Using protein language models for protein interaction hot spot prediction with limited data

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Biology

Background:

• Protein language models (PLMs) are inspired by human language models for analyzing protein sequences.
• PLMs show promise in predicting protein structure, function, and mutation impact.
• The use of PLMs for predicting residue properties from limited data, like protein-protein interaction (PPI) hotspots, was unexplored.

Purpose of the Study:

• To investigate the feasibility of using PLM-derived representations as features for predicting PPI-hotspots.
• To evaluate the performance of PLM-based methods against traditional sequence and structure-based approaches.

Main Methods:

• Utilized a dataset of 414 experimentally confirmed PPI-hotspots and 504 PPI-nonhot spots.
• Employed unsupervised learning with PLMs to extract features from protein sequences.
• Applied machine learning models using PLM-learned representations to predict PPI-hotspots.

Main Results:

• PLMs successfully capture critical functional residue attributes from evolutionary sequence information.
• PLM-based methods achieve competitive performance compared to sequence and structure-based features for PPI-hotspot prediction.
• An optimal feature set was identified, balancing information gain and preventing overfitting.

Conclusions:

• Transformer-based PLMs can extract valuable knowledge from sparse datasets, particularly for challenging tasks like PPI-hotspot prediction.
• PLMs provide a cost-effective and time-efficient alternative to experimental methods for predicting certain residue properties.
• Further research is needed to understand the interpretability of features driving residue property predictions.