Protein stability prediction improves using shared transformer decoder
StabilityArc: Decoding Protein Sequence Embeddings into Generalizable Stability Landscapes
Machine Learning
Summary
Predicting how changes in protein sequences affect stability is challenging because each protein behaves differently, yet underlying biochemical rules exist. The authors developed StabilityArc, a model that uses a shared decoder trained on many proteins to better predict these stability changes in unseen proteins. It outperforms previous models by interpreting protein sequence embeddings and considering interactions between mutations. This approach can guide initial experiments to focus on promising protein variants.
What this means in practice
- •For biotech engineers: Use StabilityArc to prioritize protein variants for lab tests, reducing costs by informing early stability predictions across diverse proteins.
- •For bioinformatics developers: Integrate StabilityArc’s decoder to improve computational pipelines that predict protein mutation stability without extra protein-specific training.
Authors
Aaron L. Feller, Andrew D. Ellington, Claus O. Wilke
Abstract
Every protein has a unique stability landscape, but the physical consequences of mutation are governed by recurring biochemical constraints. We test whether a shared decoder, trained on measurements from diverse proteins, can interpret these constraints in an unseen target, enabling cross-protein transfer for initial experimental round prescreening. We present StabilityArc , which maps frozen ESMC-600M residue representations through a shared RoPE transformer to an Lx20 matrix of substitution effects; a symmetric, contact-aware residual aids in predicting epistasis in simultaneous substitutions. In 66 strict leave-one-protein-out evaluations covering 134,794 ProteinGym variants, StabilityArc achieves 0.7134 Spearman correlation, exceeding the strongest zero-shot baseline, ProSST-2048 (0.6526), by 0.0608. We further explore the utility of this method by providing the score as a prior for Kermut, achieving Spearman correlation of 0.8280 across three supervised split schemes, improving on Kermut's reported 0.8167.