Interpretably deep learning amyloid nucleation by massive experimental quantification of random sequences

More than 50 human diseases are characterized by the deposition of specific protein aggregates in the form of insoluble amyloid fibrils. However, only a very small number of proteins are known to form amyloids with high propensity, limiting our ability to understand, predict and engineer amyloid aggregation from sequence. Here we use a massively parallel assay to quantify the amyloid nucleation propensity of >100,000 random 20 amino acid sequences. Approximately 5% of assayed random sequences nucleate the formation of aggregates, generating a very large and diverse training dataset from which to train models to predict amyloid nucleation. We use this dataset to train CANYA, a convolution-attention hybrid neural network that predicts the propensity of any primary sequence to form amyloids. CANYA outperforms previous predictors of protein aggregation on additional random sequences and out-of-sample datasets including human disease-causing amyloids, with very stable performance across diverse prediction tasks. We adapt and extend recent advances in interpretability of genomic neural networks to elucidate CANYA's decision-making process and learned grammar and to provide mechanistic insights into amyloid formation. Our results demonstrate the power of massive experimental random sequence-space exploration and provide an interpretable and robust neural network model for understanding, predicting and designing amyloid-forming proteins. Overall design: Systematic measurement of the nucleation of random 20mers peptides