AggrescanAI: Prediction of Aggregation-Prone Regions Using Contextualized Embeddings

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Neuroscience

Background:

• Protein aggregation is implicated in neurodegenerative diseases.
• Aggregation-prone regions (APRs) drive protein aggregation.
• Predicting APRs is crucial for protein engineering and disease research.

Purpose of the Study:

• To develop a novel deep learning tool, AggrescanAI, for predicting residue-level protein aggregation propensity.
• To leverage protein language models for sequence-based aggregation prediction without structural data.
• To provide an accessible and open-source tool for researchers.

Main Methods:

• Utilized ProtT5 protein language model for contextual embeddings.
• Trained the deep learning model on experimentally validated APRs and expanded datasets.
• Evaluated model performance using cross-validation and an external benchmark.
• Assessed the model's ability to predict aggregation shifts caused by mutations.

Main Results:

• AggrescanAI accurately predicts residue-level aggregation propensity directly from protein sequences.
• The tool outperforms existing state-of-the-art aggregation predictors.
• AggrescanAI successfully captures aggregation changes induced by pathogenic mutations.
• A user-friendly Google Colab notebook is available for accessibility.

Conclusions:

• AggrescanAI represents a significant advancement in sequence-based protein aggregation prediction.
• The tool enhances capabilities in neurodegenerative disease research and protein engineering.
• Deep learning and protein language models offer powerful approaches for predicting protein behavior.