AI Drug Discovery Excels at Molecules, Fails at Clinical Trials

The pharmaceutical industry's multibillion-dollar bet on artificial intelligence reaches a critical inflection point in 2026, as the first substantial cohort of AI-designed drugs enters Phase 3 clinical trials. The results will reveal whether AI's impressive gains in molecule design translate to medicines that actually help patients—or whether the investment boom targeted the wrong problem entirely.

The transformation at the front end

AI has genuinely revolutionized early-stage drug discovery. Companies like Absci can now design functional antibodies by testing fewer than 100 candidate designs, compared to the millions of compounds screened in traditional chemical libraries. Insilico Medicine reports completing the journey from target selection to clinical candidate in roughly 18 months for approximately $2.6 million—a process that historically required four to six years.

By 2022, some 150 firms were applying AI to small-molecule design alone. The capital flowing into the sector reflects confidence in the approach: Xaira launched in 2024 with $1 billion, while Alphabet's Isomorphic Labs raised $600 million in 2025. The technology works because molecule design is fundamentally a chemistry and physics problem with firm rules and vast data sets—precisely the conditions where machine learning excels.

Where the money actually goes

The problem is that designing molecules was never the expensive part of drug development. The roughly $2.6 billion required to bring a single drug to market is dominated by clinical trials and the failures within them. A Phase 1 safety study averages about $4 million, Phase 2 efficacy trials cost around $13 million, and Phase 3 confirmatory trials run $20 million or more.

According to a Boston Consulting Group analysis of approximately two dozen AI-discovered molecules in clinical trials, AI-designed drugs succeed in Phase 1 safety tests 80 to 90 percent of the time—well above the historical 50 percent rate. But in Phase 2, where drugs face their first real test of whether they work in patients, success rates drop back to the industry's typical 40 percent.

BCG concludes that AI roughly doubles the end-to-end odds of a drug reaching market, from 5–10 percent to 9–18 percent. But every percentage point of that gain comes from the cheap early stage. AI is saving money at the front of a process whose costs accumulate at the back.

Why efficacy remains the hard problem

The divide reflects a fundamental distinction. Designing a molecule involves chemistry and physics—questions with firm physical rules and enormous prior data libraries. Choosing the right biological target requires understanding the causal machinery of human biology, which remains mapped only in fragments with sparse, noisy data.

Drugs fail in two ways: the molecule is wrong, or the biological hypothesis behind it is wrong. AI has largely conquered the first problem while barely touching the second. The technology has made the cheap end of development cheaper and left the expensive, scientifically difficult end exactly where it was.

Early clinical results support this assessment. Insilico Medicine's rentosertib, an AI-discovered drug for idiopathic pulmonary fibrosis, posted positive mid-stage results in Nature Medicine in 2025. Against that stands Recursion's discontinued lead AI program in 2025 after efficacy signals failed to hold. The sector has thinned through mergers and shutdowns as markets begin recognizing that spectacular discovery success has not yet produced a single approved medicine.

Why it matters

For investors and pharmaceutical strategists, this analysis reframes where value actually lies in AI drug development. The real prize is not generating more molecules—a now-crowded space where returns are competed down—but using AI to attack Phase 2 and Phase 3 directly: validating which targets drive disease, finding biomarkers that predict patient response, and designing smarter trials. A company that reduces the efficacy failure rate would capture far more value than another molecule-design platform, because that failure rate drives the bulk of the $2.6 billion development cost.

The question of whether AI can reach into this harder problem remains genuinely open. Better data and new methods may enable AI to begin predicting efficacy, in which case the value dwarfs anything in discovery. Or the biology may represent a limit of knowledge that computation cannot shortcut, meaning much of the discovery boom is priced for a breakthrough it will never deliver.

This analysis was first reported by Michael A. Santoro writing for ProMarket.