Interpretable machine learning for genomics

Abstract

High-throughput technologies such as next-generation sequencing allow biologists to observe cell function with unprecedented resolution, but the resulting datasets are too large and complicated for humans to understand without the aid of advanced statistical methods. Machine learning (ML) algorithms, which are designed to automatically find patterns in data, are well suited to this task. Yet these models are often so complex as to be opaque, leaving researchers with few clues about underlying mechanisms. Interpretable machine learning (iML) is a burgeoning subdiscipline of computational statistics devoted to making the predictions of ML models more intelligible to end users. This article is a gentle and critical introduction to iML, with an emphasis on genomic applications. I define relevant concepts, motivate leading methodologies, and provide a simple typology of existing approaches. I survey recent examples of iML in genomics, demonstrating how such techniques are increasingly integrated into research workflows. I argue that iML solutions are required to realize the promise of precision medicine. However, several open challenges remain. I examine the limitations of current state-of-the-art tools and propose a number of directions for future research. While the horizon for iML in genomics is wide and bright, continued progress requires close collaboration across disciplines.

Fig. 1

The classic bioinformatics workflow spans data collection, model training, and deployment. iML augments this pipeline with an extra interpretation step, which can be used during training and throughout deployment (incoming solid edges). Algorithmic explanations (outgoing dashed edges) can be used to guide new data collection, refine training, and monitor models during deployment

Fig. 2

A saliency map visually explains a cancer diagnosis based on whole-slide pathology data. The highlighted regions on the right pick out the elements of the image that the algorithm deemed most strongly associated with malignancy. From (Zhang et al. , p. 237)

Fig. 3

A complex decision boundary (the pink blob/blue background) separates red crosses from blue circles. This function cannot be well-approximated by a linear model, but the boundary near the large red cross is roughly linear, as indicated by the dashed line. From (Ribeiro et al. , p. 1138)

Fig. 4

Shapley values show that high white blood cell counts increase the negative risk conferred by high blood urea nitrogen for progression to end stage renal disease (ESRD). From (Lundberg et al. , p. 61)

Fig. 5

Example rule lists for AMR prediction from genotype data. Each rule detects the presence/absence of a k-mer and is colored according to the genomic locus at which it was found. From (Drouin et al. , p. 4)