Somewhere between a promising compound and a working drug lies a graveyard. Most molecules that look brilliant in a dish — that bind beautifully to their target, that show real potency in early screens — quietly die in the body. They get shredded by liver enzymes, refused entry by the blood-brain barrier, or accumulate somewhere they shouldn’t until they become toxic. This is the ADMET problem: absorption, distribution, metabolism, excretion, toxicity. It has eaten more drug development budgets than any other single obstacle in pharmaceutical history. And AI is finally getting serious about cracking it.
For decades, predicting ADMET properties meant building shallow models on small, proprietary datasets — rules of thumb dressed up as science. Lipinski’s Rule of Five, introduced in 1997, is still quoted in introductory pharmacology courses. It is, by the standards of modern machine learning, a five-line heuristic. Useful, but nowhere near sufficient. The space of drug-like molecules is estimated at 1060; a handful of physicochemical rules cannot navigate it.
What’s changed is the convergence of three things: enormous datasets of experimental ADMET measurements accumulated across decades of high-throughput screening, graph neural network architectures that treat molecules as what they actually are (graphs of atoms and bonds, not flat strings), and pre-training strategies borrowed from large language models that let those networks absorb chemical intuition before they ever see a toxicity label. The result is a new class of models that don’t just predict whether a molecule clears the liver — they reason about the structural features driving that outcome, and they generalize across chemical scaffolds in ways that earlier QSAR models simply couldn’t.
Insilico Medicine, Recursion Pharmaceuticals, and a wave of smaller computational chemistry startups are building ADMET prediction into the earliest stages of the design loop. Instead of optimizing purely for target binding and then checking pharmacokinetics later — the traditional, expensive sequence — they’re running multi-objective optimization from the start, shaping molecules to be potent and metabolically stable and orally bioavailable simultaneously. Microsoft’s MoleculeNet benchmarks and the more recent Therapeutics Data Commons have given researchers a common scoring surface to measure progress against, and the numbers have moved substantially in the last two years.
The harder frontier is toxicity — specifically, predicting rare but catastrophic effects like hERG channel blockade (which causes fatal cardiac arrhythmia) or idiosyncratic liver toxicity that only manifests in a subset of patients with particular genetic variants. These are sparse-data problems: the failure cases are rare by definition, and the underlying biology is complex enough that even mechanistic models struggle. But multi-task learning — training a single model on dozens of toxicity endpoints simultaneously so that signal from abundant endpoints reinforces learning on sparse ones — is genuinely improving performance here. And active learning pipelines are letting wet-lab experiments be targeted at exactly the uncertain regions of chemical space where the models need more data.
None of this means the attrition crisis in drug development is solved. Clinical failure for reasons that have nothing to do with pharmacokinetics — patient selection, flawed target hypotheses, manufacturing challenges — is its own enormous problem. But if AI can reliably eliminate the compounds that were always going to fail in the body before anyone runs a single mouse experiment, the economics of early discovery change substantially. Fewer resources chasing dead ends means more shots at the ones that might actually matter.
The molecules that survive the body’s gauntlet have always been selected by expensive, painful trial and error. We’re approaching the point where that selection happens first in silico, at the speed of inference, across libraries of millions of candidates. That is not a small thing. That is potentially a restructuring of how medicine gets made.