Artificial intelligence has moved from laboratory promise to actual human trials, with one AI-designed drug now in late-stage testing for idiopathic pulmonary fibrosis. However, despite six years of clinical progress, no AI-discovered drug has yet won regulatory approval, and early success rates reveal a sobering reality: AI excels at designing molecules but struggles with the biology that matters most to patients.

What Has AI Actually Achieved in Drug Discovery?

The most clinically advanced AI-discovered asset is rentosertib, developed by Insilico Medicine using its Pharma.AI platform. The entire preclinical program, from identifying the target protein to nominating a drug candidate, took approximately 18 months and cost roughly $2.6 million. For context, the pharmaceutical industry typically spends $430 million to bring a single drug to market, though that figure includes all failures and regulatory costs across a decade or more.

In a Phase IIa trial published in Nature Medicine in 2025, rentosertib showed a dose-dependent improvement in lung function. Patients receiving 60 milligrams once daily experienced a mean increase in forced vital capacity of 98.4 milliliters compared to a decline of 20.3 milliliters in the placebo group. On July 7, 2026, Insilico announced the drug had advanced to Phase III testing, involving 320 patients across 47 centers in China with a 52-week primary endpoint. This marks the first AI end-to-end discovered drug to enter late-stage development.

Several other AI-designed drugs have reached early clinical stages. Schrödinger's SGR-1505, a MALT1 inhibitor for blood cancers, showed no dose-limiting toxicities and an overall response rate of 22 percent in Phase I testing across 49 patients. BenevolentAI's BEN-8744, a peripherally restricted PDE10 inhibitor for ulcerative colitis, demonstrated safety in 54 healthy volunteers without central nervous system adverse events, a key improvement over prior drugs in its class. However, not all programs have succeeded. Sumitomo Dainippon and Exscientia's DSP-1181 for obsessive-compulsive disorder was discontinued after Phase I results fell short of expectations.

Where Is AI Failing to Deliver?

The clinical data reveals a critical gap between molecular design and real-world efficacy. AI-native programs achieved Phase I success rates of 80 to 90 percent, substantially exceeding the historical industry average of 50 to 60 percent. This suggests AI is genuinely better at identifying molecules with favorable drug-like properties. But Phase II success rates for AI-discovered drugs align with historical norms at approximately 40 percent, indicating that target validation and disease biology, not molecular optimization, remain the primary clinical bottlenecks.

In other words, AI can design better molecules, but it cannot yet reliably predict whether those molecules will actually work in patients. BenevolentAI's BEN-2293, a topical pan-TRK inhibitor for atopic dermatitis, met its primary safety endpoint in Phase IIa but failed primary efficacy endpoints in the intention-to-treat population, though a post-hoc subgroup signal in patients with 20 percent or greater body surface area involvement requires prospective validation.

What Structural Barriers Are Slowing Progress?

Beyond clinical trial results, several systemic challenges limit AI drug discovery productivity. Model generalizability is a primary scientific concern: AI models trained on curated benchmark datasets frequently show degraded performance in prospective, real-world applications. A 2026 risk-tiered validation framework identified four validation tiers, internal machine learning reproducibility, molecular-science benchmarks, prospective experimental validation, and clinical or translational calibration, and emphasized that retrospective benchmark enrichment is an unreliable predictor of prospective performance.

Data quality and fragmentation compound these challenges. Pharmaceutical datasets are proprietary and biased toward well-studied diseases and molecular targets, limiting AI models' ability to generalize to rare diseases or novel mechanisms of action.

How Can AI Drug Discovery Teams Improve Their Odds?

• Prospective Validation: Move beyond retrospective benchmarking and test AI predictions in real-world experimental settings before committing to expensive clinical trials, ensuring models perform as expected outside controlled datasets.
• Data Sharing and Integration: Establish collaborative frameworks to pool proprietary pharmaceutical data across companies and institutions, reducing bias and improving AI model generalizability to underrepresented disease areas.
• Target Validation Focus: Prioritize AI investment in target identification and disease biology rather than molecular optimization alone, since Phase II failures indicate the bottleneck is biological validation, not chemical design.
• Multi-Omics Integration: Leverage AI platforms that combine multiple data types, such as genomics, proteomics, and network analysis, to identify targets with stronger causal links to disease, as demonstrated by Insilico's PandaOmics engine.

What Does This Mean for Patients and the Industry?

The transition of AI from preclinical research tool to operationally embedded component of pharmaceutical R&D is real and measurable. As of April 2024, eight leading AI drug discovery companies had 31 drugs in human clinical trials. Rentosertib's advancement to Phase III represents a genuine milestone, but it also underscores a hard truth: the first AI-discovered drug to win regulatory approval has not yet emerged, and the clinical evidence supporting AI's superiority over traditional drug discovery remains incomplete.

The pharmaceutical industry is betting that AI will eventually crack the problem of target validation and disease biology. The speed and cost advantages demonstrated in preclinical work are compelling. But until an AI-discovered drug achieves regulatory approval and demonstrates clear patient benefit, the gap between computational promise and demonstrated clinical success will remain substantial. Rentosertib's Phase III trial, expected to conclude in 2027 or 2028, will be a critical test of whether AI can finally deliver on its promise.