cell lung cancer and metastatic colorectal cancer. In NSCLC, predictions calibrated to published clinical trial results reproduce observed overall survival at the population level and across most key subgroups, including PD-L1 expression, histology, and genomic alterations such as STK11, KEAP1, and KRAS. In the POSEIDON trial, predicted survival curves aligned with observed outcomes at the population level and across these subgroups, despite calibration being performed only at the cohort level.

The BREAKWATER result described in the previous section represents a fully out-of-sample prediction across indications. These results demonstrate the methodology’s capabilities, not its coverage limits. The foundation model’s pre-training data already spans breast, pancreatic, and prostate cancer in addition to NSCLC and mCRC.

Reducing the data burden, not adding to it

These results highlight a limitation of traditional approaches. Data matching depends on finding patients who exactly meet a set of criteria. As those criteria become more specific, the number of usable patients declines, often to the point where estimates are unstable or unavailable. A modeling approach allows information to be shared across related patients, indications, and treatment contexts, enabling predictions in populations where direct matches are limited.

This also has implications for how data is used in development programs. Rather than requiring repeated, bespoke real-world data acquisition efforts for each new question, the model can generate calibrated predictions using existing data. Additional data collection can then be focused on areas of greatest incremental value.

In the next post, we will examine how this approach can be applied in practice to evaluate trial scenarios, compare biomarker strategies, and test assumptions before protocol finalization.

Unlearn was founded to eliminate trial and error in medicine, and Q1 was a quarter that showed just how far that mission can reach.

In our conversations with leading sponsors, one theme consistently kept coming up: the upstream work of trial planning is more painful (and costly!) than it needs to be. Planning is scattered across tools and teams, and every time an assumption shifts, the work starts over from scratch. This quarter, we launched a new product to fix it: TrialPioneer, a unified workspace that extends Unlearn’s approach earlier in the lifecycle, into the planning decisions that shape a trial long before the first patient is enrolled.

Three capabilities work together inside it. Scout handles precedent review, bringing together scientific and regulatory evidence from sources like PubMed, ClinicalTrials.gov, and drugs@FDA into a structured, shareable foundation. Hindsight enables teams to explore harmonized clinical and real-world datasets to validate assumptions about population characteristics, endpoint behavior, and benchmarks, grounding design decisions in empirical evidence rather than intuition. And SimLab lets teams evaluate scenarios across endpoints, eligibility criteria, and sample size, with reproducible outputs directly tied to their assumptions.

What I’m most excited about is what this enables before a protocol is ever finalized. That’s where design risk is highest, and where scattered assumptions and hard-to-reconstruct rationale lead to governance struggles and expensive amendments. TrialPioneer gives teams a foundation where every assumption is tied to evidence, so the rationale is built before it’s needed.

Partnerships Advancing Digital Twins Across Therapeutic Areas

This quarter also marked significant progress in our partnerships, with new collaborations that demonstrate the growing role of digital twins across clinical development.

Deepening Our Huntington’s Disease Models

We announced our use of data from CHDI Foundation, a nonprofit exclusively dedicated to developing therapeutics for Huntington’s disease. Through access to the global Enroll-HD clinical research platform, one of the richest longitudinal datasets available in neurodegeneration, we recently updated our HD DTG with the latest available data, enhancing the performance of patient-level control predictions for cUHDRS, TMS, and UHDRS. Better data means better models, and better models mean more reliable evidence for the sponsors designing trials in this space.

Expanding Our Presence in ALS

ALS remains one of the most urgent unmet needs in medicine, and this quarter, we announced two new partnerships that demonstrate how digital twins are uniquely suited to accelerate development for this disease.

We’re partnering with VectorY Therapeutics to support their PIONEER-ALS study, a Phase 1/2 trial evaluating a first-in-class vectorized antibody therapy. And we announced a partnership with SOLA Biosciences to support their Phase 1/2 study of SOL-257, an investigational gene therapy targeting a key pathological driver of ALS.

In both cases, the challenge is the same: traditional placebo-controlled designs are often impractical, yet single-arm studies leave sponsors without a reliable comparator. Digital twins fill that gap, providing an individualized benchmark for each enrolled patient, enabling sponsors to extract stronger, more interpretable signals from every participant. The result is more confident development decisions in a disease where every data point matters.

Bringing Our Approach to Oncology

We spend a lot of time talking to oncology teams, and one thing comes up constantly: predicting outcomes for the specific population you’re actually trying to enroll has become one of the hardest parts of trial design. Populations keep narrowing, standards of care keep shifting, and the tools most teams rely on weren’t built for that reality. (We wrote about this challenge here.)

This quarter, we took a major step toward solving those issues with a new pan-cancer foundation model, trained on clinical and genomic data from approximately 300,000 tumor biopsies. The model generates patient-level predictions even in the narrow, biomarker-defined subgroups where traditional data-matching approaches fall short. In validation against the BREAKWATER Phase III trial, where only five patients in available datasets matched the eligibility criteria, our predictions closely matched the observed outcomes without ever using BREAKWATER data.

We’ve published the full methodology and validation results in our whitepaper, “A Fresh Approach to Precision Oncology Decision-Making.” For sponsors navigating increasingly complex trial design decisions in oncology, we believe this work opens a new path forward.

Sharing Our Work with the Community

Our team was active on the conference circuit this quarter, presenting new research and connecting with sponsors across therapeutic areas. At ISCTM in Washington, D.C., we presented research on how digital twins can serve as external controls in non-randomized trials to reduce bias and lower variability in treatment effect estimates.

At AD/PD in Copenhagen, we shared results showing that digital twins of study participants delivered up to a 15% power boost across endpoints in an Alzheimer’s disease trial setting, and our co-founder and CSO, Jon Walsh, joined a panel on precision drug development and AI methods in trial design.

These presentations are part of our ongoing commitment to advancing the science openly and building trust with the clinical development community through peer-reviewed, evidence-based work.

Looking Ahead

This quarter showed us what happens when foundational science meets real-world application. With TrialPioneer, we’re giving leading sponsors the tools to make better design decisions, earlier and faster. With our expanding partnerships, we’re proving that digital twins belong at the center of modern clinical development, from early planning through trial execution. We’re excited about the significant momentum we’re carrying into Q2.

If you’d like to explore how Unlearn can support your clinical development program, reach out to our team.

Oncology drug development has always involved making decisions with incomplete information. What has changed over the past decade is not just the scale of that uncertainty but the stakes attached to it, driven by two converging forces: patient populations are fragmenting into ever-narrower biomarker-defined subgroups, and standards of care are evolving faster than trials can read out.

A therapy may be evaluated in patients with a specific mutation (for example, BRAF V600E or KRAS G12C), a PD-L1 expression threshold, and/or a particular treatment history and line of therapy. At the same time, standards of care evolve rapidly as new molecularly targeted therapies and modalities, immunotherapies, and antibody-drug conjugates enter clinical practice. By the time a Phase III trial reads out, the standard of care may have already shifted.

Together, these forces put development teams in a difficult situation. Expected outcomes shape nearly every aspect of trial design (i.e., sample size, endpoint selection, eligibility criteria). Yet the evidence available to estimate those outcomes was built for broader populations plus different and fragmented standards of care. Get the assumptions wrong, and the consequences are severe: an underpowered trial, an unnecessary protocol amendment, or years of follow-up spent on a study that was never going to succeed. Every one of those outcomes traces back to a decision made on faulty assumptions. And if outcome prediction is where the decision goes wrong, it follows that improving outcome prediction is where the most leverage lies.

What you really want is the granularity of patient-level data to evaluate specific design choices, combined with the recency of population-level trial findings, even for brand new treatments. Those two sources exist, but they were never designed to work together. Published clinical trial reports provide reliable outcome estimates (OS curves, PFS medians, event rates) but only as population averages for historical cohorts. Real-world datasets provide patient-level granularity, but they are observational, often confounded, and rarely aligned with the exact population and standard-of-care context a new trial will actually enroll.

Reconciling the two sources is slow, expensive, and increasingly difficult to scale as cohorts narrow…and no off-the-shelf solution exists to do it reliably.

The data problem in oncology trial design

For many development teams, pulling together the data needed to predict outcomes and support a single trial design can take months of work across clinical, statistical, and data science groups. It’s an approach that doesn’t scale when teams want to explore many possible cohorts, comparators, and endpoints.

Accessing real-world data in particular comes at a high cost. Acquiring datasets that reflect the relevant patient population and standard-of-care context can cost hundreds of thousands of dollars (or more) and in many cases those data are not commercially available at all. Even when the data can be obtained, cleaning, harmonizing, and analyzing them is time-consuming work that must be repeated from scratch every time a team wants to evaluate a new cohort, comparator, or endpoint assumption.

Why the margin for error is shrinking

A Phase III oncology trial may require hundreds of patients, dozens of sites, and years of follow-up, with total budgets often reaching tens of millions of dollars or more.

Regulatory trends are adding further pressure: sponsors are increasingly expected to have confirmatory trials already enrolling at the time accelerated approval is granted, which means committing to expensive, high-stakes designs earlier and with less time to stress-test assumptions.

As oncology populations fragment and therapies evolve, it is becoming clear that clinical development needs a new set of tools. Uncertainty, cost, feasibility, and regulatory pressure are all converging on the same bottleneck: the ability to predict outcomes reliably for a specific trial population before committing to a design. Relying on manual evidence synthesis alone is no longer enough, which is why many teams are beginning to treat outcome prediction itself as a modeling problem to be solved in a more systematic way.

In Part 2 of this series, we’ll walk through what it looks like to treat outcome prediction as a modeling problem (building on concepts explored in our oncology whitepaper, download here) and how bridging patient-level and population-level data opens up new possibilities for trial design.

We’re excited to announce the release of AD DTG 4.2, the latest update to our Alzheimer’s disease Digital Twin Generator. This release expands the biomarker data available to our model and enables early exploration of biomarkers as potential clinical outcomes.

Why Biomarkers Matter in AD Trials

Biomarkers are measurable biological molecules whose concentration shifts in response to disease processes or treatments. Unlike traditional lab values that monitor general physiology, such as electrolytes or liver enzymes, disease-specific biomarkers target molecular pathways tied to a specific condition.

In Alzheimer’s, novel biomarkers like phosphorylated tau-217 (p-tau217), amyloid beta 40 (Aβ40), and amyloid beta 42 (Aβ42) can signal neurodegenerative changes years before symptoms appear, making them valuable for detecting disease, tracking progression, and evaluating therapeutic response in clinical trials.​

Expanded Data, Expanded Possibilities

The updated data asset now includes or enriches several AD-relevant biomarkers, including: Aβ40, Aβ42, total tau (t-tau), phosphorylated tau-181 (p-tau181), p-tau217, glial fibrillary acidic protein (GFAP), and neurofilament light chain (NfL).

Incorporating these biomarkers into the data asset creates new opportunities to explore their role in inputs, outputs, or both when generating digital twins of study patients in future DTG releases. ​

An Early Exploration of ptau-217 as a Clinical Outcome

AD DTG 4.2 incorporates two biomarkers from the updated data asset, p-tau217 and GFAP.

P-tau217 is a sensitive and specific marker of Alzheimer’s pathology, correlating with amyloid accumulation, tau burden, brain atrophy, and physical degradation — and notably does not predict such changes in patients with other neurodegenerative disorders, making it highly specific to Alzheimer’s disease. This specificity makes it a compelling feature for modeling disease progression.

With this release, the model can now predict longitudinal changes in ptau-217 concentration as a clinical outcome, a capability we’re optimistic about as we continue to grow and validate the underlying dataset. This capability may help support broader biomarker modeling within the AD DTG framework for future model updates.

In addition, treating ptau-217 as the sole input to the model can provide prognostic information for other outcomes commonly measured in Alzheimer’s trials.

What’s Next

This biomarker work builds on the validated foundation of the AD DTG, which has already demonstrated meaningful impact in Alzheimer’s studies. In retrospective analyses, digital twins of study participants have supported sample size reductions of up to 15% and up to 33% in control arms using Unlearn’s EMA-qualified and FDA-supported method.

As we continue to grow and enrich our AD dataset, now trained on over 25,000 patient records spanning cognitively normal individuals through moderate Alzheimer’s disease, the model will be positioned to support increasingly robust biomarker modeling and stronger digital wins in future DTG releases.

Looking back at 2025, I’m reminded of a conversation from February that has stayed with me. A CEO pulled me aside at a conference and said, ‘Steve, when we began working together last year, I was excited, but I’ll admit, I wasn’t entirely sure what to expect. I kept those doubts to myself. But seeing Unlearn’s impact on our program, and how your work helped us make genuinely life-changing decisions for patients, was extraordinary. It proved to me that AI in clinical development truly moves the needle.”

That moment? That’s what this year was about.

Real Partners, Real Progress

We’re ending 2025 with the most customers in our history—a testament to our evolution from a forward-looking AI research company into one delivering real, measurable value in clinical development today. Sponsors are trusting us with their most important programs, and the conversations I’m having now are very different from even six months ago. They’re not asking “Does this work?” anymore. They’re asking “How fast can we implement this?” and “What else can we do together?”

One of the clearest signals of our progress is seeing sponsors return to work with us again—often expanding into additional programs and therapeutic areas. As teams experience the value of our AI-powered solutions, they start rethinking how they approach upcoming trials. That shift from skepticism to urgency tells me everything about where we’re headed.

The Evidence Is In (And It’s Good)

A standout moment this year came at AD/PD Vienna, where AbbVie shared a retrospective analysis showing how digital twins could reduce placebo arm size by 23% in Alzheimer’s disease. We also presented our own work in Parkinson’s disease, demonstrating a 38% reduction in control arm size for the same endpoint. These kinds of reductions directly translate into stronger evidence teams rely on to make high-stakes decisions.

Our Science team has been on fire this year—presenting at over twenty conferences, advancing our ML methods, and laying the groundwork for our expansion into oncology next year. When you watch a room of scientists lean in as our team presents, you know we’re building something that resonates.

Building the Team to Get Us There

If you want to know what I’m most proud of this year, it’s the team we’ve assembled.

We welcomed Krates Ng as CTO—someone who knows how to build systems that scale. Kwame Marfo came on as VP of Product whose 15 years in life sciences product leadership give us a powerful lens on what customers need. And Dr. Robert Lenz joined as senior advisor—when someone who’s run 130+ clinical studies across 50 countries tells you your approach is the right one, you pay attention.

2026: A Defining Year Ahead

After 30 years in this industry, you learn to recognize inflection points—the moments when meaningful change becomes possible. We’re entering one of those moments now.

Clinical development has never been more complex or more expensive. But with that complexity comes opportunity for teams willing to rethink how evidence is generated, how decisions are made, and how quickly new therapies can move forward. We’re not here to make trials 5% more efficient. We’re here to help teams design smarter studies, align more quickly, and make decisions with greater clarity and confidence.

Everything we built in 2025—the science, the technology, the team—is setting the stage for what comes next.

To everyone at Unlearn: Thank you for bringing your best every single day. Thank you for believing in this mission when many said it couldn’t be done. And thank you for making this the most exciting chapter of my career.

Let’s keep building.

Last month, I had the privilege of speaking at the FDA-CTTI Workshop on Artificial Intelligence in Drug & Biological Product Development. I joined Session 4, “Navigating the Future of AI in Drug Development,” alongside fellow panelists Dr. Jessilyn Dunn from Duke University and Ryan Hoshi from AbbVie, in a session moderated by Rebecca Nebel (PhRMA) and Gabriel Innes (FDA).

My talk, “Using AI for the Hard Problems,” focused on a necessary transition we need to make as an industry: the shift from human-centric to model-centric development.

For decades, drug development has been a biological, chemistry-driven field. It has also been deeply human-centric. We focused on training people, defining best practices for them to follow, and building complex systems. We wanted to understand every decision, often slowly and in ways that act against efficiency gains from software and AI.

But as I argued during my session, if we want to solve the hard problems—like finding safe, effective drugs for complex diseases or predicting climate change—we have to let the models do what they are good at. And we have to set up the business processes that allow that to happen.

In physics, we model the Large Hadron Collider—a $1 billion-a-year experiment—to understand fundamental truths about the universe. Clinical trials are a $40 billion-a-year experiment on people. The scale is massive, and relying solely on human-centric systems limits us. We need to move toward a model-centric approach where:

• Best practices ensure quality. We circumvent the “black box” fear by having rigorous rules for how models are developed and used.

• Performance rules. The attitude becomes “just use a model, it’s going to do a better job,” and interpretability becomes less relevant than accuracy.

• We use systems–cleaning, harmonizing, and using datasets; using software and AI–that make it easier to build and use models and integrate them into our decision-making processes.

But you can’t just flip a switch to model-centricity. You need proof points to overcome the high mistrust and lack of clear value demonstration that often plagues new tech.

I highlighted a specific use case we are championing at Unlearn: using AI models to predict participant outcomes in a clinical trial. By integrating these predictions into the analysis—creating what we call model-derived covariates—we can safely add statistical power and eventually reduce sample sizes. We hope using these tools becomes a de facto expectation because it is simply the better, safer way to run a trial.

One of the most encouraging moments from the discussion came during the closing remarks from Qi Liu, the Associate Director for Innovation & Partnership at the FDA’s CDER.

We often view regulators as the brakes on innovation, but in this workshop, they were helping steer the car. Qi Liu shared a perspective on digital twins that effectively validated that the agency is looking at these technologies not just as novelties, but as legitimate tools for modernizing clinical trials. Her comments underscored a critical shift: the FDA is open to advanced methodologies—provided we have the rigorous frameworks to assess them.

We left the workshop with a clear to-do list for both sides of the aisle:

• For Regulators: Continue defining the framework. As I mentioned in my talk, regulators are actually out in the lead in some ways. We need them to sharpen the boundaries for not yet versus work with us to evaluate and help define case studies that the whole community can use.

• For Sponsors: We need to acknowledge that drug development is fundamentally a computational science now. That means investing in data scientists and machine learning engineers—giving them a seat at the table, where they currently often don’t have one.

The transition to a model-centric future is inevitable. The question is no longer if AI will transform drug development, but how quickly we are willing to unlearn our old habits to make it happen.

Unlearn’s Machine Learning Scientist, Daniele Bertolini, recently partnered with the Multi-Regional Clinical Trial Center (MRCT Center) of Brigham and Women’s Hospital and Harvard to host a two-part webinar series on leveraging AI-generated digital twins of study participants and synthetic data in clinical development.

The sessions offered a comprehensive view of how Unlearn’s technology moves beyond theoretical promise to deliver actionable, regulatory-qualified benefits. Here are the core insights and highlights from the series.

In the first webinar, Daniele establishes why traditional RCTs are fundamentally inefficient: they operate on a “start-from-scratch” principle, systematically ignoring the wealth of historical data available on disease progression. Unlearn’s digital twins are probabilistic computational models that use a patient’s baseline data to predict their exact, individual disease progression under standard of care. This prediction provides a critical measure of uncertainty, allowing for highly precise modeling.

By incorporating the digital twin’s prediction as a “super covariate,” sponsors can effectively explain away a significant portion of the natural variance within the patient population. This leads directly to a more precise estimate of the treatment effect, resulting in powerful, practical clinical development outcomes: either reduced sample sizes or increased power in RCTs.

The second webinar expanded on the foundational concepts to showcase practical applications across the drug development lifecycle, emphasizing the quantitative benefits and regulatory environment. Daniele presents a retrospective analysis across multiple therapeutic areas (including ALS, Alzheimer’s, and Huntington’s) demonstrating how Unlearn’s EMA-qualified and FDA-aligned method for using digital twins in RCTs achieves a variance reduction in the treatment effect estimate of 10% to 20%, translating directly into a potential 20% to 30% reduction in the control arm without sacrificing statistical power.

The webinar then introduced solutions for challenging trial designs. For indications where randomization is unethical, such as rare diseases or pediatric trials, digital twins provide a robust solution by serving as a synthetic control arm, generating the necessary counterfactual outcome for every treated patient. The session also previewed advanced statistical methods and explored how synthetic data can be leveraged pre-trial to model the impact of different inclusion/exclusion criteria, trial duration, and endpoint choices, enabling faster and cheaper trial design optimization.

Stay tuned for details on the January panel discussion!

Quality data is at the heart of any high-performing AI model, but wrangling data into a modeling-ready state can be challenging and time-consuming. This is especially true for clinical data, which is often noisy, lacks standardization, or requires clinical expertise to assess its quality and relevancy. For clinical teams, these bottlenecks can delay trials and slow innovation.

At Unlearn, we’ve built the infrastructure, processes, and AI-based methods to transform this messy, incomplete, heterogeneous data into usable inputs for training and validating AI models.

We’ve collected de-identified, longitudinal patient data from over 1 million patients across more than 30 indications, partnering with research institutes, advocacy groups, academics, and commercial vendors to do so. Through harmonization—organizing, cleaning, and standardizing data—we’ve created a curated, robust dataset covering over 370k patients and more than 1 million patient-provider interactions—fueling 14 Digital Twin Generator (DTG) deployments across neurodegenerative, immunology, cardio-metabolic, and psychiatric indications.

It’s the kind of investment that could take years—and enormous resources—for a pharma company to build internally. By partnering with Unlearn, clinical teams bypass the “build vs. buy” dilemma and gain immediate access to validated, ready-for-use AI models. These models enable us to uncover patterns and answers that no single study or one-off analysis could provide on its own.

The Tools That Make Imperfect Clinical Data Perfect for Clinical Trials

Traditional Extract, Transform, Load (ETL) pipelines are designed for data with a consistent format and executed on a fixed schedule, which means they can afford to be rigid. But clinical data is anything but uniform. In fact, the sheer variety in structure, format, and quality of the data sources we need to harmonize necessitates a highly flexible approach to data processing.

Our data processing toolkit is designed for flexibility and scale, centering on:

• ETL Pipelines for the 80% Case: Modular, configurable pipelines handle the most common transformations—removing impossible values, standardizing lab units, handling duplicates, pivoting tables, and aggregating questionnaire scores. Each run creates an audit trail, ensuring traceability from raw data to final dataset.

• Clinical Data Scientist Input for the 20% Case: Unique, indication-specific, and study-specific challenges require expert intervention. Our clinical data scientists can inject custom code directly into pipelines, which is then subject to the same quality checks as standardized steps. This allows domain expertise to be encoded without sacrificing consistency or rigor.

• The Data API: Just as an application programming interface defines clear communication between systems, our data API standardizes how datasets are structured, formatted, and released. This allows our ML engineers to make reliable assumptions about data quality before training begins.

The result: a harmonized, transparent, and high-quality dataset that’s ready for AI model training and validation, and usable for downstream analyses.

When Data is Missing, Our Models Still Deliver

Data from past clinical trials and observational studies, even after harmonization, still contain gaps. This missingness is unavoidable in clinical data, since not every assessment is collected at every timepoint, and some records are incomplete.

Our DTGs are designed to learn from this reality. Instead of relying on naïve imputation methods, our modeling architecture is built to handle missing data natively. During training, DTGs learn patterns from incomplete longitudinal records. When applied in a new trial, the model can probabilistically infer missing values using correlations across biomarkers, clinical features, and patient trajectories.

For example, when a biomarker measurement is missing, the model infers it using correlations with related biomarkers and patient characteristics—without biasing outcomes toward an “idealized” patient.

The result: while the input datasets may contain missingness, the digital twins generated by our DTGs are complete and realistic representations of patient outcomes over time. This ensures that sponsors don’t face biased or “idealized” predictions, but instead receive digital twins that reflect the true complexity of patient data.

Model Validation That Proves Trust and Reliability

Every DTG undergoes rigorous evaluation, including cross-validation during model training to assess performance across outcomes and endpoints. This validation gives sponsors confidence that digital twin predictions remain accurate and unbiased when applied to new patients and trial settings.

For our partners, this translates into immediate access to validated, trial-ready DTGs—without the years of work it would take to build comparable infrastructure internally. Designing architectures that can robustly handle the intricacies of clinical data and validating them for trial use has taken Unlearn years of R&D. By partnering with us, sponsors bypass that burden and accelerate time to value, enabling clinical teams to focus on what matters most: designing smarter studies, making faster decisions, and moving trials forward with confidence.

This past quarter, what stands out most to me is the progress we’ve made in turning our vision for how AI will revolutionize clinical development into practical solutions sponsors are using today.

Clinical trials have always been complex, requiring input and expertise across biostatistics, regulatory, clinical, and operations teams, among others. At Unlearn, we’re innovating to bring those pieces together so programs can move forward faster, with fewer risks and better-informed decisions.

That commitment to bringing clarity and alignment to complex clinical development decisions is reflected in the partnerships we advanced this quarter.

Partnerships Powering Next Generation Clinical Trials

• This quarter, we initiated a major engagement with a leading global pharma company in Huntington’s disease. As part of the collaboration, the sponsor will use AI-generated digital twins of patients to create an external comparator arm for their Phase 1b study, providing a reliable benchmark for outcomes. In parallel, we’re supporting their Phase 2b/3 program by helping teams evaluate protocol scenarios, ground assumptions in stronger evidence, and align more efficiently across functions. The result is better informed decision-making, leading to fewer costly redesigns in one of the most complex areas of medicine.

• We also advanced a collaboration with top global pharmaceutical partners on regulatory bridging studies—trials designed to expand labels of marketed therapies in jurisdictions where they have not run registrational studies. In these studies, digital twins serve as a synthetic control arm, so that each patient can receive the experimental treatment while comparator outcomes are generated virtually. This approach provides a more efficient and patient-centric pathway to introducing existing medicines to new markets, while potentially adding hundreds of millions of dollars in annual sales to key drug franchises in their portfolio.

• We’ve also seen an increase in inbound sponsor interest in applying our AI applications to diverse development challenges in rare immunology diseases, such as generalized myasthenia gravis (gMG), as well as in metabolic disorders like obesity.

• We are building on customer interest in oncology with launching an early adopter program to partner with sponsors on solving critical pain points in cancer trials related to trial design and subgroup discovery, generating insights from fragmented trial data and better informing program decisions by boosting underpowered analyses —and invite interested teams to join us in shaping these solutions.

Growing our Leadership

We also strengthened our leadership this quarter with two key appointments:

Kwame Marfo, VP of Product

Kwame brings over 15 years of experience at Genentech and Komodo Health. He combines deep expertise in life sciences with product innovation. His leadership will guide the next phase of our platform’s evolution, ensuring our platform meets the highest standards of scientific rigor, regulatory trust, and customer value as we expand our product offerings.

Dr. Robert Lenz, Strategic Advisor

Rob has spent his career leading development strategies at scale, from his senior leadership roles at Amgen and Abbott to his time as Head of R&D at Neumora. While at Amgen, he oversaw more than 130 clinical studies in 50+ countries, enrolling tens of thousands of patients and helping drive adoption of innovative trial designs company-wide. Rob’s experience gives him firsthand insight into the challenges sponsors face in bringing new medicines to patients, and his expertise in leveraging novel technologies to advance drug development will be invaluable in our next chapter.

Looking Ahead

This quarter showed us what’s possible when we work side by side with our partners to tackle complex challenges, open new doors, and drive progress. As we carry that energy into the next quarter, we’re excited to keep pushing forward, together. If you’d like to explore how Unlearn can transform your program, reach out to our team.

Biostatistics teams are the foundation of every successful trial. Their analyses determine whether studies are credible, reproducible, and regulator-ready. They’re also the ones carrying a growing burden: balancing limited sample sizes with the demand for higher precision, navigating incomplete or messy datasets, and working under increasing pressure to deliver results faster.

Biostatistics teams already bring the rigor and judgment that guide every successful development program. What they deserve are tools that match the growing scale and complexity of today’s clinical research. That’s where Unlearn comes in.

Our AI-powered solutions extend the impact of biostatistics teams. Whether it’s harmonizing fragmented datasets so analyses aren’t constrained by missing values, simulating entire trials before the first patient is enrolled, or generating external controls that strengthen the credibility of open-label designs, the goal is always the same: to give statisticians greater confidence, flexibility, and precision within the frameworks they already use.

What Makes Our Approach Different

Unlearn’s Digital Twin Generators (DTGs) are disease-specific models designed to integrate with the statistical frameworks biostatistics teams already rely on. Each model is trained to forecast a participant’s clinical trajectory under standard of care, based only on their baseline characteristics. Because these forecasts are generated exclusively from baseline data, they can be incorporated into standard analyses to adjust for baseline risk in ways that are precise, regulatorily familiar, and immediately useful.

In practice, this gives statisticians a powerful new covariate—an individualized prediction that sharpens comparisons between treatment and control groups without requiring changes to trial design or analysis plans. When trials are underpowered, when treatment effects are subtle, or when variability is high, these forecasts provide the additional resolution needed to detect meaningful signals. They also open the door to exploring responder subgroups across endpoints and timepoints, giving biostatistics teams a clearer view of how treatments perform across diverse patient populations.

Extending Biostatistics Without Adding Burden

Unlearn’s solutions are built to make communication across clinical, statistical, and regulatory teams more straightforward. By providing transparent validation, clear documentation, and pre-specification language for SAPs, we give every group the same reference points to work from.

We also bring in scientific, statistical, and regulatory expertise to work alongside your team. That way, questions get resolved faster, expectations are aligned earlier, and outputs are easier for everyone to interpret and trust. Rather than adding to your workload, our role is to extend your team’s capacity, helping biostatistics teams move from design to analysis with fewer roadblocks and greater confidence.

The Right Tools for the Expertise You Already Have

If you’re planning a study and want to explore how AI-powered solutions can support your statistical team—from harmonizing fragmented data, to simulating trial designs, to generating regulator-ready digital twins—let’s talk.

We’ll walk you through exactly how our approach integrates with the frameworks you already use, and how it can extend the impact of your biostatistics group without adding new complexity or requiring years of infrastructure build.

Because your team doesn’t need to build AI solutions from scratch. They just need the right tools to do even more with the expertise they already have.

From patient selection to predictive modeling, leading sponsors are incorporating AI to drive efficiency gains of 30–70% in clinical development, thereby accelerating timelines and boosting success rates (McKinsey, 2023).

Yet, despite the promise, many AI initiatives introduce more risk than they mitigate—not because the technology falls short, but because implementation fails. Internal build teams may have strong technical talent, but they often lack the cross-functional structure, dedicated resources, and organizational trust needed to drive adoption. Timelines drag, regulatory confidence wanes, and new tools can go unused because no one is empowered to carry them forward. Even capable teams struggle to manage the change required to use what they’ve built.

External vendors face a different but equally limiting challenge. Many can deliver promising models, but few understand how to deploy those models into the tightly coordinated processes that bring drugs to market. Without alignment across clinical, statistical, and regulatory functions, even the most advanced solutions stall at the point of execution.

Unlearn bridges the gap between innovation and execution by embedding AI into trials with cross-functional support that directly incorporates into your team’s workflows. Here’s how.

Your Strategic AI Partner

When you partner with Unlearn, you gain an extension of your team—one that understands how clinical trial teams operate and can fill the resource or expertise gaps that can stall high-priority projects. We ensure seamless collaboration between internal and external stakeholders, allowing your scientific leaders to shine so innovation isn’t stalled by operational complexity.

Each engagement is led by a Solution Scientist, a senior scientific partner responsible for aligning goals, ensuring scientific integrity, and guiding project execution. They’re supported by a Relationship Manager who manages communication, scheduling, and project tracking throughout the partnership. Together, they coordinate a broader team of Unlearn specialists, each aligned with sponsor counterparts.

This embedded approach eliminates the disconnect that often occurs when vendors operate in silos or require sponsors to translate technical deliverables into trial-ready outputs. Unlearn’s team manages that full scope so you don’t need to reorganize internally or stand up a new AI function. We’ve already done the work to make that unnecessary.

Sponsors often struggle to align clinical, statistical, regulatory, and data functions internally. Unlearn reduces that friction by delivering an integrated team from the start, resourced and ready to deliver on complex, cross-disciplinary initiatives.

AI Built for Clinical Trials

With Unlearn, your team is equipped to lead. Our proprietary Digital Twin Generators (DTGs) forecast each trial participant’s trajectory under control across all clinical outcomes, subscales, and biomarkers, enabling your scientists to conduct hyper-granular analyses and explore “what-if” scenarios.

DTGs are disease-specific, rigorously validated, and used in trials for Alzheimer’s, ALS, Parkinson’s, and other complex indications. Combined with PROCOVA, our method for using digital twins qualified by the EMA and supported by the FDA, these tools help your team make confident, regulatory-aligned decisions.

This combination of AI modeling and regulatory coordination is rare, and it’s what allows Unlearn to operate in high-stakes environments where other AI partners fall short.

Flexible Delivery to Meet You Where You Are

• Full-Service Model: Unlearn sources the data, builds the models, and delivers results end-to-end.

• Custom Infrastructure Support: Unlearn integrates within your environment to build secure, proprietary solutions using your data.

Delivering Real World Impact

No matter how you choose to engage, Unlearn’s AI solutions are designed to fit within your existing clinical ecosystem, not force you to create a new one.

We’ve delivered for global pharma companies like J&J, AbbVie, and Roche, supporting trials from SAP development to regulatory submissions—without requiring sponsors to re-architect their internal teams.

That’s what makes us different, and why sponsors partner with us when they want to move quickly, confidently, and with clarity.

AAIC 2025 showcased an Alzheimer’s research community entering a new era. Sponsors are advancing earlier-phase trials, exploring a wider range of treatment modalities, and embracing biomarker-driven strategies for more targeted intervention. Yet while the science is accelerating, trial infrastructure is struggling to keep pace. Across sessions, speakers emphasized the need for better tools—like simulation and synthetic data—to support earlier decisions, optimize subgroup analyses, and adapt designs with greater confidence.

What Stood Out at AAIC 2025

1. Biomarker Momentum: Promise, Accessibility, and Open Questions

Multiple sessions focused on emerging biomarkers such as p-tau217, plasma GFAP, and digital measures, that make earlier and less invasive detection of dementia possible. Many of these biomarkers can now be measured in the blood, rather than through PET scans or lumbar punctures. Amyloid and tau positivity can precede symptoms by 15-20 years, reinforcing the urgent need for earlier interventions. However, researchers are still actively examining which biomarkers are most predictive of future cognitive and functional changes, including when those changes might occur and how severe they might be.

2. Early Treatment Requires Early Confidence

Speakers emphasized that anti-amyloid therapies such as lecanemab and donanemab work best at the earliest disease stages—possibly even pre-MCI. Yet, selecting the right patients remains a challenge due to the lack of sensitive cognitive diagnostics in early-stage disease. Early-phase trial strategies now require robust, evidence-based ways to assess whether a treatment signal is real—before advancing to pivotal studies.

3. Trial Design Needs a Modern Upgrade

Several presenters shared statistical strategies for improving trial efficiency, but even with more innovative designs, one key challenge looms large: enrolling enough patients, especially in early stages. As sponsors begin targeting patients in the earliest stages of disease, many of whom may not yet show symptoms, recruitment becomes exponentially harder. Sessions also highlighted the need for more flexible trial infrastructure and design strategies that can accommodate evolving biomarkers, eligibility criteria, and endpoints.

Our Take: Making Every Data Point Count in Alzheimer’s Trials

This year, we shared new research on how digital twins—AI-generated forecasts of individual patient outcomes—can optimize composite scores for greater sensitivity in early Alzheimer’s trials. Composite scores frequently serve as primary or secondary endpoints in early-phase studies but can be noisy or incomplete, sometimes capturing only a subset of relevant clinical change, such as cognition but not function. This allows sponsors to identify and reweight the most informative assessment items, tailoring composite scores to better detect true treatment effects.

Key Takeaways

• Digital twins forecast outcomes at the patient level, enabling reweighting of score components for stronger signal.

• Optimized scores improve power without requiring additional patients or changes to trial design.

• Simulation results showed power gains from 25–50% to 65% in moderately powered studies, and from 44–80% to 92% in better-powered studies.

• Composite scores tailored to the study population can better reflect meaningful clinical change, especially in early-phase populations..

• More sensitive endpoints mean clearer insights, faster interpretation of treatment effects, and greater confidence in trial outcomes.

The Path Ahead

We left AAIC energized by the momentum in the Alzheimer’s space and encouraged by the conversations we had with biopharma teams, regulators, and researchers who share our urgency.

There’s never been a more critical time to invest in smarter trial tools—and no better way to do so than by partnering with experts who’ve already built them.

[Download the full case study from AAIC 2025]