AI drug design and AlphaFold applications

The idea underneath structure prediction is older than the machine learning that finally exploited it. If two residues touch in a folded protein, a mutation at one that destabilises the contact is tolerated only when a compensating mutation appears at the other. Over hundreds of millions of years and thousands of homologues, that constraint leaves a statistical trace: pairs of columns in a multiple sequence alignment that vary together. Read enough homologues and the contact map — and with it, roughly, the fold — is recoverable from sequence alone.

What AlphaFold actually does with that signal

AlphaFold2’s advance was not discovering coevolution but refusing to treat it as a separate step. Its Evoformer passes information back and forth between the alignment representation and a pairwise residue-residue representation, then a structure module places atoms directly in 3D and the whole network is recycled through its own output. At CASP14 in 2020 the median backbone accuracy crossed roughly 1 Å, comparable to experimental error for many targets, and the field’s forty-year benchmark problem stopped being open. AlphaFold 3 extended the same architecture, with a diffusion-based generator, to complexes containing nucleic acids, ions and small molecules.

Two properties limit what follows. The model emits a per-residue confidence, pLDDT, and it is well calibrated: low pLDDT usually means the region is genuinely disordered rather than badly modelled — a useful answer, but not a structure. And the output is a single static conformation, biased toward whatever state the alignment’s homologues were crystallised in. Proteins that matter pharmacologically are usually the ones that move: allosteric switches, cryptic pockets, active/inactive kinase states. A prediction that names one of those states does not tell you which one the drug should see.

The energy barrier

The harder constraint is thermodynamic. Binding affinity depends exponentially on free energy, so at body temperature an error of about 1.4 kcal/mol — a small fraction of a single hydrogen bond — moves a predicted affinity by an order of magnitude. Empirical docking scoring functions are not close to that accuracy and were never designed to be; they rank poses, they do not measure binding. Rigorous alchemical methods such as free energy perturbation do approach roughly 1 kcal/mol on congeneric series, but only with a correct starting pose, a well-sampled protein, and GPU-hours per compound. Structure prediction has made the geometry cheap and left the energetics expensive, which is why the practical gain is in prioritising which of a billion enumerated compounds to synthesise, not in eliminating synthesis.

Where the ceiling really sits

Generative chemistry has a related boundary: a model trained on known molecules reproduces the biases of known molecules, and proposals must still be synthesisable in a finite number of steps. But the deeper limit is that clinical attrition is not a chemistry problem. Most candidates fail because the target turned out not to drive the disease, or because of toxicity that no structure of the intended target predicts. Better molecules against a wrong hypothesis fail faster, not less often. Regulators have said as much in outline — the FDA’s discussion paper on AI/ML in drug development and the EMA’s reflection paper both centre on model transparency and validation rather than on accepting in-silico evidence in place of experiment.