Model-Agnostic Interpretation of Cancer Classification with Multi-Platform Genomic Data

Machine learning models are often criticised for being black-boxes. Recent work in this field has aimed to address this criticism by developing methods to explain the underlying behaviour of machine learning models. These explanations are designed to help the end-user interpret how the models input features are used to make a prediction. Here, we present an extension to one such method, referred to as local interpretable model-agnostic explanations, to interpret multimodal tumor type classification from multi-platform genomic data. We propose a framework for transparent biomedical machine learning by leveraging interpretable dimensionality reduction to facilitate gene-wise explanations for the model behaviour. Using RNA-seq expression and single nucleotide variation (SNV) data from eight cancer types, our experimental results uncovered the models use of clinically relevant genes for cancer cell stratification. We demonstrate that model-agnostic explanations can provide valuable information to a clinician or scientist when predictive ability and interpretability are of absolute importance.