Fig. 4: Molecular convergence on a neuroactive drug–target connectivity signature predictive of anti-glioblastoma efficacy.
a, COSTAR workflow. b, COSTAR network of 127 PCY-tested drugs, 965 ePTGs and 10,573 STGs, connected by 114,517 edges. c, COSTAR method by logistic LASSO regression. See also Methods. d, COSTAR training model performance compared to PCY-based experimental ground truth. e, COSTAR connectivity (solid lines) reveals convergence of NAD (pink) and ONCD (blue) hits to key ePTGs (gray) and STGs (yellow). See Extended Data Fig. 5c for the full model. Additional proteins (white nodes) with high-confidence interactions to STGs (dashed lines) are shown. f, In silico drug screen across 1,120,823 compounds by COSTAR. Compounds are ranked (x axis) by their predicted PCY-hit probability (COSTAR score; y axis). Predicted drug hits (COSTAR-HIT; mint green) and predicted non-hits (COSTAR-NEG; black) selected for experimental validation are indicated. g, ePTGs (x axis) ranked by their integrated contribution ‘C’ to predict a hit (+1) or a non-hit (−1) (y axis) in the COSTAR model, separated for COSTAR-HITs (top) and COSTAR-NEGs (bottom) (‘d’). h, Drug–target connectivity of select COSTAR-predicted drugs (columns; n = 23 COSTAR-HIT drugs; n = 25 COSTAR-NEG drugs) to primary and secondary drug targets (rows). COSTAR subscore (heatmap color scale) is the LASSO model coefficient multiplied by the integrated connectivity of drug to target mapping. Target genes with absolute COSTAR LASSO coefficients greater than 0.1 are displayed. i, Experimental ex vivo validation by PCY of COSTAR-HIT (n = 23; mint green) and COSTAR-NEG (n = 25; black) drugs (columns) across four glioblastoma patient samples (rows) including positive (PCY-hits; pink; n = 3) and negative (PCY-negative; dark gray; n = 1) control drugs. Heatmap color scale indicates the PCY score of glioblastoma cells. One-tailed t-test; *FDR-adjusted P < 0.05. Outliers beyond color scale limits are set to minimum and maximum values. j, Receiver operating characteristic (ROC) curves (gray, n = 4 patients; mint green, mean across patients; red dashed, random classifier) describing the COSTAR validation accuracy in glioblastoma patient samples of the COSTAR-predicted drugs (n = 48 drugs; corresponding to i). FPR, false-positive rate; PCY-HIT, PCY-hit; PCY-NEG, PCY-negative; TPR, true-positive rate.
