Fig. 3

Precision-recall curves given by the generic and PARP1-specific SFs. To generate PARP1-specific ML SFs, docked poses of the PARP1-ligand complex were encoded either as GRID features (top) or as PLEC fingerprints (bottom). The resulting features were used by each of the following classification (left) and regression (right) learning algorithms: RF (purple, dashed line), XGB (green, dashed line), SVM (blue, dashed line), ANN (sienna, dashed line), and DNN (violet, dashed line). The PR curve of each target-specific ML SF is that of the training-test run giving an NEF1% equal (or closest) to the median NEF1% across 10 runs (chosen at random if multiple runs satisfy this criterion). The generic SFs are represented as solid lines in gray (Smina), light blue (CNN-Score), and salmon (SCORCH). Results are further specified in Additional file 1:Table S1