The Drug That AI Designed: What the First Genuine Cases Reveal

Drug discovery has a famous problem: it takes about twelve years and $2.5 billion to bring a new drug from initial concept to regulatory approval, and roughly 90% of drug candidates that enter clinical trials fail. These numbers have barely moved in three decades despite vast investment in technology. The industry’s response to each new tool — combinatorial chemistry in the 1990s, high-throughput screening in the 2000s, genomics in the 2010s — has followed a consistent pattern: genuine progress in early-stage candidate identification, followed by the sobering discovery that the hard part is biology, not chemistry.

AI is following the same pattern, but with greater speed and somewhat better early results.

The INS018_055 Case

Insilico Medicine’s INS018_055 is the most-cited example of an AI-designed drug because it is the furthest along and the most thoroughly documented. The target was a protein called TNIK (TRAF2 and NCK-interacting kinase), implicated in idiopathic pulmonary fibrosis — a fatal lung disease with limited treatment options. The target identification was AI-assisted: a graph neural network analyzing disease mechanism data suggested TNIK as a plausible driver.

The drug design used a generative chemistry pipeline. The system proposed molecular structures predicted to bind TNIK, filtered by predicted ADMET properties (absorption, distribution, metabolism, excretion, toxicity), and generated a ranked list of candidates. The top candidate — what became INS018_055 — was synthesized and tested experimentally. It showed the predicted binding affinity. It showed acceptable toxicity profiles in animal models. It entered Phase I clinical trials in 2022 and Phase II in 2023.

By early 2027, Phase II results are published. The drug is safe. It shows biological activity — measurable effects on fibrotic markers in patients. The effect size in Phase II was smaller than hoped, and the trial’s primary endpoint was not met. Whether it continues to Phase III depends on decisions that are made with incomplete information, as all such decisions are.

The honest interpretation: AI successfully identified a plausible target and designed a molecule that works in the predicted way, reached clinical trials on a compressed timeline (roughly 4-5 years from initial AI analysis versus the typical 12-15 years), and produced a drug that is biologically active but not, on current evidence, clearly clinically superior to existing treatments. This is not failure. It is a result — a real result, more informative than “it didn’t work” and less conclusive than “it’s approved.”

What the Broader Pipeline Looks Like

As of early 2027, there are approximately 23 drug candidates in clinical development that their sponsors describe as “AI-designed” or “AI-first.” The category is not well-defined — some involve extensive AI design of the compound itself, others use AI primarily for target identification with conventional chemistry downstream, and the marketing departments of pharmaceutical companies have incentives to overstate the AI contribution.

The most credible count of genuinely AI-designed molecules — where the computational system generated the structure rather than simply screening existing compound libraries — is closer to eight. Of these, two have Phase II data (INS018_055 and a separate cancer immunology candidate from Exscientia/Sanofi that was terminated after Phase II for insufficient efficacy). Six are in Phase I or transitioning to Phase II.

The termination of the Exscientia/Sanofi candidate is worth dwelling on. The drug was designed to modulate adenosine A2A receptors involved in immune checkpoint pathways in cancer. The Phase I data showed acceptable safety. Phase II, in a specific head-and-neck cancer indication, showed no significant clinical benefit. Exscientia’s CEO described this as the system working correctly — finding molecules with predicted properties that turned out not to be clinically sufficient — rather than as an AI failure. This is technically accurate and a good illustration of what AI can and cannot do: it can find molecules with predicted biochemical properties; it cannot guarantee that those biochemical properties translate to clinical benefit in a complex disease.

The Translation Problem Is Not an AI Problem

The 90% failure rate in clinical development is not primarily caused by poor drug design. It is primarily caused by the gap between what a drug does in controlled laboratory and animal settings and what it does in the messy, heterogeneous, co-morbidity-ridden bodies of real human patients.

This gap has multiple dimensions. Animal models for most diseases are poor predictors of human outcomes — particularly for diseases involving the immune system, the brain, and the gut microbiome, where species differences are large. The patient populations enrolled in early clinical trials may not represent the populations who will eventually receive the drug. The outcome measures used in trials may not capture what actually matters to patients. The dose-response relationship in humans may differ from predictions based on in vitro or animal data.

AI addressed exactly none of these problems. The computational tools that design better molecules cannot tell you whether that molecule will work in a human patient who has the disease you’re targeting plus diabetes, hypertension, and a gut microbiome that metabolizes your compound unexpectedly. The tools have improved early-stage candidate quality — there is reasonable evidence that AI-designed molecules have better predicted ADMET profiles than historically average clinical candidates — but the translation problem is downstream of everything AI currently does.

Where the Real Gains Are Hidden

The less glamorous and more impactful contribution of AI to pharmaceutical development is not in designing first-in-class molecules for novel targets. It is in the unglamorous work of optimization: taking a compound that is known to work and making it better.

Medicinal chemistry optimization — adjusting a lead molecule to improve its metabolic stability, increase its selectivity for the intended target over related targets, reduce its toxicity, improve its oral bioavailability — is the craft that most actual drug discovery consists of. It is iterative, expensive, and labor-intensive. AI tools for predicting the effects of structural modifications on each of these properties have materially improved the efficiency of this process at multiple major pharmaceutical companies.

This doesn’t produce press releases. “We optimized our lead compound 20% faster than we would have otherwise” is not a headline. But if it scales across the industry — and the evidence from internal reports and occasional publications suggests it is — the aggregate impact on development timelines and costs may be larger than any number of AI-first discovery programs that make it to clinical trials.

The twelve-year timeline is not going to become three years because of AI. The places where AI will shave off years are at the front (faster identification of plausible targets and leads), at the optimization stage (faster medicinal chemistry iteration), and potentially at the biomarker stage (better patient stratification for trials). The middle — the biology, the animal models, the clinical trial logistics — remains stubbornly time-consuming, for reasons that have nothing to do with the availability of a good generative model.