Figure 5.

Cross-validation testing of binding frequencies. Accuracy of each domain–ligand interaction from the representable-NR set is measured as the average AUPRC in cross-validation with up to 10 folds. (A) For each domain–ligand pair, we compute the fold change between the actual AUPRC and a baseline AUPRC corresponding to the fraction of binding positions in the hold-out set. (B) We use each positional binding frequency computed across domain–ligand pairs as a threshold to distinguish predicted binding from non-binding domain positions, and we measure the precision achieved in each held-out set of domain–ligand structural instances using each of these thresholds. Shown is the average precision computed across domain–ligand interactions in the representable-NR set at binding frequency thresholds varying from 1 (highest) to 0 (lowest).