ChemFLuc

ChemFLuc predicts firefly luciferase (FLuc) inhibitors from chemical structures to identify compounds likely to generate false-positive signals in high-throughput screening (HTS) assays.

Key Features:

• Extensive Dataset Utilization: Model development used a labeled dataset of 20,888 known FLuc inhibitors and 198,608 non-inhibitors.
• Machine Learning Algorithms: Employs a combination of three machine learning algorithms, with the best-performing model based on XGBoost.
• Molecular Descriptors: The best-performing model uses ECFP4 and MOE2d molecular descriptors.
• Predictive Performance: Achieved balanced accuracy (BA) and area under the ROC curve (AUC) of 0.878 and 0.958 on the validation set, and BA 0.886 and AUC 0.947 on the test set.
• External Validation: Performance was confirmed on three external validation sets with BA values of 0.864, 0.845, and 0.791.
• Feature Recognition and Rule Development: Applied Shapley additive explanations (SHAP) to identify structural fragments associated with FLuc inhibition and derived 16 predictive rules with a 70% correction rate.
• Comparison with Existing Models: Compared models and rules with existing prediction tools in virtual screening contexts and reported superior reliability.
• Risk Assessment in Chemical Databases: Applied the model to three curated chemical databases and identified approximately 10% of evaluated molecules as potential FLuc inhibitors.

Scientific Applications:

• HTS Prescreening: Prescreen chemical libraries to flag compounds likely to inhibit firefly luciferase and cause false positives in HTS assays.
• Virtual Screening Benchmarking: Serve as a comparative model in virtual screening workflows to assess interference from FLuc inhibitors.
• Chemical Database Risk Assessment: Estimate the prevalence of potential FLuc inhibitors within curated chemical databases to prioritize follow-up validation.

Methodology:

Modeling used a labeled dataset of 20,888 inhibitors and 198,608 non-inhibitors, trained with three machine learning algorithms (best model: XGBoost using ECFP4 and MOE2d descriptors), evaluated by BA and AUC on validation/test sets and three external validation sets, interpreted with SHAP to identify fragments and derive 16 predictive rules, and applied to three curated chemical databases.