From a map of interactions to a designed lead.

The first decision is not chemical. It is which two protein modules are worth bringing together at all — and that is a question about disease biology, answered by searching functional interaction and dependency data rather than by starting from a molecule someone already has.

This is what separates event-driven design from occupancy thinking. A conventional programme asks how tightly a drug can sit on one target; we ask which two things a cell cannot afford to have joined. Choosing a pair whose functions do not back each other up means an escape route has to solve two problems at once, rather than routing around a single blocked node.

For each chosen module we work from experimentally determined structures of that protein bound to a small molecule, and read off two things: a chemical anchor that is known to engage it, and the direction in which a bridge could leave that anchor without burying itself against the protein.

The anchoring chemistry does not have to be invented — only bridged. Building from binding events that have already been observed rather than proposed means the uncertainty in a design is concentrated in the one part that is genuinely new. That is a deliberate narrowing: a molecule with one novel element can be reasoned about, while a molecule that is novel everywhere cannot be assessed at all.

The bridge is the designed element, and it is where the geometry becomes unforgiving. A connection is only legal if it can physically leave both anchors along a route that is open to solvent and span the distance between them without forcing the two proteins into each other. Those constraints are easy to state and impossible to satisfy by eye.

So the engine enumerates candidate bridges and tests them against that gate directly, rather than proposing a plausible-looking connection and hoping the proteins accommodate it. Most proposals fail here, and they should. A gate that rejects almost everything it sees is doing the work — the alternative is a pipeline that discovers the clash later, after the chemistry has been committed to.

Surviving designs are scored on interaction energy, calibrated against complexes whose interfaces have already been experimentally determined, and reported with the settings that produced them. Proteins are then allowed to move — side chains repack, the assembly relaxes — because a structure judged rigid is judged under conditions that do not hold.

Induced fit is priced rather than vetoed. A design that requires a protein to shift is not disqualified for it; the cost of that shift is carried into the score, where it can be weighed. The validation standard behind this is set out in full on the science page →

Every pose the engine evaluates — including the failures — becomes a row of training data relating geometry to energy. This is the step that turns a calculation into a platform. The near-misses are the most informative rows in the set, and they are exactly what the structural record cannot supply, since it contains only the complexes that formed.

Trained on that record, the model begins to predict productive designs beyond the ones enumerated by hand. The enumeration teaches the model; the model then reaches molecules the enumeration would never have been run on. That is the difference between a tool that answers a question and a system that learns the grammar behind it.

What leaves the engine is a ranked set of specific, synthesisable molecules, each carrying the provenance behind its ranking — what was assumed, what was calculated, and under which settings. A shortlist without that record is a list of names.

We state the limit of these plainly: they are computationally designed and ranked candidates, not confirmed binders. Binding data is the gating milestone, and no amount of further scoring substitutes for it. The claim is that the search space was covered properly and the ranking is defensible — which is what a synthesis programme actually needs from a design engine, and considerably more than a promising hit.