eSIG-Net: an interaction language model that decodes the protein code of single mutations.

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• Proteins interact with other molecules to perform functions.
• Predicting the impact of mutations on protein interactions ('protein codes') is a significant computational challenge.
• Current methods struggle to accurately predict mutation-driven interaction changes.

Purpose of the Study:

• Introduce eSIG-Net, a novel language model for predicting mutation effects on protein interactions.
• Develop a mutation-centric model that utilizes sequence information alone.
• Improve the accuracy and mechanistic understanding of protein interaction perturbations.

Main Methods:

• Developed eSIG-Net (edgetic mutation sequence-based interaction grammar network), a language model.
• Integrated protein sequence embeddings with syntax-aware and evolution-aware mutation encoding.
• Employed contrastive learning to train the model for predicting mutation-driven interaction changes.

Main Results:

• eSIG-Net outperforms existing state-of-the-art sequence-based and structure-based prediction methods.
• The model successfully nominates causal variants responsible for altered interactions.
• eSIG-Net provides valuable mechanistic insights into mutation impacts.

Conclusions:

• eSIG-Net is an effective mutation-centric interaction language model.
• The model accurately predicts interaction-specific network rewiring from sequence data.
• eSIG-Net demonstrates generalizability across diverse biological contexts.