AI Drug Discovery Platforms in 2026: Insilico INS018_055, Recursion+Exscientia, Schrödinger, AlphaFold3, and the Clinical-Trial Bottleneck
Insilico Medicine's INS018_055 in Phase 2, the Recursion+Exscientia merger, Schrödinger's physics+ML platform, Iktos, Atomwise, AlphaFold3 from Isomorphic Labs, and why discovery wins still meet the clinical-trial bottleneck.
The most useful framing for AI drug discovery in 2026 is that the field has cleanly bifurcated. On one side are the discovery wins — molecules designed by AI that did reach the clinic, often faster and cheaper than industry benchmarks. On the other side is the clinical reality: AI discovery does not shorten Phase 2 or Phase 3, does not avoid the fundamental biology of efficacy and safety, and has not yet produced an approved drug. The 2023-2025 data — Insilico’s INS018_055 reaching Phase 2, the Recursion-Exscientia merger, the AlphaFold3 release from Google DeepMind’s Isomorphic Labs — gave the field its first honest answer on what works and what does not.
This post is the engineer-friendly view of the platforms, the milestones, and the structural reality that comes after a molecule is designed.
What “AI discovery” actually means in 2026#
The phrase has been so heavily marketed that it is worth being precise. The technical layers, roughly in order of where AI has shown clinical-stage results:
Target identification. Use multi-omics and literature data to identify which protein or pathway to drug. ML helps, but most program-deciding work is still expert biology.
Hit finding. Generative chemistry, virtual screening, or active learning over compound libraries to find starting points. This is where AlphaFold, RoseTTAFold, and the diffusion-based generative models (Boltz, Chai, RFdiffusion) have changed the field most.
Lead optimization. Iteratively improve a hit’s potency, selectivity, and pharmacokinetics. Physics-based methods (free-energy perturbation, MD simulation) plus ML for property prediction.
ADMET prediction. Absorption, distribution, metabolism, excretion, toxicity. Pure ML, mostly trained on historical assay data. Still imperfect at the edge cases that kill programs.
Clinical trial design and execution. Patient stratification, biomarker selection, recruitment optimization. AI helps, but the biology dictates the result.
The discovery wins of 2024-2025 are mostly in the first three layers. The clinical reality bottleneck shows up in the last two.
Insilico Medicine and INS018_055#
The cleanest published case study in AI drug discovery is Insilico Medicine’s INS018_055 program for idiopathic pulmonary fibrosis. The company used its Pharma.AI platform — target identification with PandaOmics, generative chemistry with Chemistry42 — to identify a TNIK inhibitor as a candidate. The published timeline from program start to IND was approximately 30 months, well under the industry benchmark of 4-6 years for the discovery-to-IND phase.
INS018_055 entered Phase 1 in 2022, Phase 2 trials began in 2023, and by 2024-2025 the program was the most-watched AI-discovered asset in clinical development. The interim readouts have been mixed in the way that Phase 2 readouts often are — signals of biological activity, the harder question of whether the magnitude justifies Phase 3 still open.
What Insilico established credibly is that AI tools can compress the discovery timeline. What is still open is whether Phase 2 and Phase 3 success rates for AI-discovered molecules are any different from industry baselines.
Recursion, Exscientia, and the consolidation#
The Recursion-Exscientia merger announced in August 2024 and completed in late 2024 was the most consequential strategic event in AI drug discovery to date. Recursion brought industrial-scale phenotypic screening — millions of cell-painting images generated weekly, with proprietary perturbation maps used to identify drug targets. Exscientia brought a stronger lead-optimization platform with several molecules already in the clinic and a notable Bristol Myers Squibb partnership.
The combined company has a meaningful clinical pipeline (Exscientia’s CDK7 inhibitor EXS-21546, several other small molecules), industrial-scale phenotypic data (Recursion’s), and a Sanofi partnership that survived the merger. The merger also signalled that pure-play AI biotechs need scale to survive — the cost of running an AI discovery engine alongside a real clinical pipeline is high enough that small companies struggle to do both.
The broader consolidation pattern through 2024 and 2025 — Schrödinger acquiring smaller biotech assets, BenevolentAI restructuring, Atomwise’s continued pivot toward partnerships rather than internal pipeline — suggests the field is settling into roughly three viable models: tools-as-software (Schrödinger), pipeline-with-platform (Recursion+Exscientia, Insilico, Isomorphic Labs), and partnership-only (most of the smaller platforms).
Schrödinger and the physics-plus-ML platform#
Schrödinger’s position is the most differentiated in the field. The company sells industry-standard physics-based simulation tools — FEP+ for free-energy perturbation, Glide for docking, Maestro for structure-based design — and has been integrating ML on top of and alongside those physics methods. Most large pharma uses some Schrödinger software somewhere; Schrödinger also has its own internal pipeline and partnership programs.
The argument for physics-plus-ML over pure ML is honest about what each approach is good at. Physics is more reliable for binding affinity prediction on novel chemotypes; ML is faster for triaging large libraries. The two combine well — ML triages a billion-compound virtual library down to the thousand that physics then evaluates rigorously. The 2024-2025 work integrating ML into FEP+ and the new generative models within the Schrödinger toolchain is the platform-software side of the story.
Iktos, Atomwise, and the partnership-only platforms#
Iktos’s Makya generative chemistry platform and the Spaya synthesis-planning tool have notable partnerships with Sanofi, Pfizer, Janssen, and Servier. Atomwise’s Atomnet structure-based virtual screening platform is similarly partnership-heavy. Neither company has the internal pipeline scale of Insilico or Recursion+Exscientia, and both have settled into a posture of providing tools and joint discovery programs rather than going independent on clinical development.
The honest read on partnership-only platforms is that they survive by being genuinely useful to pharma teams that prefer to keep clinical development in-house. The value-capture is smaller per program but the cost base is also smaller.
AlphaFold3 and Isomorphic Labs#
AlphaFold3, released in May 2024 by Google DeepMind and Isomorphic Labs, was a step change for structure prediction in drug discovery. Where AlphaFold2 predicted protein structures, AlphaFold3 predicts structures of proteins with ligands, with nucleic acids, with ions, and with other proteins — the full biomolecular complex. The accuracy improvements on protein-ligand and protein-nucleic-acid prediction matter for early-stage hit finding and structure-based design.
The Isomorphic Labs commercial arm — separate from DeepMind, set up to monetize the science — landed Novartis and Eli Lilly partnerships in early 2024 worth nearly 3 billion USD in total upfront and milestones. The 2024-2025 work has been less about model releases and more about translating AlphaFold3 into the daily workflow of drug-discovery teams — the structure of an active site, predicted with appropriate uncertainty, becomes a constraint on generative chemistry rather than a separate analysis step.
The honest caveat: AlphaFold3 is excellent at static structure prediction, less so at predicting conformational dynamics, allosteric sites, and protein-protein interfaces that move on biological timescales. The next generation of models — Boltz, Chai, RFdiffusion-AA — are working on these limitations with varying degrees of success.
The clinical-trial bottleneck#
The structural reality that every AI discovery story hits eventually: the clinical trial. AI has not meaningfully shortened Phase 2 or Phase 3 timelines. Patient recruitment still takes years. Regulatory standards are unchanged. The biology of human disease is still what it was. A molecule designed in six months still has to survive a multi-year clinical program before it becomes a medicine.
The places where AI does help in clinical development are real but narrower than the marketing suggests. Patient stratification using biomarkers — finding the right subpopulation to enroll. Synthetic control arms drawn from real-world data, where regulators allow them. Operational efficiency — site selection, recruitment prediction, monitoring. None of these compress Phase 3 by 50 percent. Most of them save months and improve the probability of a successful readout by margins that are real but not revolutionary.
The clean-eyed prediction for 2026-2028: AI will produce a meaningful number of new INDs each year — possibly dozens — and a small number of approved drugs over the rest of the decade. The headline number that matters is not “molecules entered Phase 1” but “molecules approved.” That number is small in 2026 and will not become large until the late 2020s.
What this means for technology teams in pharma#
For technology and data teams inside pharma, the practical 2026 picture is that the AI discovery question has moved from “should we” to “how do we.” The infrastructure that supports this — a chemistry-aware data lake, structured assay data going back decades, the model-serving stack for generative chemistry and structure prediction, the audit trail that the FDA will eventually ask about — is substantial work. We have helped enterprise data teams stand up the data engineering foundations underneath this kind of platform; the value is in plumbing more than in the model itself.
Related reading#
AI moves discovery faster but the clinic still decides what becomes a medicine. If your organization is building the data and platform foundations for AI-aided discovery, our data engineering team has stood up the infrastructure that this work actually needs. Tell us about the platform.