Rapid Assessment of Virtually Synthesizable Chemical Structures via Support Vector Machine Models.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Support Vector Machine (SVM) and Support Vector Regression (SVR) are established methods for quantitative structure-activity relationship (QSAR) modeling.
• Evaluating vast numbers of virtual molecules, especially those from virtual synthesis, presents a significant computational challenge.

Purpose of the Study:

• To develop an efficient SVM/SVR-based method for screening virtually synthesizable molecules by analyzing their constituent reactants.
• To enable rapid evaluation of billions of molecular combinations for drug discovery and chemical biology applications.

Main Methods:

• Implementation of reactant-wise kernel functions within SVM/SVR models for accelerated computation.
• Utilizing data augmentation techniques to enhance the performance of SVR models.
• Testing the proposed method on 120 small molecular activity datasets against 10 macromolecule targets.

Main Results:

• The proposed SVR models with data augmentation demonstrated performance comparable to standard SVR models using the Tanimoto kernel.
• An exhaustive evaluation of 6.4 x 10^12 reactant combinations was completed in 8 days on a single desktop computer.
• The method successfully enabled large-scale virtual screening without the need for sampling.

Conclusions:

• The developed SVM/SVR-based method offers a computationally efficient solution for screening large libraries of virtually synthesized molecules.
• This approach significantly advances the feasibility of large-scale virtual screening in drug discovery and chemical research.
• Reactant-wise kernel functions provide a viable strategy for accelerating QSAR model evaluations.