Figure 1.

Performance of different model explainers without optimization

A. Spearman’s correlation of gene contribution scores (upper panel) and overlap in the top 100 contributing genes (lower panel) in liver among replicates from the same pre-trained model, different pre-trained models with the same gene order, and different pre-trained models with different gene orders on CNN-based models. B. Spearman’s correlation on gene contribution scores (upper panel) and overlap in the top 100 contributing genes (lower panel) in liver among replicates from the same pre-trained model and different pre-trained models on MLP-based models. CNN, convolutional neural network; MLP, multilayer perceptron; Saliency, gradients; InputXGradient, Input x Gradient; GuidedBackprop, guided backpropagation; IntegratedGradients, Integrated Gradients; DeepLift, DeepLIFT; DeepLiftShap, approximating SHAP values using DeepLIFT; GuidedGradCam, Guided Grad-CAM; GuidedGradCam++, Guided Grad-CAM++.