Experimentally validated deep learning control of protein aggregation.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Protein aggregation reduces solubility and yield, limiting biotechnological applications.
• Identifying and modifying aggregation-prone regions is crucial for protein engineering.

Purpose of the Study:

• To develop and validate a deep neural network-based predictor, AggreProt, for identifying residue-level aggregation-prone regions in protein sequences.
• To experimentally validate AggreProt's predictions and demonstrate its utility in improving protein solubility and yield.

Main Methods:

• Developed a deep neural network model (AggreProt) for predicting aggregation propensity at the residue level.
• Validated the model on independent datasets of hexapeptides and full-length proteins.
• Experimentally tested predictions using hexapeptides from haloalkane dehalogenase LinB and proteins from the AmyPro database.

Main Results:

• AggreProt outperformed or matched state-of-the-art algorithms in predicting aggregation-prone regions.
• Experimental validation showed 79% agreement with predictions, identifying inaccuracies in existing databases.
• Mutations designed using AggreProt to suppress aggregation in LinB variants resulted in reduced aggregation, improved solubility, and up to 100% increased yield.

Conclusions:

• AggreProt is a reliable tool for predicting protein aggregation-prone regions.
• The tool aids in protein engineering by guiding mutations to enhance solubility and yield.
• AggreProt is available as a web server for the scientific community.