Exploratory finding · FLIP2 Hydro
Aggregate chemistry transfers.
Across three protein backbones.
Can a model learn from mutations in some proteins and still rank variants in a protein it did not train on? Here, a small model with 11–16 position-free chemistry features beat mutation count alone on all three held-out backbones.
The problem
Protein positions do not line up neatly.
The seven tested amino-acid positions sit in different structural contexts in these three proteins. Calling each one “site 1” or “site 7” makes the positions look comparable when they are not. The question was whether a small summary of overall chemical change could travel better.
Rank correlation · Spearman ρ
The chemistry model beats mutation count alone on every held-out protein.
The model is below the published best on FYN-SH3 and above the published table value on CspA and CI-2A. This is not an official leaderboard submission.
The correction
`VVVVVVV` was not the native sequence.
The first analysis counted non-valine residues. Canonical mapping showed that the all-valine sequence is an engineered anchor. The analysis was retracted and rebuilt with the native cores.
FLFFIIVVIIVLVIVIVVVFVMore than mutation count
Chemistry still separates variants with the same number of mutations.
Within each mutation-count group, all three permutation checks: p = 1/1001.
Why it may transfer
Keep the chemistry. Drop the position labels.
The seven randomized sites are buried, but “site 1” in one protein does not match “site 1” in another. The model summarizes chemical changes without pretending those positions line up.
Useful failure: adding naive position-specific 3D features reduced held-backbone performance in all three folds.