Identification of small molecule aggregators from large compound libraries by support vector machines

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Small molecule aggregators non-specifically inhibit proteins, hindering therapeutic development.
• Aggregators frequently cause false hits in high-throughput screening (HTS), necessitating their removal.
• Existing computational methods for aggregator identification lack validation on large compound libraries.

Purpose of the Study:

• To develop and validate a computational model for identifying small molecule aggregators.
• To assess the model's performance in screening large compound libraries and minimizing false positives.

Main Methods:

• Developed a support vector machine (SVM) model using 1319 known aggregators and 128,325 non-aggregators.
• Validated the SVM model using five-fold cross-validation, independent testing, and retrospective screening of large databases (13M PUBCHEM, 168K MDDR).
• Compared SVM performance against other machine learning methods.

Main Results:

• The SVM model achieved comparable aggregator and improved non-aggregator identification rates in cross-validation.
• Successfully identified 71% of independently discovered aggregators.
• Retrospective screening predicted >97% of compounds in large databases as non-aggregators, with only 1.14% of similar compounds predicted as aggregators.
• SVM demonstrated superior performance compared to other machine learning methods.

Conclusions:

• The developed SVM model is capable of accurately identifying small molecule aggregators from large compound libraries.
• The model shows substantial potential for reducing false hits in high-throughput screening campaigns.
• The findings support the use of computational methods for efficient and accurate aggregator identification in drug discovery.