AggNet: Advancing protein aggregation analysis through deep learning and protein language model.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Science

Background:

• Protein aggregation is crucial in biological processes and biotherapeutic development.
• Experimental methods for protein aggregation profiling are costly and time-consuming.
• Efficient computational approaches are needed to predict protein aggregation.

Purpose of the Study:

• Introduce AggNet, a novel deep learning framework for protein aggregation prediction.
• Utilize physicochemical, evolutionary, and structural data for accurate prediction.
• Identify aggregation-prone regions (APRs) in diverse proteins.

Main Methods:

• Developed AggNet using the ESM2 protein language model and AlphaFold2.
• Integrated physicochemical, evolutionary, and structural information.
• Benchmarked AggNet against existing protein aggregation prediction methods.

Main Results:

• AggNet achieved state-of-the-art performance in predicting protein aggregation.
• The model demonstrated stable predictive ability across proteins with varying secondary structures.
• Feature analysis confirmed AggNet's effective capture of peptide physicochemical properties, enhancing interpretability.

Conclusions:

• AggNet offers a computationally efficient and accurate alternative to experimental methods for protein aggregation analysis.
• The framework has practical utility in guiding protein engineering and mutation strategies for aggregation mitigation.
• AggNet enhances computational tools for predicting protein aggregation and protein engineering applications.