Fig. 4: Sensitivity to input sequences using Explainable AI.
From: Accuracy and data efficiency in deep learning models of protein expression
A DeepLIFT37 attribution scores per nucleotide position for a given test sequence and trained model. Panels show scores of 30 sequences chosen at random from the same test set employed in Fig. 3C for models trained on 75% of mutational series 21. B Attribution distances for models trained on series 21. We computed the cosine distance between DeepLIFT scores for each sequence in the test set. Distance heatmaps were hierarchically clustered to highlight the cluster structure that both models assign to the input sequences. C K-means clustering of the distance matrices in panel B. Line plots show the optimal k-means score averaged across 20 runs with random initial cluster assignments. Lower scores for all values of k suggests that the MLP clusters sequences more heavily than the CNN; we found this pattern in all but four mutational series (Supplementary Fig. S10).
