Publication
Computational drug screening is of fundamental importance in early-stage drug discovery. The current computational methods predict compound affinities to their targets based on docking or pharmacophore (PH4) hypotheses. Here, we develop Alpha-Pharm3D, a versatile deep learning method that predicts ligand-protein interactions using 3D PH4 fingerprints by explicitly incorporating geometric constraints. This comprehensive new algorithm enhances substantially not only the prediction interpretability and accuracy of binding affinities of ligands against the target protein, but also the PH4 potential for screening large compound libraries efficiently. Alpha-Pharm3D outperforms state-of-the-art scoring methods in bioactivity prediction and achieves considerable improvements in both accuracy and success rate, irrespective of data scarcity. We demonstrate the superior applicability of Alpha-Pharm3D for compound screening against the NK1R, a cancer growth and metastasis related G-protein coupled receptor, yielding nanomolar active compounds. This opens up attractive possibilities for applying PH4 fingerprints to efficiently promote scaffold hopping and ultimately accelerate ligand-based drug discovery.