Predicting Peptide Aggregation with Protein Language Model Embeddings.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Amyloid fibrils, peptide aggregates, are linked to diseases and therapeutic challenges.
• Experimental characterization of peptide aggregation is costly and data-limited, hindering model development.

Purpose of the Study:

• To introduce PALM (Predicting Aggregation with Language Model embeddings), a deep-learning model for predicting peptide aggregation.
• To leverage transfer learning from pretrained protein language models (pLMs) for enhanced prediction accuracy.

Main Methods:

• Utilized transfer learning with pLM embeddings to develop the PALM model.
• Trained PALM on the WaltzDB-2.0 dataset for peptide classification and aggregation-prone region identification.
• Evaluated PALM's performance on held-out experimental datasets and its ability to predict mutation effects.

Main Results:

• PALM demonstrated competitive performance against existing models in predicting peptide aggregation.
• The model initially struggled to predict the impact of single mutations on amyloid beta peptide aggregation.
• Training PALM on the larger CANYA NNK1-3 dataset significantly improved its performance on mutation prediction tasks.

Conclusions:

• Transfer learning with pLM embeddings enhances model performance, especially with limited data.
• Predicting the effects of single mutations on aggregation requires substantial experimental data.
• PALM offers a promising approach for studying amyloid fibril formation and disease mechanisms.