Applying interpretable machine learning in computational biology-pitfalls, recommendations and opportunities for new developments

Abstract

Recent advances in machine learning have enabled the development of next-generation predictive models for complex computational biology problems, thereby spurring the use of interpretable machine learning (IML) to unveil biological insights. However, guidelines for using IML in computational biology are generally underdeveloped. We provide an overview of IML methods and evaluation techniques and discuss common pitfalls encountered when applying IML methods to computational biology problems. We also highlight open questions, especially in the era of large language models, and call for collaboration between IML and computational biology researchers.

Figure 1:

The two main IML approaches used to explain prediction models are post-hoc explanations and by-design explanations. Each approach has its canonical workflows and popular types of IML methods: post-hoc explanations are model-agnostic and are applied after a model is trained while by-design explanations are typically built into or inherent to the model architecture.

Figure 2:

How do we assess explanations, which attribute importance scores to features of an input, generated by an IML method? IML methods are typically evaluated for the faithfulness of their computed feature importance scores as compared to a known ground truth mechanism and the stability of computed feature importance scores (e.g., as denoted by error bars) across varied input data.

Figure 3:

An overview of three common pitfalls of IML interpretation in biological contexts and how to avoid these pitfalls. 1) Only considering one IML method. Consideration of multiple IML methods can inform the downstream interpretation of the outputs. 2) IML output disconnected from biological interpretation. Oftentimes, a post-processing step is necessary to enable interpretation of the IML output, particularly when the method is applied to sequence or pixel-level data. 3) Cherry-picked presentation of results. Many prior works do not present a complete picture of the extent to which the IML output reflects known biological mechanisms.