Subscribe now

I attended two events back-to-back: the Alchemistry Workshop on Free Energy Methods (May 4–6) and the inaugural CoFold Summit (May 6).

The first is the established annual conference for physics-based drug design, with participation and sponsorships from companies like Schrödinger, AstraZeneca, Cresset, OpenBioSim, etc. The second brought together the teams building deep learning co-folding models — Isomorphic Labs, Boltz, OpenFold, RoseTTAFold, SandboxAQ, and others. Same city, same week, overlapping audiences.

Frontier AI models alone won’t get you far in biology. We need to invest in physics-grounded tools and methods.

Anyway, speaking about free energy perturbation methods, they have become the industrial workhorse for binding affinity prediction, and GPU acceleration has moved them from supercomputers to everyday pharma workflows. But FEP needs good starting structures, and it struggles with structurally diverse compounds coming out of generative AI pipelines. On the co-folding side, models like Boltz-1x are making real progress on the chemical validity of predicted poses. But they still default to well-represented binding sites from training data and can’t reliably score what they generate. Allosteric pockets, novel targets, anything underrepresented, still a major challenge.

The pattern kept coming up in different sessions and hallway conversations. Co-folding generates structural hypotheses from sequence alone. Physics-based methods provide rigorous validation. Neither works well in isolation.

The companies doing interesting work at this interface — SandboxAQ, Genesis Molecular AI, Iambic, Schordinger, etc., seem to get this. The real progress is hybrid: learning-based generation feeding into physics-based refinement.

I think the current hype around general-purpose AI agents obscures something important about the pharma and biotech realm of AI progress. In drug discovery, binding is fundamentally a physics problem. Models trained on data can approximate physics, but they can’t replace it, not yet. The teams investing in both sides are the ones to watch.

Now, since conferences were specifically focused on FEP and co-folding methods, I decided to share a couple of trends in those areas here:

Observations about free energy perturbation methods

FEP has become the workhorse for binding prediction. Free energy perturbation calculations can now reliably predict how well a molecule binds to a protein target. What used to require supercomputers and deep specialist knowledge now runs on a few GPUs, thanks to better hardware and better classical force field parameterization.

The shift is from artisanal to industrial. Drug discovery moved from hand-designed molecules to combinatorial libraries in the late ‘90s, and FEP is following the same trajectory, from carefully hand-tweaked individual calculations to routine bulk triaging of large compound sets before synthesis.

The binding problem is increasingly solved; everything else is not. FEP handles potency prediction relatively well, but druglikeness, metabolism, PK, toxicity, and crystal polymorphs remain poorly amenable to computation. So FEP doesn’t replace medicinal chemistry judgment yet; it removes one major bottleneck while the others persist.

The generative chemistry + FEP synergy is the frontier, but it’s hard. AI-generated molecules tend to be structurally diverse (not congeneric series), which means you need absolute binding free energy (ABFE) calculations rather than relative ones. ABFE is more expensive and less accurate. That’s the current bottleneck for combining generative AI with physics-based validation at scale.

Ease of use matters for adoption. The article argues (via Cresset’s Flare product) that automation, error-checking, and cloud access are what turn a specialist method into an everyday tool across organizations.

Here is what Dmitry Lupyan, Research Leader at Schrodinger, got to say about the current state of FEP:

Observations about co-folding methods

Orthosteric binding works reasonably well; allosteric does not, yet. Co-folding methods reliably place ligands in the main (orthosteric) binding site but consistently fail to find allosteric pockets, instead defaulting to the orthosteric site. This is a training data bias problem because orthosteric sites arguably dominate the protein data bank (PDB).

Boltz-1x is the chemical validity winner. Only 1.5% of its predicted ligands had any PoseBusters issue (default settings), versus 56% for Boltz-1, 93% for NeuralPLexer, and 85% for RoseTTAFold. Under stricter criteria, everything degrades significantly.

Prevalence in training data correlates with success. When allosteric sites are well-represented in the PDB (like GCK), predictions improve. But it’s not the whole story — some well-represented allosteric sites still fail.

The dual-ligand trick helps but does not solve the problem. Submitting two copies of the allosteric ligand improved sampling (50% placed correctly), but you still can’t reliably distinguish the correct pose from incorrect ones without external scoring.

The core tension with co-folding methods: These methods show promise as potential replacements for docking and even FEP, but they currently lack physics-based scoring, produce ensembles of unknown quality, and need significant post-processing before they’re useful for prospective drug design.

📢 Announcement: I’ll be at HLTH Europe in Amsterdam this June (15-18) as an invited journalist/science writer. It is arguably Europe’s largest healthcare tech event, with 5,000+ attendees, one in three at the C-suite level.

I’ll be covering what’s actually being said in the hallways, not just on the stages. If you’re attending, let me know, happy to connect in person! And if there’s a specific topic or company you would want me to dig into while I’m there, drop it below.

Subscribe now

Company Picks

During the event, I talked to several company reps and founders, including , and so I decided to summarize some of the interesting companies in this space:

SandboxAQ

It is an enterprise AI company spun out of Alphabet in 2022, currently valued at $5.75B after raising ~$950M. Their drug discovery division, AQBioSim, combines generative AI with physics-based molecular simulation — what the company calls Large Quantitative Models (LQMs).

The key technical claim is their proprietary Absolute Free Energy Perturbation method (AQ-FEP), which, according to SandboxAQ predicts binding affinities without requiring reference compounds, making it applicable to structurally diverse libraries rather than just congeneric series.

The company says it can profile over 20,000 ligands per day at scale using this approach. They report partnerships with AstraZeneca, Sanofi, and UCSF, among others.

Genesis Molecular AI

Founded in 2019 in California (originally as Genesis Therapeutics, rebranded to reflect the AI focus). The company's core platform is GEMS (Genesis Exploration of Molecular Space), which, according to Genesis, combines proprietary deep learning models with physics-based molecular simulation for small molecule drug design.

Their flagship model, Pearl, is a 3D diffusion foundation model for protein-ligand structure prediction that the company claims outperforms AlphaFold 3 on binding pose prediction, notably, trained using large-scale synthetic data generated from physics simulations, not just experimental PDB structures.

Genesis has raised over $300M from investors including a16z, NVIDIA, Fidelity, and BlackRock. They report active collaborations with Gilead, Eli Lilly, and Incyte — the Incyte partnership was recently expanded to cover at least five additional targets, with Incyte sharing proprietary experimental data to further train GEMS. Nate Gruver from Genesis presented at the CoFold Summit in Session 2 on predicting properties beyond structure.

Iambic Therapeutics

Founded in 2019, headquartered in San Diego. A clinical-stage company whose platform combines two main proprietary AI models: NeuralPLexer, a co-folding model for predicting protein-ligand complex structures directly from sequence, and Enchant, a multimodal transformer that according to the company predicts clinical and preclinical endpoints from small, noisy datasets. The company describes its approach as physics-informed — integrating physical principles into AI architectures to improve data efficiency and enable broader exploration of chemical space. Iambic claims to complete full design-make-test cycles on a weekly cadence through tight integration of AI-generated designs with automated high-throughput chemistry and biology. They report their lead oncology program went from program start to clinic in under 24 months. Partnerships include a multi-year collaboration with Takeda announced in early 2026 (potentially worth over $1.7B in milestone payments) and a technology collaboration with Revolution Medicines. Matt Wellborn from Iambic presented at the CoFold Summit

Nostrum Biodiscovery

Founded in 2015 in Barcelona as a joint spin-off of the Barcelona Supercomputing Center (BSC) and the Institute for Research in Biomedicine (IRB Barcelona), with participation from the University of Barcelona and ICREA. Co-founded by Victor Guallar, Modesto Orozco, and Robert Soliva.

The company's core technology is PELE (Protein Energy Landscape Exploration), a Monte Carlo-based molecular modeling algorithm for protein-ligand docking, binding site prediction, and protein surface exploration. Their commercial platform, NostrumSuite, integrates PELE with AI-driven molecular modeling for virtual screening, hit-to-lead optimization, and applications across small molecules, antibody design, targeted protein degradation, and nucleic acid therapeutics.

According to the company, their ALScreen platform combines AI and molecular modeling for virtual screening of both predefined and ultra-large compound libraries. Nostrum describes itself as bridging physics-based simulation and AI — notably, they are rooted in HPC and biophysical simulation rather than coming from the deep learning side.

Apheris

A Berlin-based company co-founded by Robin Röhm that provides federated computing infrastructure for drug discovery. The core premise is that pharma companies hold proprietary structural and molecular data they can't share due to IP constraints, but that data is exactly what co-folding and ADMET models need to improve.

Apheris claims to solve this by bringing computation to the data rather than moving data, enabling multiple organizations to collaboratively train and benchmark AI models without exposing proprietary datasets.

They provide the technology layer for the AI Structural Biology (AISB) Network, an industry-led collaboration that, according to the company, includes AbbVie, Astex, AstraZeneca, Boehringer Ingelheim, Bristol Myers Squibb, Genentech, Johnson & Johnson, Sanofi, and Takeda. One of the network's flagship projects is fine-tuning OpenFold3 on proprietary structural data from multiple pharma companies — without that data leaving each organization — in collaboration with Mohammed AlQuraishi's lab at Columbia.

Their thesis is a bit different from what most AI biopharma companies are doing. Instead of designing new molecules, Biossil uses AI to dig through late-stage clinical failures and figure out which patient subgroups those drugs should have actually been tested on. Ten molecules were acquired while in stealth mode over three years. Trials running in everything from glioblastoma to Alzheimer’s.

That’s not drug repurposing in the classic sense — taking an approved drug and finding it a new indication, like thalidomide going from its original (disastrous) use to multiple myeloma, or metformin being studied in cancer.

Biossil is doing something more subtle: same molecule, same disease, just a more precisely defined subset of patients. The argument is that many drugs “failed” trials only in the “aggregate”, averaged across a heterogeneous population where a real signal got buried.

And they’re not alone. A cluster of companies, each with different technical approaches and varying levels of clinical evidence, is converging on a shared conviction: the pharma industry’s 90%+ clinical failure rate isn’t just a scientific problem. It’s partly an analytical one. The tools to find the right patients simply weren’t good enough, until now.

This is a piece about that convergence. We’ll map who’s doing what, how the approaches differ, what’s actually been validated, and whether the thesis holds up under scrutiny.

In this issue: The Logic of Drug Rescue — The Landscape: Who’s Doing What — A Closer Look at the Frontrunners — The Roivant Precedent — What Doesn’t Work (Yet) — Looking Ahead

💊 The Logic of Drug Rescue

Before we profile the companies, it’s worth understanding why this thesis is surfacing now and why it’s distinct from what came before.

Drug repurposing has a long history. Sildenafil started as a cardiovascular drug before becoming Viagra. Thalidomide was rehabilitated remarkably, decades after its teratogenic disaster in the 1960s, as a treatment for multiple myeloma. These are cases where an approved (or previously studied) molecule found a genuinely new indication.

What companies like a newcover Biossil, as well as more established players like Lantern Pharma, Pathos AI, and BPGbio, are doing is different. They are not necessarily changing the target disease, but the target patient. The hypothesis: within a trial population that produced a negative aggregate result, there are subgroups of patients who responded, and whose response was masked by the statistical noise of everyone who didn’t.

This isn’t a new idea conceptually. Post-hoc subgroup analysis has been part of clinical trials for decades. What’s new is the scale and sophistication of the AI being applied: multimodal foundation models trained on hundreds of petabytes of data, causal inference engines, spatial transcriptomics paired with pathology imaging, and multi-agent systems reasoning across publications and biomarker data.

The question is whether the analytical tools have finally caught up to the biological complexity… or whether we’re just building fancier ways to p-hack.

Subscribe now

🗺️ The Landscape: Who’s Doing What

The companies working in this space share a thesis but diverge significantly in their technical approaches, therapeutic focus, and maturity. Here’s how the landscape breaks down.

⭐ Biossil

The freshest entrant. Biossil emerged from stealth with $70M co-led by Founders Fund and OpenAI, and a portfolio of ten molecules acquired quietly over three years. Their approach centers on reanalyzing late-stage clinical failures to identify patient subgroups where a meaningful treatment signal was hidden by population heterogeneity. Trials are running across glioblastoma, Alzheimer’s, and other indications.

What makes Biossil notable is the breadth of their bet: ten molecules across multiple therapeutic areas, funded by investors who have not traditionally played in biopharma. The OpenAI connection signals a belief that general-purpose AI capabilities, not just domain-specific biostatistics, can crack the patient stratification problem. That’s an interesting bet, though one that remains unproven clinically.

Details on their technical platform are still limited. We’ll be watching for specifics on what data they’re training on, how their models identify subgroups, and, most critically, whether their approach produces prospectively validated biomarkers or just retrospective correlations.

⭐ Lantern Pharma (Nasdaq: LTRN)

Lantern has been working on a related playbook for years, making it one of the most useful reference points for whether the thesis actually holds up in the clinic. Their RADR AI platform identifies abandoned clinical-stage drugs and matches them to patient subgroups most likely to respond. The focus is oncology.

The most tangible proof point right now is LP-300, a candidate in development for never-smoker non-small cell lung cancer (NSCLC) — a molecularly distinct type of the disease with poor outcomes and no approved therapies focused on this specific population. Lantern has just announced that the FDA raised no objections to key proposed protocol amendments for the Phase 2 HARMONIC trial, a meaningful de-risking step.

Three protocol changes are worth noting:

• Focused enrollment on the EGFR exon 21 L858R subgroup — the molecular subset, accounting for roughly 40% of EGFR-mutant NSCLC globally, where current therapies leave the largest unmet need, according to the company, and where LP-300’s preliminary data have been most differentiated.

• Extended dosing from a maximum of 6 to 8 cycles.

• Transition from a randomized to a single-arm design — intended to accelerate enrollment and sharpen the clinical signal in a genomically defined subgroup.

On the AI side, Lantern has been developing withZeta.ai, a multi-agentic system derived from RADR that extends the platform’s mechanistic modeling capabilities. For LP-300, withZeta has been used to interrogate the drug’s mechanistic potential in the L858R setting — reasoning across publications and biomarker observations to surface insights that informed the development strategy. It’s an interesting example of AI agents being used not just for patient selection but for mechanistic hypothesis generation.

⭐ Pathos AI

Pathos is arguably the company with the deepest data moat in this space, thanks to its roots in the Tempus ecosystem. Founded by executives from Tempus (Eric Lefkofsky’s healthcare AI company), Pathos claims to have access to over 200 petabytes of multimodal oncology data linked to patient outcomes, reportedly 50 times the size of The Cancer Genome Atlas, the largest public genomic dataset in oncology.

Their thesis mirrors Biossil’s: drugs fail because they were tested in the wrong patients, with the wrong assumptions, in trials that couldn’t answer the real question “who benefits, and why?” But Pathos is focused exclusively on oncology and is building what it describes as the largest foundation model in the field.

In April 2025, Pathos entered a major three-way collaboration with AstraZeneca and Tempus to build a multimodal oncology foundation model, with $200 million in data licensing and model development fees flowing to Tempus. The foundation model is being built on Tempus's repository, which includes 7.3 million de-identified patient records, including 1.4 million with imaging data, 1.3 million with genomic information, and 260,000 with full transcriptomics profiles.

Pathos is also running its own clinical programs. In March 2025, they dosed the first patient in a Phase 1b/2a trial of pocenbrodib (a CBP/p300 inhibitor) in metastatic castration-resistant prostate cancer. They also acquired Known Medicine, which builds patient-specific 3D cell cultures and uses AI to predict drug responses prospectively — an attempt to close the loop between computational prediction and wet-lab validation.

Funding: $365 million in a Series D (May 2025), at a $1.6 billion valuation.

⭐ BPGbio

BPGbio takes a different technical angle: Bayesian causal AI, built on their NAi Interrogative Biology platform. Rather than relying on pattern recognition across large datasets, their approach aims to infer causal relationships, not just correlations, between patient characteristics and treatment response.

The platform integrates one of the largest non-governmental biobanks (over 100,000 clinically annotated patient samples) with deep multi-omic and clinical data, running on the Frontier exascale supercomputer at Oak Ridge National Labs.

The clearest demonstration of the approach in the drug rescue context comes from a multi-arm Phase Ib oncology study involving 104 patients across multiple tumor types. NAi’s models, trained on tissue and blood-derived multi-omic data, identified biological signatures predicting response to BPM31510 — and BPGbio used those insights to prioritize glioblastoma multiforme (GBM) and pancreatic cancer as the most compelling indications. The causal framing is that if the model can distinguish “patients who happened to respond” from “patients who responded because of a specific biological mechanism,” the resulting biomarkers should be more robust in prospective validation.

BPGbio is further along clinically than many AI-native biotechs. The company has completed enrollment in a Phase 2b GBM trial, with topline results expected in Q3 2026, and sought FDA guidance in late 2025 for a potential expedited regulatory path in GBM. They have multiple Phase 2 clinical trials underway — making them, by their own account, one of the first companies worldwide to advance multiple Phase 2 programs developed using causal Bayesian AI.

Share

🔹 Others in the Neighborhood

The companies above aren’t the only ones in this territory. Several others touch the thesis from adjacent angles:

🔹 NOETIK trains AI models on massive datasets of paired pathology images and spatial transcriptomics to find hidden biological subtypes among trial participants and predict which patients will respond. The pairing of imaging and spatial transcriptomics is technically ambitious and could surface subgroups invisible to genomics-only approaches.

🔹 Ignota Labs (London) takes a complementary but distinct approach. As CEO Sam Windsor has noted, sometimes the drug just isn’t good enough and Ignota focuses on fixing fundamental safety issues in the chemistry itself to give failed drugs a second chance. Patient stratification and molecular optimization are different interventions for the same problem (the 90%+ failure rate), and both are needed.

🔹 Formation Bio (New York, valued at ~$1.7 billion) acquires stalled clinical-stage drugs and uses AI to run trials more efficiently, optimizing patient recruitment, protocol design, and site management. In November 2024, they launched Muse, an AI tool for clinical trial recruitment, in partnership with OpenAI and Sanofi. Formation’s overlap with the “drug rescue” thesis is real but less precise: they’re improving trial execution, not fundamentally reanalyzing who should be in the trial. Closer to operational arbitrage than to computational patient stratification.

🔹 Origent Data Sciences uses machine learning to build patient-level predictive models for disease progression, then identifies cohorts within failed trials where treatment effects can be demonstrated. Their ForecastOne platform is specifically designed for drug rescue in neurodegenerative diseases — a space where population heterogeneity is especially pronounced.

📜 The Roivant Precedent

The idea of finding value in pharma's abandoned assets isn't new.

Roivant Sciences, founded in 2014 by Vivek Ramaswamy, was built on the thesis that the pharmaceutical industry was full of abandoned assets that failed not because of efficacy problems but because of strategic deprioritization. They licensed shelved drugs, housed them in independent subsidiaries (”Vants”), and pushed them through development.

The model worked, sometimes. The most famous case, buying GSK’s Alzheimer’s candidate intepirdine for $5 million via Axovant, failed in Phase 3. But Roivant learned and evolved. The company has since grown to a ~$20 billion market cap by pivoting to a precision focus on immunology and inflammation, and its AI story now lives in VantAI (a spinout building the Neo-1 model for molecular glue design), not in clinical failure analysis.

Roivant’s original model was a financial and operational bet, spot undervalued assets, give them focused management, and move fast. It wasn’t a computational bet on finding hidden responder subgroups in trial data per se. The companies profiled above are making a different bet: that AI can extract signal from noise in ways that traditional biostatistics couldn’t.

Roivant CEO Matt Gline has been candid about his skepticism of AI drug discovery, arguing it faces a fundamental problem by solving only one or two of the roughly 150 hard problems in preclinical development. But it was aimed at de novo drug discovery, not at the more focused application of AI for patient stratification in existing clinical data.

⚠️ Still Open Questions

We need to be honest about the risks and limitations.

The p-hacking problem is real. Post-hoc subgroup analysis is one of the oldest and most dangerous tools in clinical research. If you slice a trial population enough ways, you will find a subgroup that responded, by chance. The key question for every company in this space is: can your AI-identified subgroups be validated prospectively? Retrospective signal discovery is table stakes. Prospective confirmation is where most of these approaches will succeed or fail.

Regulatory uncertainty. The FDA has frameworks for enrichment strategies and biomarker-driven trial designs, but there’s no well-trodden regulatory path for “we reanalyzed a failed trial with AI and found a responding subgroup, now we want to run a new trial in just those patients.” Each company is navigating this largely ad hoc. Lantern’s FDA interaction on the HARMONIC trial amendments is a positive signal, but one data point doesn’t make a precedent.

Small subgroups, small markets. Patient stratification is a precision medicine play. By definition, you’re narrowing the addressable population. Some subgroups will be commercially viable (Lantern’s L858R NSCLC population of 65,000–80,000 relapsed patients per year is meaningful). Others may be too small to justify the cost of a dedicated clinical program. The economics of drug rescue only work if the subgroup is big enough, or if the development cost is low enough, to justify the investment.

Data access and quality. These approaches are only as good as the data they’re trained on. Pathos has an enormous advantage through the Tempus relationship (200+ petabytes), but most failed trials sit in corporate vaults, and the patient-level data needed for subgroup reanalysis is rarely publicly available. Companies that can’t access high-quality, multimodal, longitudinal patient data are building on thin foundations.

The “drug just isn’t good enough” problem. As Ignota’s CEO, Sam Windsor pointed out, sometimes patient stratification isn’t the answer; the molecule itself has fundamental issues. A drug with genuine safety problems or an insufficient therapeutic window won’t be rescued by finding better patients for it. The companies building these AI platforms need to be disciplined about walking away from molecules that don’t warrant rescue.

Self-reported early data. Most of the clinical results we’ve seen so far — including Lantern’s LP-300 PFS data — are company-reported, from early-stage trials, in small patient numbers. This is expected at this stage of maturity, but we shouldn’t confuse preliminary signals with validated outcomes. The real test comes in registrational trials with pre-specified subgroups and independently adjudicated endpoints.

🔭 Looking Ahead

So is “rescuing failed drugs with AI” a category now?

The capital providers certainly believe so. Between Biossil ($70M), Pathos ($365M Series D, $1.6B valuation), Formation Bio ($600M+, $1.7B valuation), and Lantern (public, running clinical trials), there’s real money behind the thesis, from investors ranging from traditional life science VCs to Founders Fund and OpenAI.

But a category needs more than capital. It needs clinical proof points. Here’s what to watch:

Near-term (2026–2027):

• Lantern’s HARMONIC trial data in the narrowed L858R NSCLC subgroup. This is one of the most concrete tests of the thesis: a drug that was broadly tested, AI-identified a specific molecular subgroup, and a redesigned trial is now running in just that population. If LP-300 produces confirmatory data, it’s a powerful proof of concept for the entire space.

• Pathos’s pocenbrodib Phase 1b/2a data in mCRPC — particularly whether their biomarker-defined subgroups show differential response.

• Biossil’s first clinical readouts from any of its ten programs.

Medium-term (up to 2028 and beyond):

• Whether any AI-identified subgroup leads to a regulatory filing. This would be the true inflection point — a drug that failed in a broad population, succeeded in an AI-defined subgroup, and got approved for that subgroup.

• The maturation of multimodal foundation models in oncology (the AstraZeneca-Tempus-Pathos collaboration). If these models can reliably predict responders across tumor types, the implications extend far beyond drug rescue.

• Whether pharma companies begin systematically reanalyzing their own shelved assets with these tools, either internally or through partnerships. The volume of failed late-stage programs sitting in corporate vaults is enormous.

The open questions:

• Can retrospective AI-driven subgroup discovery produce biomarkers robust enough for prospective trial enrichment? This is the central scientific question.

• Will regulators create clearer frameworks for AI-informed trial redesign, or will each program remain a bespoke negotiation with the FDA?

• Is there a sustainable business model here, or will drug rescue remain a niche strategy for specific asset classes? The economics depend heavily on how cheaply you can acquire failed assets, how efficiently AI can identify the right subgroup, and how large that subgroup turns out to be.

These are early days. The tools are getting dramatically more powerful, including multimodal models, massive patient datasets, causal inference frameworks, and agentic AI systems. But the clinical validation is thin, with only a handful of companies running trials. None has yet produced the definitive proof point: a failed drug, rescued by AI-driven patient stratification, approved by regulators.

We think this is a space worth watching closely, because the underlying logic is sound, the unmet need is enormous (90%+ failure rates, billions in sunk R&D), and the technical capabilities are likely approaching what the problem demands. The next 18–24 months of clinical data will tell us whether this is a genuine new paradigm or an expensive lesson in the limits of computational biology.

As always, if you’re working in this space or watching it from the inside, we’d love to hear what you’re seeing. Leave a comment!

Leave a comment

Read more

Read more

Read more

A team at Stanford recently posted a preprint describing a system they called “Virtual Biotech”: a coordinated squad of AI agents organized to mirror a real drug discovery company, complete with a virtual Chief Scientific Officer, specialized scientist agents, and over 100 tools for querying biomedical databases.

For their headline demonstration, they deployed over 37,000 agents in parallel, each one tasked with annotating a single clinical trial, linking therapeutic targets to genomic and single-cell transcriptomic features. The resulting dataset spans 55,984 trials. The analysis turned up what the authors call previously unreported associations: drugs targeting cell-type-specific genes were 40% more likely to advance from Phase I to Phase II, 48% more likely to reach market, and showed 32% lower adverse event rates.

In another case study, the system pulled together genetics, transcriptomics, and clinical data on B7-H3 in lung cancer and landed on an antibody-drug conjugate strategy, the same bet several pharma companies are already running in the clinic. It also flagged liabilities and differentiation angles. The whole thing reportedly cost $46 in API credits and took less than a day.

The Virtual Biotech is one lab’s preprint, but it lands in what we recently likened to a “gold rush”—agents are being deployed across clinical operations, translational biology, antibody design, and regulatory workflows. Major pharma companies are in an apparent compute arms race, stacking GPU clusters and billion-dollar AI partnerships within months of each other. Startups backed by hundreds of millions are launching agent-focused platforms. NVIDIA’s Jensen Huang even went so far as to declare agentic AI “the new computer” at this year’s GTC.

Whether the implementations match is another question. In a recent experiment, researcher Liang Chang asked—”Can AI make better decisions than pharma executives?” and sent AI agent teams back to a pivotal 2012 decision in oncology, the BMS vs. Merck biomarker strategy that ultimately decided the Keytruda-Opdivo war, and found that both Claude and GPT independently recommended the same path BMS took. The path that lost.

The agents produced rigorous analysis, identified the exact competitive threat, and still followed the consensus. As Chang put it: “AI can give you the best possible analysis. It can’t give you the courage to go against it.”

What can AI agents do today, where are they falling short, and why is everyone building them?

Subscribe now

🤖 Why agents, and why now?

Read more

Read more

GTC 2026 brought that energy straight into life sciences. NVIDIA CEO Jense…

Read more

Read more

⚠️ The most immediate one is patient recruitment. Far back in 1979, the father of clinical pharmacology Louis Lasagna observed that the pool of eligible patients shrinks by 90% the moment a trial opens, only to reappear once it closes. Lasagna’s Law remains as relevant as ever: according to a 2022 article, 11% of trial sites enrol zero participants and nearly 90% of trials face meaningful delays. With Phase II and III trials costing roughly $40,000 per day, the financial toll is brutal.

⚠️ Another issue is clinical attrition. Research from VU Amsterdam found that between 2012 and 2019, the share of trials successfully completing each phase declined steadily—particularly at Phase II. In the first half of 2024, nearly a third (32%) of trials were discontinued at Phase II—a 56% rise compared to pre-pandemic levels. Combined with stagnant rates of Phase III initiation over that same decade, the picture is one of a systemic bottleneck: trials that begin are increasingly unlikely to see the finish line.

⚠️ Rare disease research presents its own distinct challenge. As the FDA’s Rare Disease Evidence Principles note, shrinking patient populations make it progressively harder to generate reliable efficacy data through conventional designs—especially placebo-controlled trials, where enrolling enough participants to reach statistical significance can be close to impossible.

⚠️ Apart from operational intricacies, there is the ethical dilemma. Randomized controlled trials remain the gold standard for evaluating new therapies, but randomization isn’t always defensible. When an effective treatment already exists, assigning patients to a placebo raises serious moral questions, e.g. HIV cure trials with the antiretroviral treatment interruption.

The question, then, is whether parts of the control process can be simulated rather than physically recruited.

Digital twins are emerging as a compelling response. Last October, Sanofi Ventures invested in a digital twin platform developer QuantHealth, bringing its total funding to $30M. In 2025, the FDA announced plans to phase out animal testing requirements for monoclonal antibodies in favor of human-relevant methods, including AI-driven computational models—with the EMA moving in the same direction. Both industry and regulators, it seems, are taking this technology seriously.

👥 How it Works

A digital twin is a virtual replica of a physical object, continuously updated with real-world data so it mirrors the original’s behavior in real time. The concept, first applied by NASA in the 1960s, has since migrated from engineering into healthcare.

The applications are wide-ranging: optimizing industrial processes as Eli Lilly did to boost production of their GLP-1 drugs, predicting equipment failures, streamlining supply chains, and accelerating product development.

In a clinical trial patients are generally divided into two groups, also known as arms. The intervention arm receives the experimental treatment; the control arm receives a placebo, standard-of-care treatment or sham, serving as the baseline against which results are measured. Randomized controlled trials (RCTs) are the gold standard because randomization minimizes bias, but that randomization isn’t always flawless.

The traditional workaround of external controls drawn from historical trials, health records, or registries all carry their own limitations. For instance, data like those don’t include underrepresented groups or don’t account for placebo effect due to their observational nature.

Digital twins go a step further: using AI models augmented with historical data, they generate individualized predictions of how a patient might respond under different treatment scenarios. When used to simulate outcomes for patients who do not receive the experimental therapy, these models can produce a synthetic control arm.

These trial-level twins build on a foundation of patient-specific digital twin modeling (virtual replicas of individual physiology shaped by genomics, imaging, and clinical history) which we covered in our earlier overview of biological and patient-specific twins.

In clinical trials, an AI-powered digital twin typically operates in three steps:

• Build virtual patients — AI integrates biomarkers, imaging, genetics, and real-world evidence to generate synthetic profiles capturing the full variability of real populations.

• Run simulated trials — virtual cohorts replace placebo groups or test experimental therapies in silico, probing efficacy and safety without exposing patients to unnecessary risk.

• Optimize continuously — trial parameters like dosing and sample size are continuously refined in real time, anchored by validation against real-world data.

A less computationally demanding synthetic control arm approach uses AI-generated patient data based on registries, and real-world evidence but unlike DTs not modelling it on a particular individual. The appeal is sharpest in rare diseases, where finding enough eligible control patients is often impractical. The FDA, EMA, and NICE have all endorsed the approach, and it’s gaining traction: recent Phase II/III myeloma and lymphoma trials have leaned on external control data, and in at least one case (blinatumomab for acute lymphoblastic leukemia), a synthetic control arm helped support accelerated regulatory approval. AstraZeneca used over 300M synthetic patient records to advance its clinical trials, allegedly saving up to $100M per drug in development.

Synthetic control arms built from historical data have already supported label expansions and accelerated approvals with alectinib, blinatumomab, palbociclib among them. AI-generated individualized digital twins, however, have not yet served as primary evidence in a completed approval.

🦾 An Industry Arm

Read more

Read more

Shortly prior to last Christmas the European Commission published a proposal of a European Biotech Act, a strategic initiative aimed at setting up a regulatory framework to strengthen the life sciences sector across the EU. The document has been mostly positively received by the sector leaders as a needed step towards fostering local biotech innovation. Brussels isn’t stopping there. Commission President Ursula von der Leyen pitched EU-Inc, a pan-European company structure meant to solve what many see as the EU’s core startup problem of navigating 27 different bureaucratic regimes. The proposal would let one register in 48 hours, fully online and in English.

The Case for Urgency

Why does it matter? Europe gave the world its first blockbuster pharmaceutical (Aspirin in 1899) and just 30 years ago produced half of all new treatments globally. Today, that share has fallen to just one in five. Even though the EU biotech industry has grown twice as fast as the overall union’s economy over the last decade, it struggles to convert the world’s top science into commercially viable products.

Europe holds a comparable share of the top 10% most-cited biomedical research to the US and China, yet lags significantly behind in venture investment — a gap caused by underdeveloped private equity markets and fragmented, complex regulatory frameworks. The disparity is also visible in listing trends, with 66 of the 67 EU companies that went public over the past six years choosing foreign stock exchanges.

To address this, the Biotech Act includes measures like:

Read more

A newly launched U.S. bra…

Read more

Transformer architectures trained on biological data are beginning to predict drug response, generate therapeutic hypotheses, and identify which patients are likely to benefit from which treatments (part of a broader push that includes early attempts at virtual cell models) tasks that previously required years of wet-lab iteration. Some of that work is still early, though a handful of results have made it far enough through validation to be worth paying attention to.

• Google Research, Google DeepMind, and Yale spent much of 2025 scaling C2S-Scale, a language model that reads single-cell RNA data as text; the 27-billion-parameter version, released in April, came in October with wet-lab validation of a model-generated hypothesis about making immune-”cold” tumors visible to T cells.

• A collaboration between Microsoft Research, Providence Health, and the University of Washington took a complementary approach: GigaTIME, published in Cell in December, routinely converts pathology slides into virtual immune-protein maps, surfacing over 1,200 significant associations across 14,256 patients.

• At Davos in January, Demis Hassabis now put Isomorphic Labs‘ first trials, primarily oncology candidates, at end of 2026; the company followed this month with IsoDDE, a general-purpose drug design engine that reportedly doubles AlphaFold 3’s accuracy, already deployed across its oncology programs.

Not all of it is language-model work.

Read more

Separately, a new benchtop sequencer announcement from San Diego kept the “falling sequencing …

Read more

Just a day earlier, MIT Technology Review published its annual 10 Breakthrough Technologies list. This year, three of the highlighted technologies were in genomics: personalized gene editing, embryo scoring, and gene resurrection. From there, it seems like genomic applications are moving more into the mainstream technology discourse.

Another just-in data point from a few days ago is a report out of San Diego, where Element Biosciences says its newly announced VITARI benchtop sequencer can deliver a whole genome for $100, positioning it as a lower-cost alternative to Illumina’s high-throughput systems.

Looking at these and many of last year’s developments, genomics come into view as an integrated technology wave that extends from data generation to interpretation, intervention, and biological reconstruction.

With those early-2026 pings as a starting point, let’s do a selective pass through a few genomics patterns that seem to be carrying momentum into 2026.

Read more

Read more

At first glance, these headlines span different companies and medical areas. But they share a common thread: each centers on advanced therapeutic modalities (ATMs)—a new generation of medicines that go beyond the limits of conventional drugs.

According to BCG, eight of the top ten best-selling biopharma products in 2025 are new-modality drugs, and the global pipeline value for these therapies has reached $197B. ATMs are becoming a more established part of the industry and are noticeably contributing to its growth.

Reject Tradition, Embrace Modernity

Read more

Read more

In DealForma’s figures cited by CEO Chris Dokomajilar, deal flow between large-cap biopharma and Chinese biopharma accelerated in 2024-2025. In 2025, big pharma completed 18 in-licensing and asset purchase deals (just one in 2020) from Chinese companies with $50M+ upfronts, totaling $57.3B in deal value and $3.9B in upfront cash and equity. By 2026, China continues to emerge as a major source of globally licensable, clinical-stage biotech assets, backed by an increasingly complete innovation stack, even as new policy constraints complicate cross-border data flows and outsourcing.

In late January, speaking at the Asian Financial Forum in Hong Kong, executives from Merck and Amgen pointed to China as a likely early approval market for fully AI-designed drugs. Merck China president Marc Horn suggested that 2026 could mark the shift from AI-assisted discovery to compounds designed end-to-end by AI entering regulatory pipelines, citing China’s patient datasets, clinical execution, and the government’s recent “AI Plus” policy push. Amgen’s chief medical officer Paul Burton pointed to a similar timeline, seeing 2026 as a year when AI-driven and human genetics–led discovery could begin translating more directly into drug candidates.

For perspective, among recent big pharma deals involving Chinese companies, this year’s JPM week had AbbVie’s $5.6B partnership with RemeGen around a bispecific oncology asset. Looking back at just 2025, Pfizer licensed a bispecific from 3SBio with $1.25B upfront, AstraZeneca entered a multi-year $5.3B AI-enabled small-molecule discovery collaboration with CSPC Pharmaceuticals, and GSK’s x Jiangsu Hengrui agreements included $500M upfront and up to about $12B in potential milestones.

From Generics to Innovation

Read more