Most of this site describes what our model represents. This page describes what it does, the framework that takes a question about disease biology and returns a specific molecule, ranked and ready to synthesise. The order matters more than any single step: we let the proteins move before we ask what can join them, because a search run on rigid structures has already excluded most of the answers.
We don't screen for luck. We compute the molecule that has to work, then let the model generalise it.
Six steps, from which two proteins, through what can physically join them, to a molecule worth making.
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 known to engage it, and the set of directions 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 concentrates the uncertainty 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.
This is the step most pipelines treat as a refinement and we treat as the foundation. Backbone motion, side-chain repacking, local adjustment and anchor rotation are run as one inseparable procedure over the entire landscape, not a filter applied to a shortlist, and not a top-ranked sample. Every composite state the procedure produces is kept.
The reason is a lesson we learned the expensive way. A triage over a handful of arrangements is not a verdict about the rest. Sample first and flex second and the flexibility only ever sees what the rigid search already liked, which is precisely the set least in need of it. Repacking is carried forward as coordinates rather than used as a veto, because a rejection made under rigid rotamers is not a result.
Two commitments fall out of this. Flexing each partner on its own yields a rigidity profile, how much each one's interface actually moves, and a plastic partner is docked onto a rigid one, not the reverse. And each partner is evaluated against a field built from its own flexed coordinates: freezing one side while the other moves makes the calculation depend on which protein you happened to call first, which is a correctness bug rather than an approximation.
Connectability is a topological question, not a matter of two chemical groups happening to face each other. What matters is whether the exit directions available to each anchor can cross or touch once the proteins are arranged, an adapter absorbs the rest. Searching only for anchors already pointing at one another manufactures false negatives, and the shortest bridge worth considering is set by the direct exit-to-exit distance rather than by any assumption about geometry.
That splits the design space into two branches rather than one ladder: a linker spanning a gap, and a fusion in which the two anchor bodies coincide. They fail differently, burial of an exit blocks fusion while leaving the linker branch open, so they are searched as complements. Most proposals fail this gate, and they should. A gate that rejects almost everything is doing the work; the alternative is discovering the clash after the chemistry has been committed to.
A proximity drug does not act on two proteins. It acts inside a machine, target, effector, the scaffolding that holds them, and the designed molecule between. Treating the problem as two bodies and a tether ignores the parts of that machine most likely to be in the way.
So the design module works in the full multi-body assembly: bridge geometry and fusion chemistry are proposed against every component present, with the surrounding machinery treated as excluded volume rather than left out of the picture. A bridge that is legal between two isolated partners and illegal inside the assembly is a bridge you want to find now, not after synthesis.
Status, plainly: this module is built and queued. It runs once the current exhaustive full-flexibility sweep completes, that sweep supplies the landscape it operates on, and running it earlier would only reproduce the sampling error the previous step exists to eliminate.
The obvious way to rank proximity designs is by computed interaction energy. We moved away from it, and the reason is worth stating: at the close contacts these molecules are built to create, an interface energy is dominated by steric penalty and by exactly where the anchor was placed. It responds sharply to things that a modest repack would relieve, and it is quietest about the thing we actually care about, whether the whole assembly holds together.
So the ranking question became a structural one. Given only the designed molecule and the sequences of the proteins it is meant to join, does an independent structure-prediction model rebuild the assembly the molecule was designed to produce? A design that reconstitutes its intended architecture from scratch has demonstrated something an energy number cannot: that the arrangement is recoverable without being told the answer. One that does not is telling us something too.
Two disciplines make this a measurement rather than a vibe. We rank on agreement with the designed architecture, not on a predicted affinity, because the affinity heads of current co-folding models score a cropped pocket and cannot see a multi-body assembly at all. And the instrument is control-gated like any other: it must rebuild solved complexes we already know before its verdict on an unsolved one carries weight, and it must be shown to separate designs that differ only in their bridge. A model that scores everything highly has not ranked anything.
What leaves the engine is a ranked set of specific, synthesisable molecules, each carrying the provenance behind its ranking. We state the limit plainly: these 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 validation standard behind all of this is set out on the science page →
Six steps, run the same way on every target pair, which is what makes the seventh programme cheaper than the first.
Biology picks the pair. Two modules whose functions don't back each other up, so escape has to solve both at once.
Anchors and their available exit directions read off experimentally determined structures, so novelty sits in one place.
Backbone, side chains and anchor rotation moved together across the whole landscape. Never a top-ranked sample.
A topological test on exit directions, searched as two branches, linker and fusion, because they fail differently.
Bridge and fusion designed inside the full multi-body complex, with surrounding machinery as excluded volume.
Ranked on whether an independent model rebuilds the intended assembly, not on a predicted affinity.
We would rather publish the state of the work than a version of it that has aged well in a slide.
The full-flex procedure of Step 03 is executing across the retained landscape on both programmes. It replaces an earlier sampled approach that we retired after auditing how narrow it was, and it is the input every later stage depends on.
The Step 05 design module, geometry- and flexibility-driven assembly with linker and fusion design in full multi-body context, is built and waiting on the sweep above. No results are claimed from it.
A pipeline for assemblies of three proteins held by one designed molecule, one body beyond everything in the clinic. Each additional body multiplies the arrangements to enumerate and adds interfaces that must be compatible at the same time, while being excluded volume for the rest, so the search was rebuilt rather than extended. The flexed-coordinate treatment, the gates and the controls carry over unchanged. It is producing designs and a provisional application is in preparation; there is no binding or cellular data behind it and none is claimed.
The Step 06 ranking instrument is being qualified against solved complexes before it ranks anything unsolved, recovery on knowns first, then a demonstration that it separates designs differing only in their bridge.
Every arrangement the engine evaluates, including the rejections, is retained as training data for a learned representation of interaction geometry. It runs alongside the design pipeline rather than gating it, and nothing in the programmes depends on it yet.
The governing framework, its measured constants and its claim discipline are versioned and in force, with positive controls on solved complexes passing under the same settings as the work they validate.
Why we publish it this way: a design engine is judged on whether its verdicts hold, and a verdict is only as good as the sampling behind it. Saying which stage is running and which is queued costs us nothing we would want to keep, and it is the same standard we apply internally, where an insufficiently sampled negative is recorded as not searched to depth, never as not there.
A design that works once may be luck; a design process that runs target after target is an engine. If you evaluate computational chemistry for a living, the parts worth interrogating are the sampling depth, the gates and the calibration, and those are the parts we would rather discuss early than late. Provisional patent applications covering the platform have been filed, and further provisionals covering a four-body assembly architecture and its design pipeline, and partner-directed designs, are in preparation.
See what it is running on →
·
Talk to the team →