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

Area of Science:

• Medicinal Chemistry
• Computational Drug Discovery
• Machine Learning in Cheminformatics

Background:

• Similarity-based virtual screening is crucial for early drug discovery, relying on molecular fingerprints.
• Existing molecular representations may not fully capture target-specific information for enhanced screening.

Purpose of the Study:

• To develop a novel strategy for generating domain-specific molecular fingerprints using neural networks.
• To evaluate the performance of these neural fingerprints in similarity searches against established methods.

Main Methods:

• Training neural networks on target-specific bioactivity datasets to generate domain-specific fingerprints from neural network activations.
• Evaluating five neural network architectures and their generated fingerprints on a large kinase-specific bioactivity dataset.
• Comparing performance against extended-connectivity fingerprint (ECFP) and autoencoder-based fingerprints.

Main Results:

• The proposed neural fingerprint demonstrated superior performance in similarity searches compared to ECFP and autoencoder methods.
• The neural fingerprint maintained high performance even when training data lacked specific target information.
• Surprisingly, fully connected neural networks using ECFP4 as input outperformed Graph Neural Networks (GNNs).

Conclusions:

• Domain-specific neural fingerprints offer a powerful new molecular representation for drug discovery.
• This approach enhances similarity searching capabilities, particularly in target-specific contexts.
• The developed neural fingerprinting method provides a valuable, freely available tool for researchers.