Figure 3. Annotation of enzyme active sites at amino-acid resolution.

(A) Precision and recall of active site residue identification for all correct predictions in the validation set. Each orange dot represents a single protein, and the four sampled proteins in (B–E) are labeled with colored dots. For comparison, DeepFRI performance is represented as a precision-recall curve, where the blue line is the average over all proteins and the shaded error bar is the standard deviation. Four sampled structures, representing active site annotations by PARSE across proteins with diverse performance characteristics and enzymatic activities: (B) dehydroneopterin aldolase, (C) succinate dehydrogenase, (D) type-II hexokinase, and (E) asparaginase. In all examples, correctly identified active site residues are shown as green sticks. Correctly identified catalytic residues are shown as green spheres, and catalytic residues which are not identified by PARSE are shown as yellow spheres. Residues annotated by PARSE but not present in the reference site from CSA are shown as yellow sticks. The backbone cartoon is colored by DeepFRI’s gradient-weighted class activation map score, from blue (low) to red (high).