ProBASS - a language model with sequence and structural features for predicting the effect of mutations on binding...

Area of Science:

• Computational biology
• Biophysics
• Structural biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Mutations in PPIs can alter protein functionality and lead to diseases.
• Existing methods for predicting mutation-induced binding free energy changes (ΔΔG_bind) often lack precision.

Purpose of the Study:

• To develop a precise computational model for predicting the impact of mutations on PPI binding affinity.
• To leverage advanced protein language models (PLMs) for enhanced prediction accuracy.

Main Methods:

• Utilized ESM2 and ESM-IF1 protein language models to generate sequence and structure-based embeddings for PPI mutants.
• Fine-tuned a prediction model, ProBASS, using a large dataset of experimental ΔΔG_bind values.
• Validated ProBASS performance on datasets with varying mutation types and experimental conditions.

Main Results:

• ProBASS achieved high correlations with experimental ΔΔG_bind values (0.83 ± 0.05 for single, 0.69 ± 0.04 for double mutations on same PDB).
• Demonstrated strong predictive performance (0.81 ± 0.02) on a diverse dataset of 2325 single mutations across 132 PPIs.
• Outperformed existing state-of-the-art methods in predicting mutation effects on PPI binding affinity.

Conclusions:

• Integrating extensive experimental ΔΔG_bind data to refine pre-trained PLMs is a successful strategy for precise prediction.
• ProBASS offers a broadly applicable tool for predicting mutation effects on PPIs, aiding protein engineering and design.
• The model can be further improved with additional experimental data.