Every technology wave builds its own vocabulary. Large language models gave the wider world "transformer," "tokenizer," and "attention" as household-adjacent terms within a couple of years. AI-driven biology is going through the same process right now, just with less mainstream visibility -- a specific, increasingly standard stack of foundation models that discovery teams compose together the way a web developer composes libraries. Here's a field guide to the pieces, and how they fit together.
Structure prediction: AlphaFold and its family
AlphaFold2 solved single-protein structure prediction from amino acid sequence alone -- a problem structural biologists had worked on for half a century -- to an accuracy competitive with slow, expensive experimental methods like X-ray crystallography and cryo-EM. AlphaFold3 extended the same underlying approach to predict how proteins interact with small molecules, DNA, RNA, and other proteins, which is the far more directly useful capability for drug design: knowing not just what a target protein looks like, but how a candidate drug might bind to it. Open alternatives like OpenFold and ESMFold have made comparable capability broadly accessible outside the labs that built the originals, accelerating adoption across the entire industry rather than concentrating it in one company.
Protein language models: ESM and the sequence-as-language idea
The core insight behind Meta's ESM family (culminating in ESM3, now developed independently at EvolutionaryScale) is that evolution has already run an enormous, natural experiment: hundreds of millions of protein sequences across every organism ever sequenced, each one either surviving natural selection or not. Train a large transformer model on that corpus the same way a language model trains on text, and it learns something like a grammar of what makes a protein sequence biologically viable -- well enough to predict the effect of a mutation, generate entirely novel functional proteins, or flag which regions of a sequence matter most for a specific function, often without ever being explicitly told the underlying biophysics.
De novo protein design: RFdiffusion
Where AlphaFold answers "what structure does this sequence produce," RFdiffusion -- from David Baker's lab at the University of Washington, built on the same diffusion-model mathematics that power image generators -- answers the inverse question: "design me a novel protein structure with this specific function," such as a binder that locks onto a chosen target with high affinity. This is genuinely generative design in three-dimensional space rather than in pixels or text tokens, and it's the technology most directly responsible for the recent explosion in de novo binder and biologic design -- proteins that don't exist in nature at all, created to order for a specific therapeutic purpose.
Small-molecule generative chemistry
The same diffusion and flow-based generative techniques get applied to small, drug-like molecules -- proposing novel three-dimensional molecular structures conditioned on the shape and chemistry of a target binding pocket, rather than generating pixels conditioned on a text prompt. Companies like Chai Discovery and Genesis Therapeutics have built platforms around this approach, aiming to generate synthesizable, drug-like candidates directly against a validated target structure rather than screening a fixed library of existing compounds.
Single-cell and genomic foundation models
Models like Geneformer and scGPT are trained on millions of single-cell transcriptomic profiles -- essentially, a snapshot of which genes are active in individual cells across many tissues and disease states. This lets researchers predict how a cell's gene expression would shift in response to a genetic or chemical perturbation without physically running the experiment first, which is directly useful for target identification: it helps narrow down which genes or pathways are most likely to matter for a given disease before committing wet-lab resources to test them.
How these compose into one pipeline
None of these models work in isolation in a real discovery program -- their real power shows up when they're chained together. A genomic foundation model narrows a disease down to a shortlist of plausible targets. A structure-prediction model resolves what those targets look like in three dimensions, including how they interact with candidate binders. A generative chemistry or protein-design model proposes molecules or biologics shaped to fit. Predictive ADME/Tox models triage those candidates before any of them reach a wet lab. Each stage compounds the value of the one before it -- which is exactly why the "AI drug discovery stack" increasingly looks like a connected pipeline rather than a single breakthrough model, and why teams that can move evidence cleanly between these stages, rather than re-deriving it by hand at every step, have a structural advantage.
And as the previous pieces in this series have argued, that same principle -- connected evidence beats isolated point tools -- doesn't stop at the discovery stage. It's just as true moving from preclinical characterization into first-in-human dosing, and from a single trial's interim analysis into the next one. The foundation models are the new building blocks. What you build with them still depends on keeping the evidence connected all the way through.
