Molecules by Machine

In February 2020, a small British company called Exscientia announced that its AI system had designed a drug candidate for OCD — DSP-1181 — in twelve months. The standard timeline for that stage of research is four to five years. The story ran in every major newspaper. The word “breakthrough” appeared in most of them.

DSP-1181 entered Phase I trials in Japan. And then, in 2022, the program was quietly discontinued. Not because the molecule was toxic. Not because the AI had hallucinated chemistry. The compound simply didn’t perform well enough in early human data to justify continuation — the same fate that meets roughly 90 percent of all Phase I candidates regardless of how they were discovered. The AI had found a molecule. The molecule, like most molecules, lost.

That story, more than any triumphalist press release, tells you what you need to know about where AI drug discovery actually stands in 2026.

The Scorecard

As of mid-2026, somewhere between 70 and 90 AI-designed or AI-substantially-assisted drug candidates have entered human clinical trials, depending on how strictly you define “AI-designed.” The number itself is genuinely impressive — in 2019, it was effectively zero. The companies driving this include Insilico Medicine, Exscientia (now merged with Recursion Pharmaceuticals in a deal that closed in early 2025), Relay Therapeutics, Schrödinger, and a handful of large pharma internal programs at Novartis, AstraZeneca, and Pfizer.

Related reading

Insilico’s INS018_055, targeting idiopathic pulmonary fibrosis, is the most advanced of the purely AI-originated programs. It completed Phase IIa in late 2025 with results that were cautiously positive — meaningful reduction in decline of lung function compared to placebo, with a tolerability profile that kept the trial population intact. That’s not a cure. It’s a data point. Phase IIb enrollment is underway, and the field is watching the timeline closely because IPF has no approved disease-modifying therapy that actually slows progression in more than a fraction of patients.

Relay Therapeutics’ RLY-4008, a FGFR2 inhibitor targeting intrahepatic cholangiocarcinoma, reached Phase II with a 78 percent objective response rate in a heavily pretreated population — a number that would be exceptional by any standard of discovery, AI or otherwise. Relay’s approach leans heavily on dynamic structural biology (understanding proteins in motion rather than in static crystallized form), and AI tools are deeply embedded in their modeling pipeline. Whether you call RLY-4008 “AI-designed” or “AI-assisted” is largely semantic; the computational contribution was substantial and early.

Schrödinger’s SGR-1505, a MALT1 inhibitor for B-cell lymphomas, posted Phase I data in 2025 showing responses in a disease where options are limited. Phase II is enrolling. Separately, Schrödinger published internal benchmarks in 2024 showing that their physics-based FEP+ platform, combined with ML scoring, predicted binding affinities within 1 kcal/mol of experimental values for a diverse compound set — the kind of predictive accuracy that, if it holds in prospective programs, genuinely compresses experimental cycles.

Then there are the failures, which receive systematically less coverage. Exscientia’s EXS21546, a small molecule for head and neck cancer that was supposed to demonstrate AI’s capacity to address difficult binding sites, was discontinued after Phase I showed insufficient efficacy signals. Benevolent AI, which went public on Euronext Amsterdam in 2022 and then suffered a spectacular stock collapse through 2023, had its lead asset in atopic dermatitis (BEN-2293) fail Phase IIa. The company restructured aggressively and pivoted toward licensing its platform rather than running its own trials. Cyclica, a Toronto-based AI drug discovery company, shut down entirely in late 2024 after its licensing model failed to generate the revenue needed to sustain operations. The landscape has been tidied by attrition.

What “AI-Designed” Actually Means

The phrase is doing more work than it should. In most cases, AI tools contributed meaningfully to one or several stages of a multi-year, multi-team process. Generative models proposed candidate structures. Docking simulations screened out obvious failures. ADMET prediction tools (absorption, distribution, metabolism, excretion, toxicity) flagged compounds with unfavorable pharmacokinetics before they went into animals. Retrosynthesis AI mapped plausible synthetic routes so chemists could assess cost and feasibility.

None of this is the same as an AI “designing” a drug the way a human architect designs a building — with intentional choices at every step and accountability for the whole. The molecules that are in trials emerged from hybrid processes where AI contributed genuine signal, mostly in hit identification and lead optimization, and human chemists, pharmacologists, and clinicians made the decisions that governed which signals to pursue.

This distinction matters less for celebrating progress than for understanding where the technology fails. AI models are exceptionally good at interpolating within chemical spaces that are well-represented in training data. They are much less reliable when the target is genuinely novel — a protein with no known ligands, a disease mechanism that hasn’t been extensively studied in the literature — because the model has little to extrapolate from. The successes so far cluster around target classes that are well-characterized: kinases, GPCRs, ion channels. The hard frontier remains hard.

The most honest framing is that AI has become a standard tool in the medicinal chemist’s toolkit, deployed alongside traditional methods rather than replacing them. A 2025 survey of R&D heads at the top 20 pharmaceutical companies found that 18 of 20 had integrated at least one commercial AI platform into their discovery workflows, but only 4 described it as the primary driver of their lead generation strategy. The others used it as a component of a broader process, which is probably the correct epistemic position.

Speed vs. Novelty

Related reading

The strongest honest claim for AI in drug discovery is speed and cost reduction in the early phases, not fundamental novelty of output. Insilico ran from target identification to IND-enabling studies in INS018_055 in roughly 30 months. A conventional program might take 5-6 years to reach the same point. That compression is real. If it holds across a larger portfolio, it represents a meaningful advance for patients who are waiting for treatments.

The weaker claim — that AI will find drugs humans never could, exploring regions of chemical space too vast for traditional methods — remains theoretically appealing and empirically unproven at the clinical level. Generative models can certainly propose structures outside known pharmacology. Whether those structures survive the brutal empirical filter of biology, toxicology, and human physiology is a different question, and one that trials are still answering.

One telling data point: the drugs showing the best early clinical signals are almost uniformly in indications where there was already rich prior data. Oncology, pulmonary fibrosis, hematologic malignancies — these are areas saturated with published biology, clinical trial results, and structural data. The AI is, in a meaningful sense, standing on decades of human scientific work. That’s not a criticism; it’s an observation about where the leverage actually lives. The genuinely novel chemical spaces where AI might find drugs no human ever considered require training data that doesn’t exist yet, because the biology in those spaces hasn’t been thoroughly explored experimentally.

The Comparison Problem

Evaluating AI drug discovery against conventional methods is genuinely difficult because the counterfactual doesn’t exist. Nobody runs the same target, in the same organization, with the same resources, once with AI and once without. What the field has instead is historical comparison — AI programs reaching Phase I in 18-30 months versus historical averages of 5-6 years — and portfolio statistics, which are still too small to be reliable.

The attrition rates so far don’t look dramatically different from historical norms. Of the roughly 70-90 candidates that have entered trials, approximately 60 percent are still active as of mid-2026. That’s consistent with early-stage pharma attrition broadly, not dramatically better. If AI’s primary contribution were improving candidate quality — reducing Phase I failures because the molecules were better selected and had fewer liability flags — you’d expect to see a measurable improvement in success rates relative to historical baselines. That signal hasn’t emerged clearly yet, though the dataset is still too small to be conclusive.

What has emerged is that AI materially shortens the pre-clinical timeline, which means companies can run more shots on goal with the same resources. Whether more shots at the same hit rate beats fewer shots at a higher hit rate is a portfolio optimization question that will take another decade to resolve empirically. The economics of drug development, where Phase III failures cost $300 million or more and dominate total R&D budgets, mean that improving pre-clinical speed matters much less than improving Phase II and III success rates. AI hasn’t demonstrated that yet.

The Data Quality Problem

One thing the industry doesn’t discuss openly enough: the quality of the data AI models are trained on is variable and often worse than the headline database sizes suggest. ChEMBL, the largest public bioactivity database, contains roughly 17 million activity measurements across 2 million compounds, assembled from published literature. That sounds comprehensive. It isn’t. Published assay data suffers from systematic biases — results that work are published, results that don’t aren’t; compounds from a small number of privileged scaffolds are heavily overrepresented; assay conditions that generate clean data are overrepresented compared to messy biology. A model trained on this data learns the biases as well as the signal.

Proprietary company databases are larger and less biased in some respects, but they’re siloed. The competitive dynamics of pharmaceutical R&D mean that the companies with the best data — the ones who’ve run the most assays over the most diverse compound collections — share very little of it publicly. This creates an asymmetry: AI drug discovery startups that lack proprietary data are training on biased public data, while large pharma programs with massive internal datasets can build more reliable models.

Related reading

The Next Phase

The genuinely interesting frontier in 2026 is not small molecules — it’s biologics and, specifically, protein-based therapeutics. AlphaFold and its successors have changed what’s possible in antibody engineering and enzyme design, and several programs using AI-designed protein scaffolds have entered IND-enabling studies in late 2025 and early 2026. These don’t yet have Phase I data, but they represent the category where AI’s structural prediction capabilities translate most directly into experimental design. The argument that AI can find proteins with properties that evolution never optimized for is more defensible than the argument that it can find small molecules beyond known pharmacology.

The other frontier is combination — using AI not just for molecule design but for patient stratification, trial design, and biomarker identification. Several of the Phase II programs showing the strongest signals are using AI-derived biomarkers to select patient populations likely to respond. That’s arguably where computational tools are contributing the most clinical value right now, even if it generates fewer headlines than “AI designs drug.”

The machine has produced molecules that are in human bodies, doing biology. That is a real milestone. What it hasn’t yet done is break the fundamental economics of drug development — the long timelines, the brutal attrition, the enormous cost of late-stage failure. The scorecard at the end of this decade will be more informative than the one at mid-2026. For now: genuine progress, appropriate skepticism, and a lot of Phase II data still to come.