Positional SHAP (PoSHAP) for Interpretation of machine learning models trained from biological sequences

Abstract

Machine learning with multi-layered artificial neural networks, also known as “deep learning,” is effective for making biological predictions. However, model interpretation is challenging, especially for sequential input data used with recurrent neural network architectures. Here, we introduce a framework called “Positional SHAP” (PoSHAP) to interpret models trained from biological sequences by utilizing SHapely Additive exPlanations (SHAP) to generate positional model interpretations. We demonstrate this using three long short-term memory (LSTM) regression models that predict peptide properties, including binding affinity to major histocompatibility complexes (MHC), and collisional cross section (CCS) measured by ion mobility spectrometry. Interpretation of these models with PoSHAP reproduced MHC class I (rhesus macaque Mamu-A1*001 and human A*11:01) peptide binding motifs, reflected known properties of peptide CCS, and provided new insights into interpositional dependencies of amino acid interactions. PoSHAP should have widespread utility for interpreting a variety of models trained from biological sequences.

Author summary

Machine learning enables biochemical predictions. However, the relationships learned by many algorithms are not directly interpretable. Model interpretation methods are important because they enable human comprehension of learned relationships. Methods likeSHapely Additive exPlanations were developed to determine how each input alters the model prediction. However, interpretation of models trained from biological sequences remains more challenging; model interpretation often ignores ordering of inputs. Here, we train machine learning models using biological sequence data as an input to predict peptide collisional cross section, and to predict peptide binding affinity to major histocompatibility complex (MHC) isoforms. To enable positional interpretation of our predictions, we add indexes to the inputs to track SHAP explanations calculated from the models. Our results demonstrate that positional interpretation of models recapitulates known biochemistry and reveals new biochemistry. This positional SHAP (PoSHAP) conceptual framework provides a foothold for interpretation of other models trained from biological sequences.