Hybrid-Enhanced Siamese Similarity Models in Ligand-Based Virtual Screen.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Artificial intelligence in pharmacology

Background:

• Information technology is crucial in modern drug development, with virtual screening (VS) accelerating compound evaluation.
• Similarity searching, a key VS task, relies on structural similarity correlating with biological activity.
• Existing methods struggle with structurally heterogeneous compounds, limiting their effectiveness.

Purpose of the Study:

• To develop an improved molecular similarity search model for virtual screening.
• To enhance the retrieval of structurally diverse compounds.
• To overcome limitations of current similarity search techniques in drug discovery.

Main Methods:

• Investigated deep learning-based Siamese similarity models, specifically Enhanced Siamese Multi-Layer Perceptron (SMLP) and Siamese Convolutional Neural Network-one dimension (SCNN1D).
• Developed and tested various designs of a hybrid model combining SMLP and SCNN1D.
• Conducted experiments using real-world datasets to evaluate model performance.

Main Results:

• The hybrid model demonstrated superior performance compared to individual SMLP and SCNN1D models.
• The combined approach showed improved retrieval effectiveness, particularly for structurally heterogeneous molecules.
• Experiments confirmed the enhanced recall capabilities of the novel hybrid model.

Conclusions:

• The proposed hybrid deep learning model significantly advances molecular similarity searching in virtual screening.
• This approach offers a more effective solution for identifying potential drug candidates from diverse chemical structures.
• The findings suggest a promising direction for improving computational drug discovery pipelines.