Massive experimental quantification allows interpretable deep learning of protein aggregation.

Area of Science:

• Biochemistry
• Computational Biology
• Genomics

Background:

• Protein aggregation is implicated in over 50 human diseases and poses challenges in biotechnology.
• Existing computational methods for predicting protein aggregation are limited by small, biased training datasets.

Purpose of the Study:

• To address the data shortage in protein aggregation prediction.
• To develop an accurate and interpretable computational model for predicting protein aggregation from sequence.

Main Methods:

• Experimentally quantified the aggregation propensity of over 100,000 protein sequences.
• Trained a novel convolution-attention hybrid neural network (CANYA) on the generated dataset.
• Applied genomic neural network interpretability analyses to understand the model's decision-making process.

Main Results:

• The large-scale experimental dataset revealed the limited performance of existing prediction methods.
• The developed CANYA model accurately predicts protein aggregation from sequence.
• Interpretability analyses provided insights into the learned grammar and decision-making of CANYA.

Conclusions:

• Massive experimental analysis of random sequence spaces is powerful for advancing biological predictions.
• CANYA offers an interpretable and robust neural network for predicting protein aggregation.
• This work provides a valuable resource and tool for understanding and mitigating protein aggregation-related issues.