Energy-based Neural Networks as a Tool for Harmony-based Virtual Screening.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Energy-Based Neural Networks (EBNNs) model variable relationships using an 'energy' function.
• Virtual screening is crucial for identifying potential drug candidates.

Purpose of the Study:

• Introduce and evaluate a 'harmony search' procedure for virtual screening using EBNNs.
• Compare harmony search performance against traditional similarity search methods.

Main Methods:

• Harmony search was applied to Hopfield Networks (HNs) and Restricted Boltzmann Machines (RBMs).
• Performance was benchmarked against Tanimoto coefficient-based similarity search using AUC and 1%-enrichment rates for 20 targets.
• Evaluated energy, free energy, reconstruction error, and Tanimoto coefficients.

Main Results:

• Harmony search performance was comparable or superior to similarity search, particularly for several targets.
• Harmony search demonstrated high computational efficiency in predictions.
• The method effectively revealed and utilized regularities in active compound structures.

Conclusions:

• Harmony search is a computationally efficient and effective method for virtual screening with EBNNs.
• The approach offers flexibility, interpretability, and the ability to capture structural regularities.
• Harmony search presents a promising alternative to conventional similarity search techniques in drug discovery.