In silico approaches have become indispensable for drug discovery as well as drug repositioning and adverse effect prediction. We have developed the eF-seek program to predict protein-ligand interactions based on the surface structure of proteins using a clique search algorithm. We have also developed a special protein structure prediction pipeline and accumulated predicted 3D models in the Structural Atlas of the Human Genome (SAHG) database. Using this database, genome-wide prediction of non-peptide ligands for proteins in the human genome was performed, and a subset of predicted interactions including 14 PDZ domains was then confirmed by NMR titration. Surprisingly, diclofenac, a non-steroidal anti-inflammatory drug, was found to be a non-peptide PDZ domain ligand, which bound to 5 of 15 tested PDZ domains. The critical residues for the PDZ-diclofenac interaction were also determined. Pharmacological implications of the accidental PDZ-diclofenac interaction are further discussed.