How AI Is Changing Drug Discovery, and Where the Hype Outruns the Evidence

Artificial intelligence is doing real work in the earliest stages of drug discovery. It helps rank which biological targets are worth pursuing, it generates and scores candidate molecules faster than medicinal chemists working alone, and it predicts some properties of those molecules before anyone runs an assay. What it has not yet done is change the part of the process that decides whether a drug reaches patients: proving a compound is both safe and effective in humans. The honest summary is that AI has compressed the front of the pipeline and left the expensive, failure-prone back half largely intact. Most of the current debate is a fight over how much that front-end compression is worth.

What AI genuinely contributes

Three contributions are supported by evidence rather than by press release.

The first is target and structure work. Tools that predict protein structures now do so at a scale and speed that was impossible a decade ago, which gives chemists a starting model for how a drug might fit a binding pocket. This is a real acceleration, and it has hard edges. Predicted static structures are strongest for rigid, well-behaved proteins and weaker where the biology is dynamic, where large conformational changes occur, or where membrane proteins sit inside a lipid environment the model does not fully represent. A predicted structure is a hypothesis about shape. It is not proof that engaging the target changes the disease, which is a separate and much harder question that structure prediction does not answer.

The second is molecule generation and optimization. Generative models can propose large numbers of candidate structures and rank them against properties chemists care about, such as predicted binding, solubility, or the presence of chemical groups associated with toxicity. Used well, this narrows a search space that used to be explored by slower, more manual iteration. The models are strongest where they have abundant, clean training data and weakest where they are asked to extrapolate into chemistry that looks nothing like what they were trained on.

The third is property prediction, often grouped under ADMET, meaning absorption, distribution, metabolism, excretion, and toxicity. Predicting these earlier can help teams drop weak candidates before they consume years of laboratory work. The value is real, but the predictions are correlational and improve the odds of picking a good molecule rather than guaranteeing one.

Across all three, the pattern is consistent. AI is good at generating, ranking, and prioritizing hypotheses. It is not a substitute for the experiments that test them.

Where the claims outrun the results

The most cited evidence for AI in drug discovery is a set of clinical success rates, and reading them carefully matters because the headline and the footnote point in different directions.

An analysis of AI-discovered molecules that reached trials reported Phase 1 success rates in the range of 80 to 90 percent, well above the historic industry average of roughly 40 to 65 percent. That is a striking number, and it is the one most often quoted. Phase 1 mainly tests safety and tolerability, so a plausible reading is that AI is genuinely good at designing molecules with clean, drug-like properties that behave predictably in early human exposure.

The same analysis reported Phase 2 success rates of around 40 percent, on a small sample, in line with historic averages. Phase 2 is where efficacy is tested, where a drug has to actually change the disease. This is the stage where most drugs have always failed, and the early AI cohort is failing at a similar rate. The advantage that shows up so clearly in Phase 1 appears to fade exactly where the hardest question begins.

Two cautions belong next to those numbers. The sample of AI-originated drugs deep into trials is still small, so the Phase 2 figure could move in either direction as more programs report. And as of the most recent public data, no drug discovered or designed primarily by AI has completed the full path to regulatory approval. Individual high-profile candidates have been discontinued after longer-term data failed to confirm earlier signals, which is a normal part of development and a useful reminder that early promise is not late-stage proof.

The reason the back half of the pipeline resists compression is biological, not computational. A model can propose a molecule that binds a target beautifully. Whether hitting that target meaningfully changes a disease in a heterogeneous human population, without unacceptable harm, is a question that only a well-designed trial can answer. That is the wall, and no current AI clears it.

The evaluator stance

The framework that governs human testing has not been softened by any of this. Good Clinical Practice, the ICH E6 standard, was substantially revised, with E6(R3) reaching its final step in early 2025 and adoption following across major regions through that year. The revision explicitly modernizes trials around risk-based thinking, digital health tools, and data-driven oversight, but it does not lower the bar for evidence. The companion guidance on trial design and on statistical principles, ICH E8 and E9, still applies. A trial that leans on AI to enroll patients, monitor sites, or analyze data is still a trial, and it still has to pre-specify what would count as success and defend that analysis to a regulator.

For anyone assessing a claim, a few questions separate substance from marketing. What exactly did the model do, and at which stage. Is the reported advantage measured against a fair baseline, or against a flattering one. Is the endpoint a computational score, a preclinical readout, or a real clinical outcome in humans. And crucially, has the same rigor been applied to the failures as to the wins, because a platform that only reports its survivors tells you very little.

AI has earned a durable place in early discovery. It makes the search for candidates faster and, in some hands, smarter. What it has not done is repeal the arithmetic of clinical development, where efficacy and safety in real people remain the deciding evidence. Treating AI as a powerful accelerant for the front of the pipeline is accurate. Treating it as a shortcut past the trials is not.

This article is educational and is not medical advice.