Fig. 3

Performance evaluation of IG, EIG, and Deep SHAP on splicing data. a Number of significant meta-features identified by different methods: simple gradients, original IG (O-L-IG), and EIG nonlinear paths (see main text) with a zero baseline. b Number of significant meta-features identified by different EIG baselines (explained in text) with a linear path in the original feature space. c Number of significant meta-features identified by different EIG paths with an encoded-zero baseline. d Number of significant meta-features identified by different EIG paths with three median-constitutive baseline points. e Enrichment of known brain regulatory features in significant features identified by the best two EIG paths each from encoded-zero and median-constitutive baselines and Deep SHAP. f Splicing prediction for differential inclusion in the brain with increasing subsets of significant meta-features identified by EIG with latent-linear path (H-L-IG) and median baseline from constitutive splicing events (orange line) or random meta-features (gray line). The x-axis shows the number of meta-features. The y-axis shows the AUC-ROC for differential inclusion on the test set. Significant features passing one-sided t test, Bonferroni adjusted p value ≤ 0.05. g Splicing prediction for differential inclusion in the brain using features found to be significantly generated by different combinations of paths (O-L-IG, O-N-IG, H-L-IG, H-N-IG) and baselines (zero, encoded-zero, median, k-means, close, random) and simple gradients. x-axis shows the different interpretation methods, and y-axis shows the AUC-ROC for differential inclusion in the brain on the test set