Probabilistic Substructure Mining From Small-Molecule Screens.

Area of Science:

• Chemical informatics
• Computational chemistry
• Bioinformatics

Background:

• Identifying overrepresented substructures is crucial for understanding structure-activity relationships in chemical informatics.
• Current deterministic substructure miners require high confidence in molecular activity data, limiting their application.
• Noisy or probabilistic activity data is common in high-throughput screening.

Purpose of the Study:

• To develop a novel probabilistic substructure miner, pGraphSig, capable of handling noisy molecular activity data.
• To evaluate the performance of pGraphSig against existing deterministic methods.

Main Methods:

• Introduced pGraphSig, a probabilistic graph-based substructure mining algorithm.
• Applied pGraphSig to benchmark datasets from small-molecule high-throughput screens.
• Compared the efficacy of pGraphSig in identifying overrepresented structures against a deterministic substructure miner.

Main Results:

• pGraphSig demonstrated superior performance in identifying overrepresented substructures compared to a deterministic approach.
• The probabilistic miner effectively handled datasets with molecules labeled with probabilities of activity.
• Successful application on real-world high-throughput screening data.

Conclusions:

• Probabilistic substructure mining offers a robust alternative for analyzing chemical data with inherent uncertainty.
• pGraphSig provides a more effective tool for structure-activity relationship studies when dealing with noisy high-throughput screening data.
• This approach enhances the ability to discover key molecular features from large, imperfect datasets.