Supportinng information of methods.

More details on implementation, evaluation metrics, docking methods, compound desgin and synthesis and additional results in teh dark chemical genomics sequence exploration. Table A: Model architecture configuration. Table B: 65 compounds tested for selective dual DRD1/3 antagonists. Table C: Undruggable human disease-associated proteins selected by ProtalCG. Table D: Predicted ligands for the transcription factors and transcription activity related proteins. Table E: Chemicals interacted with undruggable human proteins. Table F: Targets predicted by PortalCG for AI-10–49, fenebrutinib, PF-05190457 and their docking score from Autodock Vina. Table G: Functional Annotation enrichment for human proteins in Tbio selected by PortalCG. Table H: Chemicals interacted with human proteins in Tbio. Table I: Functional Annotation enrichment for undruggable human disease proteins without Tbio selected by PortalCG. Table J: Top ranked diseases associated with undruggable human proteins excluding Tbio selected by PortalCG. Table K: Chemicals interacted with undruggable human proteins excluding Tbio. Fig A: Model performance breakdown to each class. In the main text, overall evaluation across positive and negative classes are reported, such as F1, ROC-AUC, PR-AUC. Here is a breakdown of performance in each class, where class0 is negative, i.e. not binding, class1 is positive, i.e. binding. against DISAE as baseline. Fig B: Performance comparison of PLD+SIGN and Autodock Vina Performance comparison of PLD+SIGN and Autodock Vina using the same OOD-test set as in main text: ROC-AUCs of Autodock Vina and PLD+SIGN are 0.535 and 0.569, respectively. PR-AUCs of Autodock Vina and PLD+SIGN 0.398 and 0.433, respectively. Fig C: t-test comparison. t-test comparison. The p-values for both ROC-AUC and PR-AUC are close to 0 against DISAE as baseline. Fig D: Stress model selection performance curves against DISAE as baselinE. Fig E: Ratio 1:50 for DUD.E. The ratio of inactive CPIs vs active CPIs under different Tanimoto coefficients of chemical similarities in the training data. The ratio of total inactive CPIs vs active CPIs is 1:1. Fig F: Histogram equalization results The left panel shows the original distribution of distance real values; to formalize a multi-class classification where each class has equal probability, histogram equalization transforms the distribution to the right panel of 10 bins, each as a class. (PDF)