Figure 7.

Binding predictions for complete human proteome. BioLiP: experimental annotations, HBI: homology-based inference, bindEmbed21DL: predictions. Data: human proteome from Swiss-Prot³⁶ with 20,386 proteins. For 3121 human proteins (15%), any binding annotation was experimentally known; 2211 of those were reliable (resolution $\le$ 2.5 Å; darkblue bar for BioLiP). For 10,526 proteins (52%), binding annotations could be inferred using HBI at EVAL ≤ 1 (light blue). Of those, 7973 proteins were not previously annotated (blue bar); for 7199 proteins without binding annotations, binding annotations could be inferred at EVAL ≤ 10^–3 following the protocol of bindEmbed21HBI (dark blue). Therefore, BioLiP + HBI allowed annotating some binding in 52% of the human proteome. bindEmbed21DL predicted binding residues for 18,663 proteins (92%) (light blue); no annotations were previously known (neither through BioLiP nor HBI) for 8510 (blue). Highly reliable predictions (probability ≥ 0.95) were possible for 1751 proteins without previous binding annotations (dark blue).