BindScape
Predicts whether a drug and a protein will bind, using molecular fingerprints and protein language model embeddings, no 3D structure or docking involved.
+ Detail- Close
BindScape
I tested it across more than 456,000 drug-protein pairs against 495 human kinases. Adding the protein embedding actually hurt performance on new targets compared to using the fingerprint alone. The protein representation turned out to be the real bottleneck for generalizing to targets the model hadn't seen before.
0.982
AUROC, random split
0.781
AUROC, held-out target
456K+
pairs tested







