ChemFLuo

ChemFLuo predicts and prescreens blue and green fluorescent compound interference in chemical libraries to reduce fluorescence-driven false positives in compound screening.

Key Features:

• Prescreening capability: Prescreens large datasets to identify potential blue and green fluorescent interferences in compound collections.
• Training datasets: Models were built from two high-quality datasets comprising 4,906 blue and 8,632 green fluorescent compounds.
• Algorithms and representations: Predictive capability integrates three machine learning algorithms combined with seven distinct molecular representations.
• Model performance — blue fluorescence: Best blue model achieved balanced accuracy (BA) 0.858 and AUC 0.931 on the validation set, and BA 0.823 with AUC 0.903 on the test set.
• Model performance — green fluorescence: Green model achieved BA 0.810 and AUC 0.887 on the validation set, and BA 0.771 with AUC 0.852 on the test set.
• Representative substructures: Identified 22 representative substructures associated with blue fluorescence and 16 associated with green fluorescence.
• Validation and benchmarking: Reliability assessed via comparisons with other fluorescence detection tools and application to external validation sets and large molecule libraries.

Scientific Applications:

• Early-stage drug discovery: Filter potential fluorescent interferents from screening libraries to reduce false positives during hit identification.
• Compound screening assays: Identify compounds with inherent blue or green fluorescence that may confound assay readouts.
• Chemical library design: Guide selection and curation of high-quality chemical libraries by flagging undesirable fluorescent compounds.

Methodology:

Models were trained using three machine learning algorithms with seven molecular representations on datasets of 4,906 blue and 8,632 green fluorescent compounds; performance was evaluated using balanced accuracy (BA) and AUC on validation and test sets; 22 blue and 16 green representative substructures were identified; validation included comparisons with other fluorescence detection tools and testing on external validation sets and large molecule libraries.