Positional SHAP for Interpretation of Deep Learning Models Trained from Biological Sequences
AbstractMachine learning with artificial neural networks, also known as “deep learning”, accurately predicts biological phenomena such as disease diagnosis and protein structure. Despite the ability of deep learning to make accurate biological predictions, a challenge is model interpretation, which is especially challenging for recurrent neural network architectures due to the sequential input data. Here we train multi-output long short-term memory (LSTM) regression models to predict peptide binding affinity to five rhesus macaque major histocompatibility complex (MHC) I alleles. We adapt SHapely Additive exPlanations (SHAP) to generate positional model interpretations of which amino acids are important for peptide binding. These positional SHAP values reproduced known rhesus macaque MHC class I (Mamu-A1*001) peptide binding motifs and provided insights into inter-positional dependencies of peptide-MHC interactions. Positional SHAP should find widespread utility for interpreting a variety of models trained from biological sequences.