Running a platform that tracks clinical trial data has given me an unusual vantage point on AI drug discovery: I see a lot of confident claims about what AI is doing in drug development, and then I see the actual trial registrations, protocols, and data. The claims and the reality have a gap that matters for anyone trying to understand what "AI-discovered drug" actually means in clinical practice. Insilico Medicine's INS018_055 is frequently cited as the first drug "entirely designed by AI" to reach Phase 2 clinical trials — and that claim is substantially true and genuinely significant. But AI designed the molecule; humans ran every pre-clinical assay, wrote the IND, conducted the Phase 1, and are running the Phase 2. The AI contribution is real, but the framing of AI replacing drug development is misleading in ways that affect how this field should be evaluated.
This article is for informational purposes only and does not constitute medical advice. The drugs discussed are in clinical trials and are not approved for general use. Do not seek to access trial compounds outside of a clinical trial setting.
Summary
Insilico Medicine's INS018_055 (TRAF2- and NCK-interacting kinase [TANK] inhibitor for IPF) is in Phase 2 clinical trials — designed using their Pharma.AI platform encompassing target identification, de novo molecule generation, and ADMET prediction. Recursion Pharmaceuticals REC-4881 (MAP4K4 inhibitor for Neurofibromatosis 1) and REC-2282 (EZH2 for NF2) are in Phase 1/2. Exscientia's DSP-1181 (AI-designed OCD treatment) completed Phase 1. BenevolentAI's baricitinib COVID-19 repurposing (approved) is the most commercially successful AI drug discovery outcome to date. AlphaFold 2's protein structure predictions have materially reduced the barrier to structure-based drug design for previously undruggable targets. The 2026 active AI drug pipeline includes approximately 160 programs in IND-enabling or clinical stage across major AI drug discovery companies.
ClinicalMetric Analysis
- The relevant measure of AI drug discovery is not "how many AI-designed drugs have reached clinical trials" but "what fraction of AI-designed drugs succeed in clinical trials versus non-AI drugs." The conventional industry Phase 1→approval success rate is approximately 10%. If AI-designed drugs succeed at 10%, the AI contribution is largely cosmetic — it may reduce the cost or time to reach Phase 1, but it isn't materially improving the odds. The meaningful hypothesis is that AI can improve clinical success rates by better predicting which molecules will have the right combination of selectivity, ADMET properties, and target engagement. This hypothesis has not yet been testable because most AI programs only entered clinical trials in 2021–2024 and don't have enough data for a rate comparison. By 2028–2030, we will have enough Phase 2 outcome data to evaluate this hypothesis properly. Until then, the AI drug discovery field is running on pre-clinical rationale and investor enthusiasm — not clinical evidence that AI improves drug success rates.
- AlphaFold 2's protein structure predictions have had a more immediate and demonstrable impact on drug discovery than generative AI molecule design — but it's quieter because it's harder to market. AlphaFold 2 predicted accurate structures for virtually every protein in the human proteome. For drug discovery, this matters because structure-based drug design — docking small molecules into a target's binding pocket — requires knowing the 3D structure of the protein. Structures that previously took years and millions of dollars to solve via X-ray crystallography are now available for free. This has materially accelerated hit identification for targets where experimental structure was previously unavailable. KRAS G12C (long considered undruggable) benefited from computational work enabled by structural predictions. AlphaFold's contribution to drug discovery is real, large, and largely unrecognized in mainstream coverage of "AI drugs" because it doesn't produce a molecule — it creates an enabling resource.
- Recursion's phenotypic screening approach is methodologically different from Insilico's generative chemistry — and the distinction matters for predicting what diseases each approach is best suited for. Insilico uses AI for target identification and de novo molecule generation around a specific target. This works when you have a good hypothesis about the target — i.e., when traditional medicinal chemistry would work anyway, just more slowly. Recursion uses large-scale cellular imaging (millions of cells per week, phenotypic screening) to identify drugs based on cellular phenotype changes without requiring a pre-specified target hypothesis. Their RXRX program maps disease biology at phenotypic scale and uses AI to find cellular perturbations that mimic a healthy phenotype. This approach potentially works in diseases where the biology is poorly understood and a specific target isn't known — which is the majority of diseases. Both are genuine approaches with different strengths; they shouldn't be lumped together under "AI drug discovery."
INS018_055: What "AI-Designed Drug" Actually Means Here
INS018_055 is a small molecule inhibitor of TANK-binding kinase 1 (TBK1), designed for idiopathic pulmonary fibrosis (IPF). Insilico Medicine used their Pharma.AI platform at three stages: PandaOmics identified TBK1 as a novel IPF target from multi-omics data analysis; Chemistry42 generated de novo molecules designed to bind TBK1's binding pocket; and ADMET.ai predicted pharmacokinetic and toxicity properties to prioritize candidates. The entire target identification and initial molecule design required approximately 18 months — versus a typical 4–6 years for traditional early discovery.
Pre-clinical results were published in Nature Biotechnology: INS018_055 showed selectivity for TBK1, anti-fibrotic activity in bleomycin mouse and humanized IPF models, and acceptable ADMET profile. Phase 1 (NCT05154240): safety, PK confirmed in 48 healthy volunteers and 12 IPF patients. Phase 2 (NCT05975983): ongoing in IPF patients, primary endpoint FVC change at 24 weeks, enrollment ~60 patients. Results expected 2026. The field's attention is entirely on whether this drug will demonstrate clinical efficacy — the "AI designed it" question becomes irrelevant once the Phase 2 data is available.
Recursion: Phenotypic AI Meets Rare Disease
Recursion Pharmaceuticals takes a fundamentally different approach. Their operating system runs approximately 2.2 million experiments per week using automated imaging of cellular phenotypes under thousands of small molecule and genetic perturbations. The company has generated what they call a "phenomics map" — a multi-dimensional characterization of how cells respond to thousands of perturbations across hundreds of disease-relevant cell lines. AI (graph neural networks, self-supervised learning) identifies disease-relevant cellular signatures and drugs that reverse them.
REC-4881 (MAP4K4 inhibitor for Neurofibromatosis 1-related plexiform neurofibroma): Phase 1 completed, Phase 2 being designed. REC-2282 (EZH2 inhibitor for NF2-mutant meningioma): Phase 1 ongoing. Recursion's Novartis collaboration focuses on rare disease targets identified through the phenomics approach. Their most important commercial output to date may not be any individual drug but the Recursion OS platform, which Roche/Genentech licensed in a $150M+ deal.
BenevolentAI and the Repurposing Model
BenevolentAI's most commercially significant contribution was identifying baricitinib (an approved JAK inhibitor for RA) as a potential COVID-19 treatment, based on AI analysis of its ability to inhibit AAK1 — a kinase involved in viral endocytosis. This prediction was made in February 2020, before most clinical trials had enrolled. The COV-BARRIER Phase 3 trial confirmed mortality benefit in hospitalized patients on high-flow oxygen; baricitinib received FDA EUA for COVID-19 in 2020 and is now an approved treatment. This is the most commercially successful AI drug discovery outcome to date, precisely because it operated in the repurposing space — where clinical safety data already existed — rather than requiring de novo development. Repurposing remains the lowest-risk AI drug discovery application because the chemical and safety information already exists.
AlphaFold and Structure-Based Discovery
Isomorphic Labs (DeepMind spinout) and Relay Therapeutics are using AlphaFold-enabled structure-based drug design for previously undruggable targets. Relay's RLY-4008 (FGFR2 inhibitor targeting a specific exon 17 mutation) demonstrated 81% ORR in FGFR2 fusion-positive cholangiocarcinoma in a Phase 1/2 trial — an unusually high response rate for a kinase inhibitor, attributed to the exquisite selectivity enabled by structure-based design. This is a case where computational structural biology (informed by AlphaFold) demonstrably contributed to clinical success through selectivity that wouldn't have been achievable with purely empirical medicinal chemistry.
Participating in AI Drug Discovery Trials
AI drug discovery programs follow the same clinical trial framework as conventional drugs — they have INDs, IRB oversight, informed consent, and standard eligibility criteria. From a patient perspective, participation in a trial of an AI-designed drug is identical to participation in any other Phase 1 or Phase 2 trial of a novel agent. The AI origin of the drug is irrelevant to eligibility, safety monitoring, or treatment protocol. Search ClinicalTrials.gov for "Insilico," "Recursion," "Exscientia," "BenevolentAI," or specific drug names (INS018_055, REC-4881, DSP-1181) to find active enrollment opportunities.