Why Does the Hermes Binding Model Nickname a Protein?

Ian Quigley says the Hermes binding model nicknames a protein by memorizing a short variable stretch of amino acids, instead of looking at the binding pocket or the conserved domain.

Hermes has cross-attention, so a map can show which amino acids the model uses. The first hope was the binding pocket. A more cautious hope was every residue, because the model did not yet know what mattered. It did neither. It looked at short sequences away from the functional part and away from the stretch that looks similar across many proteins. Biologists use that conserved domain to group families and guess function. Hermes looked outside it, at non-conserved stretches where the letters vary more. He thinks the model can remember 10 or 12 amino acids that differ in that spot and skip the physics. That matches co-folding models that shove a molecule into a pocket after the pocket is broken.

The counter they tried is mutations at scale. Two proteins look almost the same, one small difference changes the behavior a lot, and a nickname for the variable stretch then predicts badly. After training on those mutants, including the one on the plot they were showing, Hermes spreads its attention much more evenly. The lesson is general, not one gene. The nickname strategy has to stop working.