How accurately our designs hit the site they were aimed at
A protein binder is designed against a specific patch on a target — the site that matters biologically. A design that binds somewhere else is not a near miss; it is the wrong molecule. This measures how often ours land where they were sent, against an external reference we did not create.
The reference is the Adaptyv Bio EGFR round 2 release: 380 third-party designs whose binding was measured in a wet laboratory. Because the measured binders show where a working binder actually contacts this target, they provide a standard our own designs can be held to.
The result, and its limit
On targets in the range our platform is suited to, our designs engage the requested binding site as accurately as designs that were independently confirmed to bind. Outside that range, accuracy falls away — on the smallest targets and on the largest, it reaches chance or below.
| Target size | Targets | Accuracy vs chance | Reliable |
|---|---|---|---|
| Under 150 residues | 7 | 0.9× | no |
| 150–299 residues | 18 | 5.3× | yes |
| 300–599 residues | 6 | 11.5× | yes |
| 600 residues and above | 4 | 0.2× | no |
Every one of the 24 targets between 150 and 599 residues performed above chance. Seven of the eleven outside that range did not. We publish the range because it is a real limit, and because a customer choosing a target deserves to know it before committing compute rather than after.
The large-target limit is ours, not the problem's. On this same 621-residue target, third-party designs reached the site at 8.4× while ours reached 0.5×. Large targets are demonstrably addressable; our pipeline does not yet address them well. That is work in progress, not a boundary of what is possible.
What this does not say
Hitting the right site is necessary for a binder to work. It is not sufficient, and it is not a prediction of binding. In the same analysis we tested whether this measure predicts which designs bind, and it does not — no better than the structural-confidence scores we already publish, which is to say not reliably. We never report an affinity from a computational run, and this changes nothing about that.
This is also one target. The range above is drawn from 35 targets across our own catalogue, but the external standard it is calibrated against is EGFR alone.
Method
For each folded complex we take the residues the binder contacts within 5 Å, and compare them with the residues the design was aimed at. Accuracy is reported against chance, because a larger requested site is easier to hit: a 46-residue site on a 510-residue protein is 9% of the surface, so landing on it 78% of the time is 8.6× what random contact would give. Without that correction the numbers would flatter whichever target asked for more.
Contains information from the Adaptyv Bio EGFR round 2 dataset, made available under the Open Database License (ODbL) v1.0: opendatacommons.org/licenses/odbl. Our ranking metrics for the same designs are published separately: beltroni_metrics_egfr_round2 ↗.