Classifying stabilizing vs. destabilizing mutations instead of regressing ΔΔG
- Recasts mutation stability prediction as classifying whether a change stabilizes or destabilizes, rather than estimating exact energy values.
- Compares wild-type and mutant proteins through a shared-weight twin network, fusing local structural context around the mutation with protein language model features via cross-attention.
- Tested retrospectively on the independent S669 benchmark; no prospective experiments are described.
- Reported to beat the other methods evaluated, including on detecting stabilizing mutations.
💡 Points Based on the abstract
Would a binary stabilizing/destabilizing call actually help on my project? Maybe for triage, but I'd still want a magnitude to pick between candidates. S669 is the sensible test, yet check how much of it overlaps the training sets and whether the sparse stabilizing class is handled fairly. Without any measured follow-up, I'd treat this as a ranking filter to try against FoldX or Rosetta on our own mutants.
BlindDDG: A Blind Evaluation-Based Method for Protein Mutation Stability Prediction ↗