Figure 5.

Annotations from low-resolution structures supported through reliable predictions. (A) Our development set (DevSet1014) contained the PDB structure 1A7W^31,32 for the DNA-binding protein HMf-2 (UniProt ID: P19267). No DNA/nucleic acid binding was annotated in that structure, but our new method, bindEmbed21DL, reliably predicted (probability $\ge$ 0.95) four residues to bind nucleic acids. Shown is the PDB structure 5T5K^31,33 for the same protein that has a resolution of 4.0 Å and annotations of DNA-binding, including the four most reliable predictions (dark red). Overall, 10 of 13 (77%) residues annotated as DNA-binding in 5T5K were also predicted by bindEmbed21DL (shown in lighter red; blue residues indicate experimental annotations which were not predicted). (B) For the ribonuclease P protein component (UniProt ID: Q9X1H4), four residues were predicted with a probability $\ge$ 0.95 (indicated in dark red), none of these matched the annotations in the PDB structure 6MAX^31,34. However, those four residues were considered as binding according to the two low-resolution structures 3Q1Q (3.8 Å)^31,35 (visualized) and 3Q1R (4.21 Å)^31,35. In total, those structures marked 21 binding residues; 15 of those 21 (71%) were correctly predicted (light red; blue residues observed to bind but not predicted). These two examples highlighted how combining low-resolution experimental data and very reliable predictions from bindEmbed21DL could refine those annotations and/or help designing new investigations.