eSIG-Net: Accurate prediction of single-mutation induced perturbations on protein interactions using a language model

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Protein interactions are crucial for cellular functions.
• Single mutations can disrupt these interactions, leading to diseases.
• Predicting mutation effects on protein interactions is computationally difficult.

Purpose of the Study:

• To develop a novel deep learning model for predicting the impact of single mutations on protein interactions.
• To introduce eSIG-Net (edgetic mutation Sequence-based Interaction Grammar Network), an "Interaction Language Model" that uses only sequence information.

Main Methods:

• eSIG-Net integrates diverse protein sequence embeddings.
• A mutation-encoding module incorporates syntax and evolutionary data.
• Contrastive learning is utilized to assess mutation-induced interaction alterations.

Main Results:

• eSIG-Net demonstrates superior performance compared to state-of-the-art sequence-based and structure-based methods.
• The model accurately predicts mutational impacts on protein interactions.
• eSIG-Net successfully identifies causal variants and clarifies their functional roles.

Conclusions:

• eSIG-Net is a pioneering "interaction language model" for predicting mutation-driven interaction changes using sequence data.
• The model shows high accuracy and generalizability across different biological contexts.
• This tool aids in understanding disease mechanisms driven by protein interaction perturbations.