Figure 7 from DeePathNet: A Transformer-Based Deep Learning Model Integrating Multiomic Data with Cancer Pathways

Ablation study of DeePathNet using transformer-only model and plain neural networks (MLP). A, Box plots showing the predictive performance from the independent test set across the 549 drugs for random forest (green), MLP (dark gray), DeePathNet with randomly wired pathways (light gray), and DeePathNet (pink). Outliers for R² are hidden for clearer visualization. B, Bar plots showing predictive performances of the four models from cross-validation on cancer type classification. C, Bar plots showing predictive performance from the independent test set on breast cancer subtype classification. CIs represent 95% CI of the mean (n = 10 experiments). D, Down-sample analysis showing DeePathNet with cancer pathways performs better on drug response prediction with smaller training sample sizes compared with DeePathNet with random wired pathways. The solid line represents the mean R², and CIs represent 95% CIs of the mean (n = 10 experiments). ***, P < 0.001, paired t test.