The First Wave of AI-Designed Drugs Hits Human Trials: A 2026 Status Report on Insilico, Isomorphic, and Recursion

The decade-long promise that artificial intelligence would compress drug discovery from a 12-year, $2.6 billion endurance test into something faster, cheaper, and more clinically accurate is now generating its first meaningful body of human evidence. As of May 2026, more than 75 molecules originally designed or substantially shaped by AI methods are in active human clinical trials worldwide, according to a Boston Consulting Group analysis tracking biotech disclosures. The lead candidates have advanced through Phase 2, several have reported their first proof-of-concept data, and the field is now far enough along to ask the question that biology, not algorithms, ultimately answers: do these molecules work in patients?

Presented By Our Partners

The answer, so far, is mixed in exactly the way veteran drug developers predicted it would be. AI has dramatically accelerated early discovery and reduced the number of synthesized molecules required to identify a candidate. It has not yet rewritten the laws of human biology that make Phase 2 the most failure-prone stage in the pipeline. Insilico Medicine’s INS018_055 for idiopathic pulmonary fibrosis, the first wholly AI-designed drug to reach Phase 2, has produced encouraging interim data. Recursion Pharmaceuticals has seen two of its lead programs stumble in Phase 2 readouts. And Isomorphic Labs, the DeepMind spinout building on AlphaFold’s protein structure breakthroughs, has signed its first multibillion-dollar pharma partnerships but has yet to put a molecule in a human.

This is the field’s first real proof point, and the data are starting to settle the most consequential debate in modern pharmacology: how much of drug discovery is a computational problem, and how much is, and will remain, a biological one.

The Inflection Point: Why 2026 Is the First Real Read

For most of the past decade, AI drug discovery lived in the realm of decks, press releases, and preclinical mouse studies. The pivotal shift came in 2023, when Insilico dosed the first patient in its Phase 2 IPF trial using a molecule whose target was nominated by a machine learning algorithm and whose structure was generated by a generative chemistry model. That milestone, repeatedly cited in industry analyses including a 2024 Nature Reviews Drug Discovery commentary from Jayatunga, Ayers, Bruens, Jayanth, and Meier, did not prove that AI works. It proved that AI-designed molecules could clear the regulatory and chemistry hurdles required to test the hypothesis in humans.

The 2026 inflection is different. Multiple programs are now far enough along to produce signal data, the kind of human evidence that either validates or falsifies the underlying computational thesis. The Boston Consulting Group analysis, which has tracked AI-discovered molecules since 2017, finds that the AI-native pipeline now includes roughly 24 Phase 2 assets and at least three programs that have reached Phase 3 or pivotal study design. Sun, Ayers, and colleagues at BCG have argued that Phase 1 success rates for AI-discovered molecules are now running at approximately 80 to 90 percent, materially higher than the historical industry average of 40 to 65 percent, although they caution that the absolute numbers remain small and the populations selected to date may be enriched for tractable targets.

The Phase 2 question is the one that matters. Historically, only about 30 percent of molecules that enter Phase 2 advance to Phase 3, and the dominant failure mode is lack of efficacy, not safety. If AI-discovered drugs maintain anything close to industry-average Phase 2 attrition, the field will have demonstrated speed and cost gains without solving the deeper problem of biological prediction. If they materially outperform, the entire economics of pharmaceutical R&D could shift.

Insilico Medicine and the IPF Question

Insilico Medicine, founded in 2014 by Alex Zhavoronkov and now headquartered between New York and Hong Kong, has been the most aggressive proponent of end-to-end AI drug discovery. Its candidate INS018_055, originally branded ISM001-055, is a small-molecule inhibitor of TNIK, a kinase implicated in fibrotic disease. The molecule’s target was nominated by Insilico’s PandaOmics platform, which mines transcriptomic and proteomic datasets for novel disease drivers, and its chemical structure was generated by Chemistry42, the company’s deep generative chemistry engine.

In May 2024, Insilico published interim Phase 2a data in Nature Biotechnology and reiterated the findings in subsequent investor disclosures: the molecule was well tolerated across a 12-week dosing period and produced numerical, though not statistically powered, improvements in forced vital capacity, the standard surrogate endpoint in IPF. The 2025 Phase 2b enrollment, with a 60-patient cohort across Chinese sites, completed dosing in late 2025, and Insilico has signaled that interim Phase 2b data are expected in the third quarter of 2026.

The IPF program matters not just on its own merits but as a methodological proof point. The conventional industry benchmark cited in McKinsey’s 2023 analysis is that a single new chemical entity typically requires synthesis and testing of 5,000 to 10,000 compounds to identify a viable preclinical candidate. Insilico has reported that INS018_055 was identified after synthesizing fewer than 80 compounds, a reduction of roughly two orders of magnitude. The discovery-to-IND timeline was approximately 18 months, against an industry average of 3 to 6 years. Whether those efficiency gains translate into clinical efficacy is the question Phase 2b is designed to answer.

Isomorphic Labs: The DeepMind Spinoff Lands Its First Partnerships

Isomorphic Labs, founded by DeepMind chief executive Demis Hassabis in 2021 and built on the foundation of the AlphaFold protein structure prediction system, has taken a different commercial path. Rather than building its own pipeline, the company has positioned itself as a platform partner to large pharmaceutical companies, with the thesis that structure-first drug design enabled by AlphaFold 3 and its successor models can dramatically improve hit rates in lead optimization.

In January 2024, Isomorphic announced two landmark partnerships, one with Novartis worth up to $1.2 billion in milestones and one with Eli Lilly worth up to $1.7 billion. Both deals cover multiple targets and span the discovery and optimization phases. Through 2025, Isomorphic disclosed that its platform was actively engaged in lead optimization for oncology and cardiovascular programs at both partners, and the company raised a $600 million Series A in April 2025 led by Thrive Capital and GV, with Alphabet retaining its majority stake.

Isomorphic has not yet put a molecule into a human, and Hassabis has been notably circumspect about timelines. In a January 2025 Financial Times interview, he framed the company’s ambition as building the first end-to-end computational drug discovery engine capable of predicting clinical efficacy from a target hypothesis, a problem he described as a decade or more out at minimum. The 2026 question for Isomorphic is whether either the Novartis or Lilly programs will produce a development candidate that advances to IND filing on a timeline materially shorter than industry benchmarks.

Featured Partner

Invest in the Infrastructure Behind Modern Medicine

As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.

Learn More →

Recursion’s Phase 2 Reality Check

If Insilico and Isomorphic represent the optimistic case for AI drug discovery, Recursion Pharmaceuticals has supplied the field’s first sobering data. The Salt Lake City company, founded in 2013 by Chris Gibson, Blake Borgeson, and Dean Li, pioneered the use of high-content cellular imaging combined with deep learning to characterize compound effects at scale, and its 2021 IPO raised $502 million.

Recursion entered 2025 with three lead programs in Phase 2: REC-994 for cerebral cavernous malformation, REC-2282 for neurofibromatosis type 2, and REC-4881 for familial adenomatous polyposis. Each program had been advanced based on AI-driven phenotypic screening and target nomination. The clinical data, when they arrived, were largely disappointing.

In November 2024, Recursion reported that REC-994 had failed to meet its primary efficacy endpoint in the SYCAMORE Phase 2 trial. In April 2025, REC-2282 produced equivocal data in the POPLAR Phase 2/3 trial, with company analyses suggesting potential benefit in a subset of patients but no clear win on the primary endpoint. The REC-4881 FAP program produced more encouraging Phase 1b data in mid-2025, showing dose-dependent polyp reduction, but the program remains early.

Recursion’s 2024 merger with Exscientia, the UK-based AI drug discovery company that had previously partnered with Sumitomo Pharma, consolidated two of the field’s most prominent platforms and gave the combined entity a pipeline of more than 10 clinical assets. The strategic logic, as articulated by Chris Gibson on the company’s Q4 2024 earnings call, was that the field had moved past the point where any single AI platform could deliver enough shots on goal to validate the model; scale and pipeline breadth were now the dominant requirements.

The Recursion experience underscores a point that veteran drug developers have made consistently. The bottleneck in drug discovery is not synthesizing molecules. It is identifying the right biological target and predicting whether modulating that target will produce a clinically meaningful effect in heterogeneous human populations. AI has materially accelerated the former problem. The latter remains largely unsolved.

The Foundation Model Era Arrives in Pharma

The most important architectural shift in 2026 is the broad migration from task-specific machine learning models to foundation models, large neural networks pretrained on enormous datasets that can be fine-tuned for specific drug discovery tasks. AlphaFold 3, released by Google DeepMind and Isomorphic in May 2024 and described in a Nature paper by Abramson, Adler, Dunger, and colleagues, expanded the original AlphaFold’s protein structure prediction to model the interactions between proteins, small molecules, DNA, RNA, and ions in a single unified framework. The model’s release reset the field’s baseline for what is achievable in structure-based drug design.

The competitive response has been rapid. Chai Discovery, a 2024-founded startup with backing from OpenAI and Thrive Capital, released Chai-1 in September 2024, an open-weights model that approaches AlphaFold 3’s performance on several benchmarks. Boltz-1, released by MIT and collaborators in November 2024, made open-source foundation models for protein interaction prediction broadly available. By early 2026, every major pharmaceutical company had stood up internal foundation model teams, and a 2025 IQVIA survey reported that 78 percent of large-cap pharma companies were running foundation models in at least one production drug discovery workflow, up from 22 percent two years earlier.

The clinical payoff from foundation models is still years out, because the molecules these systems are now helping to design have only recently entered preclinical development. The 2027 and 2028 IND filings will be the first cohort that meaningfully reflects this architectural shift.

Generate Biomedicines and the Protein Design Wave

While much of the AI drug discovery conversation has focused on small molecules, the more dramatic technical progress in 2024 and 2025 has been in generative protein design. Generate Biomedicines, the Cambridge, Massachusetts company founded in 2018 by Flagship Pioneering, has built a generative platform called Chroma that designs novel protein structures conditioned on a desired function. The company’s lead candidate, GB-0669, an inhalable monoclonal antibody for COVID-19 protection in immunocompromised patients, entered Phase 1 in early 2024 and reported acceptable safety and pharmacokinetic data in late 2025.

The broader significance of the Generate program is the demonstration that an entirely de novo protein, designed by a generative model rather than discovered through immunization or natural sequence mining, can clear the early safety hurdles required for clinical development. The field’s other prominent protein design entrant, Xaira Therapeutics, launched in April 2024 with $1 billion in committed capital from ARCH Venture Partners and Foresite Capital and is building on the RFdiffusion protein design framework originally developed in the Baker lab at the University of Washington.

David Baker’s 2024 Nobel Prize in Chemistry, shared with Demis Hassabis and John Jumper for computational protein design and structure prediction, was the field’s first formal recognition that AI-driven protein engineering had crossed a scientific threshold. The Nobel committee’s citation explicitly framed the breakthrough as foundational for next-generation medicines.

The Honest Numbers: What AI Has and Has Not Changed

The most rigorous attempt to quantify AI’s actual impact on drug discovery has come from a series of Boston Consulting Group analyses published between 2022 and 2025. Their consistent finding is that AI-discovered molecules have entered Phase 1 with success rates of approximately 80 to 90 percent, compared to industry historical averages in the 40 to 65 percent range. The Phase 1 advantage appears to be real and may reflect better target selection, cleaner molecules with fewer off-target liabilities, and more accurate preclinical-to-clinical translation predictions.

Phase 2 data are too sparse to draw firm conclusions, but the early signal is less favorable. The BCG cohort includes a small number of Phase 2 readouts, and the success rate so far is broadly consistent with industry historical averages of around 30 to 40 percent. This is the pattern that veteran observers including Derek Lowe at In the Pipeline and Bernard Munos at FasterCures have predicted for years. AI helps most where computational prediction is closest to the underlying problem, in molecular structure generation, target identification, and early ADMET profiling. It helps least where the rate-limiting bottleneck is the biology of human disease, which remains poorly understood for most major indications.

The cost story is harder to characterize. Insilico and several other AI-native companies have claimed total preclinical R&D costs of roughly $10 to $30 million per development candidate, against industry averages of $50 to $100 million. These numbers are difficult to audit and likely reflect selection effects, with AI companies pursuing targets that are inherently more tractable. The discovery-to-IND timeline gain is more defensible, with multiple AI-native programs producing IND-ready candidates in 18 to 30 months versus industry averages of 4 to 6 years.

The 2026 Scorecard

As of May 2026, the AI drug discovery field can credibly claim three concrete achievements. First, it has produced its first wave of clinical-stage assets, with more than 75 molecules in active human trials. Second, it has demonstrated faster and cheaper preclinical development, with reductions of approximately 50 to 70 percent on both time and cost for the lead AI-native companies. Third, it has built a foundation model substrate that is now broadly available across the industry, ensuring that the next generation of molecules will benefit from substantially better computational tools than the current cohort.

The field has not yet demonstrated three things that will determine whether AI drug discovery becomes the dominant paradigm in pharmaceutical R&D or a useful adjunct to traditional methods. First, no AI-discovered drug has yet reached FDA approval, and the first approval, most likely from the Insilico IPF program or one of the Recursion oncology assets, is unlikely before 2028. Second, Phase 2 success rates remain unproven at scale, and the field needs at least 50 to 100 Phase 2 readouts before a robust statistical comparison to industry baselines can be made. Third, no AI platform has yet demonstrated the ability to predict clinical efficacy from a target hypothesis with materially better accuracy than human-driven hypothesis generation, which is the holy grail Hassabis and others have set as the field’s true objective.

What This Means For You

If you are a patient or a long-term observer of medicine, the practical takeaways from the 2026 state of AI drug discovery are these.

The first cluster of clinically meaningful AI-discovered drugs will likely reach FDA approval in the 2028 to 2030 window. Idiopathic pulmonary fibrosis, certain oncology indications, fibrotic kidney disease, and several genetically defined rare diseases are the most probable areas of first approval. Patients with these conditions should expect to see enrollment opportunities in late-stage AI-native trials over the next 18 to 36 months, and the National Institutes of Health ClinicalTrials.gov database is the most reliable place to identify them.

For chronic disease management, the immediate practical impact of AI drug discovery in 2026 is minimal. The drugs you take today, and the drugs you will take in the next three to five years, were almost entirely developed with conventional methods. The treatment decisions that affect your health over the next half decade should be made on the basis of currently available evidence, not on the promise of AI-discovered therapies that are still in trials.

For longer-term health planning, the most consequential shift is likely to come in the molecular characterization of disease rather than in the speed of drug approval. AlphaFold and its successors are giving researchers structural insight into proteins that have been considered undruggable for decades, including KRAS variants, multiple intracellular protein-protein interaction targets, and several membrane-bound receptors implicated in neurodegeneration. The therapeutic implications of this work will play out over 10 to 20 years and could materially expand the universe of treatable diseases.

For anyone making investment, career, or research decisions in life sciences, the 2026 data argue for a balanced view. The AI drug discovery thesis is more validated than skeptics suggested in 2020 and less validated than evangelists claimed. The companies most likely to succeed over the next decade are those that have built deep biology capabilities alongside their computational platforms, that have access to high-quality proprietary data, and that have the operational discipline to advance multiple shots on goal in parallel. The era of the single-platform AI drug discovery pure-play is likely over. The era of AI as a foundational layer in every serious drug discovery effort has firmly begun.

The next 24 months will be the most important the field has yet seen. By mid-2028, AI drug discovery will either have produced its first FDA-approved medicines and a Phase 2 success rate that materially exceeds industry averages, or it will have settled into a more modest role as a useful efficiency tool that has not yet rewritten the economics of pharmaceutical innovation. The first wave is in the water. The data that will decide which story is true are coming.

Free Daily Briefing

The Latest Longevity Science.
Delivered Every Morning.

Join researchers, physicians, and health professionals getting daily breakthroughs in AI-driven medicine, epigenetics, and longevity research.

Support the research that powers this editorial

Subscribe Free

No spam. Unsubscribe anytime. We respect your inbox.