Similarity-Based Virtual Screen Using Enhanced Siamese Multi-Layer Perceptron.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Traditional drug development is time-consuming and expensive.
• Virtual screening (VS) aids drug discovery by measuring molecular similarity.
• Existing ligand-based virtual screening (LBVS) methods struggle with structurally heterogeneous molecules.

Purpose of the Study:

• To enhance similarity searching performance, particularly for structurally heterogeneous molecules.
• To leverage the capabilities of Siamese networks for complex data analysis in drug discovery.

Main Methods:

• An enhanced Siamese multi-layer perceptron architecture was developed.
• The architecture incorporates two similarity distance layers and a fused layer.
• Node pruning based on signal-to-noise ratio was applied to optimize the model.
• The method was validated on benchmark datasets: MDDR (DS1-DS3), MUV, and DUD.

Main Results:

• The proposed Siamese network method demonstrated superior performance compared to the standard Tanimoto coefficient (TAN) and other existing methods.
• The model achieved improved similarity searching for structurally heterogeneous compounds.
• Effective node reduction was possible without compromising recall performance.

Conclusions:

• The enhanced Siamese network offers a more effective approach for virtual screening, especially for diverse molecular structures.
• This method can accelerate drug discovery by improving the efficiency of identifying potential drug candidates.
• Model optimization through node pruning maintains high performance while potentially reducing computational cost.