Improved Scaffold Hopping in Ligand-Based Virtual Screening Using Neural Representation Learning.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep learning excels at automated feature extraction, yet its application in ligand-based virtual screening (LBVS) is limited by the need for large, target-specific datasets.
• Traditional LBVS methods often require extensive training data, hindering their utility in prospective drug discovery scenarios.

Purpose of the Study:

• To develop a generally applicable deep learning tool for LBVS that overcomes the limitations of traditional approaches.
• To create a learning framework that leverages historical screening data across multiple targets for a universal molecular representation.

Main Methods:

• A novel neural network architecture was designed, utilizing molecular graphs as input.
• A learning framework was developed to create a hyperdimensional feature space where compounds with similar biological profiles are positioned closely.
• This representation was learned by simultaneously training on historical screening data from diverse targets.

Main Results:

• The developed model demonstrates exceptional generalization to unseen compounds and targets.
• Outperforms popular fingerprinting algorithms in three standard LBVS benchmarks without target-specific training.
• The learned molecular representation shows superior performance in scaffold hopping and is largely orthogonal to existing fingerprints.

Conclusions:

• A validated framework for learning a target-agnostic molecular representation for LBVS has been established.
• The approach enables effective screening with as few as one query compound, offering significant value from large screening data repositories.
• The implementation is publicly available, facilitating broader adoption in drug discovery.