PatchProt: hydrophobic patch prediction using protein foundation models.

Area of Science:

• Computational biology
• Protein structure prediction
• Machine learning in bioinformatics

Background:

• Hydrophobic patches on protein surfaces are crucial for interactions and implicated in disease.
• Predicting these patches from protein sequences is a significant computational challenge.
• Foundation models and multitask deep learning offer potential solutions for data gaps and improved prediction.

Purpose of the Study:

• To develop a novel method for predicting exposed hydrophobic patches on protein surfaces.
• To leverage a leading large language model (Evolutionary Scale Models - ESM-2) and parameter-efficient fine-tuning.
• To enhance model representation by incorporating related local and global prediction tasks.

Main Methods:

• Utilized the Evolutionary Scale Models (ESM-2) foundation model.
• Employed a parameter-efficient fine-tuning approach for efficient model training.
• Integrated multitask deep learning, training on local (residue) and global (protein) tasks.

Main Results:

• Developed PatchProt, a model that accurately predicts hydrophobic patch areas.
• PatchProt outperforms existing methods in predicting primary tasks like secondary structure and surface accessibility.
• Training on related local tasks demonstrably improves predictions for more complex global tasks.

Conclusions:

• Fine-tuning foundation models with multitask learning is highly effective for protein property prediction.
• PatchProt sets a new benchmark for sequence-based protein property prediction.
• This approach highlights the potential of enriching model representations through related task training.