Introduction
Artificial intelligence (AI) has moved from a theoretical accelerant to a demonstrable contributor to pharmaceutical development. As of July 2026, multiple AI-designed or AI-discovered drug candidates have entered Phase 1–3 human trials, generating peer-reviewed clinical evidence and attracting more than half a billion dollars in pharma partnership payments. This review, intended for medical professionals, synthesizes the current clinical pipeline, profiles the leading companies, assesses scientific validation, and examines the regulatory landscape across the United States, European Union, and China.
Key definitions: An AI-designed drug is one for which AI contributed to target identification, molecule generation (de novo design), or both. Generative AI refers to algorithms that create novel chemical structures rather than simply selecting from existing libraries. Machine learning (ML) denotes statistical models trained on experimental data to predict biological or chemical properties. Target identification is the process of selecting a disease-relevant biological molecule to modulate. Model validation refers to testing whether an AI model performs reliably in independent, real-world datasets.
The Clinical Pipeline: What Has Reached Human Trials?
As of mid-2026, the most advanced and best-evidenced program is rentosertib (formerly ISM001-055), developed by Insilico Medicine. Rentosertib is a first-in-class oral inhibitor of TNIK (TRAF2- and NCK-interacting kinase), a fibrosis-relevant serine/threonine kinase. Critically, both the target and the molecule were generated using Insilico's Pharma.AI platform—specifically PandaOmics for target discovery and Chemistry42 for generative chemistry—making it the most transparently documented "end-to-end AI" drug in the clinical literature. The Phase IIa GENESIS-IPF trial (NCT05938920), a randomized, double-blind, placebo-controlled study enrolling 71 patients with idiopathic pulmonary fibrosis (IPF) across 22 Chinese sites, met its primary safety objective and demonstrated a mean forced vital capacity (FVC) improvement of +98.4 mL at 60 mg once daily versus −20.3 mL in the placebo group over 12 weeks. Dose-dependent reductions in serum fibrosis biomarkers (collagen 1A1, fibronectin, matrix metalloproteinase-10) correlated with FVC improvement, supporting biological plausibility. Phase III (NCT07687459) was initiated in July 2026, enrolling 320 patients across 47 centers in China, with annualized FVC decline rate over 52 weeks as the primary endpoint 376.
Beyond IPF, the pipeline spans oncology, immunology, and inflammatory disease. Recursion Pharmaceuticals—which completed its merger with Exscientia in November 2024—operates the largest named AI-native clinical portfolio. REC-4881, a MEK1/2 (mitogen-activated protein kinase 1/2) inhibitor in familial adenomatous polyposis (FAP; NCT05552755), represents the first published clinical validation of Recursion's full-stack AI Operating System: 75% of evaluable patients showed reductions in total polyp burden, with a 43% median reduction at 12 weeks and a durable 53% median reduction sustained after cessation of therapy. Grade ≥3 adverse events occurred in 15.8% of patients, with no Grade ≥4 treatment-related events 1415. Generate Biomedicines' GB-0895, a computationally engineered long-acting anti-TSLP (thymic stromal lymphopoietin) monoclonal antibody with an approximately 89-day half-life, entered Phase 3 (SOLAIRIA-1 and SOLAIRIA-2 trials) in December 2025, targeting approximately 1,600 adults and adolescents with severe asthma across more than 40 countries 3940. BenevolentAI's BEN-8744, a peripherally restricted PDE10 (phosphodiesterase 10) inhibitor for ulcerative colitis—a target identified by the Benevolent Platform with no prior direct associations in the literature—completed Phase Ia (NCT06118385) with a favorable safety profile, no serious adverse events, and pharmacokinetics supporting twice-daily dosing 3438.
Table 1. Representative AI-Designed / AI-Discovered Drug Candidates in Clinical Development (2024–2026)
| Drug | Sponsor | Target | Therapeutic Area | Modality | Phase | Key Clinical Evidence | AI Platform |
|---|---|---|---|---|---|---|---|
| Rentosertib (ISM001-055) | Insilico Medicine | TNIK | IPF | Small molecule | Phase III (initiated Jul 2026) | FVC +98.4 mL vs −20.3 mL (placebo) at 12 wks (Phase IIa) 637 | PandaOmics + Chemistry42 |
| REC-4881 | Recursion | MEK1/2 | Familial adenomatous polyposis | Small molecule | Phase 1b/2 | 75% polyp reduction; 43% median reduction at 12 wks; 53% median reduction sustained after cessation of therapy 14 | Recursion OS |
| REC-617 | Recursion | CDK7 | Platinum-resistant ovarian cancer | Small molecule | Phase 1/2 | Promising safety/preliminary efficacy (Nov 2025) 14 | Recursion OS |
| Zovegalisib (RLY-2608) | Relay Therapeutics | PI3Kα (mutant) | HR+/HER2− metastatic breast cancer | Small molecule | Phase 3 (ReDiscover-2) | Phase 1/2: median PFS 10.3 mo; ORR 39% (all pts) 17 | Dynamo platform |
| GB-0895 | Generate Biomedicines | TSLP | Severe asthma | Monoclonal antibody | Phase 3 | Phase 1: ~89-day half-life; sustained biomarker reduction ≥6 months 3940 | Generative biology platform |
| SGR-1505 | Schrödinger | MALT1 | Relapsed/refractory B-cell malignancies | Small molecule | Phase 1 | Favorable safety; responses in CLL and WM; FDA Fast Track + Orphan Drug (WM) 18 | Physics+AI platform |
| BEN-8744 | BenevolentAI | PDE10 | Ulcerative colitis | Small molecule | Phase 1 (completed) | No SAEs; PK supports BID dosing; ex vivo cytokine reduction 3438 | Benevolent Platform |
| ABS-201 | Absci | PRLR | Androgenetic alopecia / Endometriosis | Antibody | Phase 1/2a | Initiated Dec 2025; interim readout expected 2H 2026 16 | Generative AI protein design |
| zasocitinib (TAK-279) | Nimbus (Schrödinger-enabled) | TYK2 | Immune-mediated disease | Small molecule | Phase 3 | Physics-enabled design; late-stage validation 7 | Schrödinger platform |
CLL: chronic lymphocytic leukemia; WM: Waldenström macroglobulinemia; PFS: progression-free survival; ORR: objective response rate; BID: twice daily; SAE: serious adverse event; PK: pharmacokinetics.
A broader review of AI-discovered molecules through December 2023 identified 24 that had completed Phase 1, with approximately 80–90% success—substantially exceeding the historical industry average of 40–65%—suggesting AI platforms are effective at producing drug-like molecules with acceptable pharmacokinetics and tolerability. However, Phase 2 success rates approximate 40%, matching historical averages, indicating that efficacy validation remains dependent on target biology and trial design rather than discovery method 7.
The Company Landscape: AI-Native to AI-Assisted
Table 2. Major AI-Enabled Drug Discovery Companies and Platform Characteristics
| Company | Platform Type | Clinical Assets (Representative) | Key Pharma Partnerships | AI Attribution Level |
|---|---|---|---|---|
| Insilico Medicine | AI-native; end-to-end generative AI | Rentosertib (Phase III); ISM9528 (preclinical) | Eli Lilly, Sanofi, Bora Pharma, CMS 13 | Highest: both target and molecule explicitly AI-generated 637 |
| Recursion + Exscientia | AI-native OS; phenomics + ML | REC-4881, REC-617, REC-1245, REC-3565, REC-4539 | Sanofi ($134M to date), Roche/Genentech ($213M) 1415 | High for target discovery (Recursion OS); AI contribution to chemistry varies by asset 14 |
| Generate Biomedicines | Generative biology platform | GB-0895 (Phase 3) | Undisclosed 39 | High: antibody computationally engineered for ultra-high affinity and extended half-life 39 |
| Absci | Generative AI protein design | ABS-201 (Phase 1/2a), ABS-501 (preclinical) | Large pharma (undisclosed) 16 | High for antibody design; no independent verification in retrieved materials |
| Relay Therapeutics | Dynamo platform (cryo-EM + computational) | Zovegalisib (Phase 3) | Elevar Therapeutics 17 | Moderate: structure-based/computational design; not strictly generative AI |
| Schrödinger | Physics+AI computational platform | SGR-1505 (Phase 1), SGR-3515 (Phase 1) | Lilly TuneLab, Manas AI, Otsuka, Takeda 18 | Moderate: physics-enabled; zasocitinib cited as late-stage validation 718 |
| Genesis Molecular AI | Foundation models (GEMS platform) | 5+ Incyte targets (discovery-IND) | Incyte ($120M upfront, $232M/program milestones) 19 | High for platform; no clinical assets yet |
| BenevolentAI | Benevolent Platform (data integration + ML) | BEN-8744 (Phase 1 complete) | Novartis 2021 | High for target discovery (PDE10 identified by AI); molecule design less disclosed 34 |
Scientific Validation: What Can and Cannot Be Attributed to AI?
The strongest causal case for AI contribution remains rentosertib, where Insilico has publicly documented AI-driven target identification and molecule generation, corroborated by peer-reviewed Phase IIa results in Nature Medicine 6. Recursion's REC-4881 provides the first clinical validation of an AI operating system in patient outcomes, with durable polyp reduction in FAP 14. GB-0895's engineered 89-day half-life and sustained TSLP blockade demonstrate that generative AI can optimize complex pharmacokinetic properties in biologics 39.
However, several important limitations apply. First, development timelines, while accelerated (18 months from initiation to preclinical candidate nomination for rentosertib), lack industry-wide standardized benchmarks; most companies focus on newsworthy proofs of concept rather than systematic transparency 8. Second, Phase 1 success does not predict Phase 2 efficacy; the ~40% Phase 2 success rate for AI-discovered molecules matches conventional drug development 7. Third, clinical registries rarely document AI provenance, making independent validation of AI claims difficult 9. Fourth, many companies market "AI-driven pipelines" without fully disclosing whether the specific target, scaffold, or optimization step was AI-generated versus AI-assisted—a critical distinction for credible attribution 5.
Regulatory Challenges: FDA, EMA, and NMPA
Regulatory frameworks are rapidly evolving. In January 2025, the FDA published draft guidance on "Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products," drawing on over 500 submissions with AI components from 2016 to 2023 and feedback from more than 800 public comments. In January 2026, the FDA and EMA jointly published 10 guiding principles for good AI practice in drug development 942. China's NMPA issued comprehensive implementation guidance in April 2026 ("Implementation Opinions on 'AI + Drug Regulation'") targeting a 2030 strategic vision for AI-integrated regulation 11.
Table 3. Regulatory Challenges and Practical Implications Across Major Jurisdictions
| Challenge | FDA Approach | EMA Approach | NMPA (China) Approach | Clinical/Physician Implication |
|---|---|---|---|---|
| Model transparency and explainability | Documentation of architecture, training data, validation methods; AI as auxiliary, not replacement 9 | Human-centric framework; data provenance and traceability required 10 | Algorithm transparency mandated; full-process record-keeping required 11 | IND submissions must justify AI-derived hypotheses with mechanistic or empirical evidence |
| Data provenance and bias | Audit trails, data quality documentation; bias assessment across diverse populations 9 | Bias detection across demographic groups; algorithmic transparency 10 | Standardized data submission; high-quality datasets and vertical LLMs required 11 | Trial design must address AI-driven patient selection bias; diverse training data is essential |
| Good Machine Learning Practice (GMLP) | Draft guidance 2025; CDER AI Council (est. 2024) coordinates implementation 9 | EMA/FDA joint 10 principles (Jan 2026): independence of training/test sets, fit-for-purpose reference standards, deployed model monitoring 42 | Human-machine collaboration with traceability; AI maintains auxiliary role 11 | Compliance increases cost; smaller AI companies may face disproportionate regulatory burden |
| Validation of AI hypotheses | Functional preclinical validation required before IND; clinical target engagement data necessary 9 | Model performance required across diverse settings; comprehensive lifecycle guidance 10 | Implementation of validation protocols by 2030 11 | Clinicians should expect Phase 2b/3 evidence before adopting AI-discovered therapies |
| Reproducibility and IP | Documentation required; open-source not mandated 9 | Independent validation emphasized; methodology disclosure required 10 | Data security and sensitive data management mandated 11 | Proprietary AI models limit independent replication; regulatory uncertainty for follow-on drugs |
Clinical and Investment Implications
For physicians and trial investigators, AI-designed drugs require the same rigorous Phase 2b/3 efficacy and safety evaluation as conventionally discovered compounds. The elevated Phase 1 safety profile does not indicate superior efficacy—Phase 2 success rates remain comparable to historical benchmarks 7. Biomarker-guided patient selection (e.g., PIK3CA mutation for zovegalisib, TNIK pathway activation for rentosertib) may enhance trial efficiency, but these strategies require transparent validation and regulatory acceptance 12.
For biopharma R&D and investors, partnership revenue has become a critical validation signal. Recursion's $500+ million in milestone and upfront payments across all its partnerships, and Genesis Molecular AI's $120 million upfront from Incyte, signal pharma confidence in AI platforms, but long-term clinical success remains unproven 141519. Emerging technologies—including AI-enabled digital twins qualified by the EMA for Phase 2/3 trials and synthetic control arms—may further streamline development 12.
Near-term milestones through 2026–2028 include rentosertib Phase III primary endpoint readout, zovegalisib Phase 3 data in HR+/HER2− breast cancer, GB-0895 SOLAIRIA trial results, ABS-201 interim efficacy data in androgenetic alopecia (2H 2026), and first clinical data from Genesis Molecular AI–Incyte collaboration targets. These readouts will collectively determine whether AI's early-phase promise translates into regulatory approvals—the definitive measure of clinical value.
Conclusion
By July 2026, AI-driven drug discovery has achieved genuine early clinical validation. Rentosertib's Phase IIa results and Phase III initiation, Recursion's FAP data, Generate Biomedicines' Phase 3 launch in severe asthma, and BenevolentAI's completed Phase I in ulcerative colitis collectively demonstrate that AI-native platforms can produce clinically active molecules. However, causal attribution of clinical success to AI remains challenging, standardized development benchmarks are lacking, and Phase 2 efficacy rates have not yet exceeded historical averages. Evolving regulatory frameworks from the FDA, EMA, and NMPA are establishing essential standards for model transparency, data integrity, and human oversight. Medical professionals should interpret AI-designed drugs through the lens of rigorous clinical evidence—scrutinizing AI attribution claims, demanding functional target validation, and awaiting Phase 3 outcomes before drawing conclusions about the long-term impact of AI on pharmaceutical innovation and patient outcomes 6789101137.