Using Domain-Specific Fingerprints Generated Through Neural Networks to Enhance Ligand-Based Virtual Screening

Similarity-based virtual screening is a fundamental tool in the early drug discovery process and relies heavily on molecular fingerprints. We propose novel strategy of generating domain-specific fingerprints by training neural networks target-specific bioactivity datasets using activation as new representation. The network expected to combine information already known bioactive compounds with unique structure doing so enrich fingerprint. evaluate this large kinase-specific dataset. A comparison five architectures their well-established extended-connectivity fingerprint (ECFP) an autoencoder shows that our produces better results similarity search. Most importantly, performs well even when specific targets are not included during training. Surprisingly, while Graph Neural Networks (GNNs) thought offer advantageous alternative, best performing were based traditional fully connected layers ECFP4 input. freely available at: https://github.com/kochgroup/kinase_nnfp.