Does ensembling improve feature attributions from sequence-to-activity models?

Sequence-to-activity models take as input DNA sequence and predict genomic activities such as transcription factor binding and gene expression. Applying explainable AI (xAI) methods such as DeepLIFT to these models has recently led to breakthroughs towards many genomic problems, including transcription factor binding grammar and predicting effects of genetic variants. However, there remains significant uncertainty about the reliability of sequence-to-activity interpretations. Thus, we need accurate probabilistic measures of confidence to distinguish reliable from unreliable interpretations. Towards this end, researchers have recently aimed to characterize variability across ensembles of S2A models. However, previous work has focused on using model ensembles to improve the model predictions themselves. Here, we aim to evaluate whether model ensembles can also be used to improve feature attributions from post-hoc xAI methods. We find that ensembling attributions from multiple models improves downstream applications, including identifying transcription factor motifs and predicting regulatory genetic variants. We show that forming an ensemble using Monte Carlo Dropout (MCDropout) gets near to, but does not match, the performance of training multiple models, at much less train-time computational cost.