Enhancing molecular property prediction through data integration and consistency assessment.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Data science

Background:

• Machine learning models face challenges from data heterogeneity and distributional misalignments, impacting predictive accuracy.
• Preclinical safety modeling in drug discovery is particularly susceptible due to limited data and experimental constraints.
• Existing benchmark datasets exhibit misalignments and inconsistent annotations, hindering reliable model development.

Purpose of the Study:

• To investigate data misalignments and inconsistencies in public ADME datasets used for preclinical safety modeling.
• To highlight the limitations of data standardization and the necessity of rigorous data consistency assessment (DCA).
• To introduce AssayInspector, a novel tool for systematic DCA across diverse scientific datasets.

Main Methods:

• Analysis of public ADME datasets to identify property annotation inconsistencies and distributional misalignments.
• Development of AssayInspector, a model-agnostic package utilizing statistics and visualizations for DCA.
• Evaluation of AssayInspector's capability to detect outliers, batch effects, and discrepancies.

Main Results:

• Significant misalignments and inconsistent property annotations were found between gold-standard and benchmark ADME datasets.
• Data standardization did not consistently improve predictive performance, underscoring the need for pre-modeling DCA.
• AssayInspector effectively identifies data inconsistencies, facilitating more reliable model training.

Conclusions:

• Rigorous data consistency assessment is crucial for robust machine learning in preclinical safety and other scientific domains.
• AssayInspector provides a systematic approach to DCA, enhancing the reliability of integrated datasets.
• The principles of DCA are applicable to federated learning and cross-domain data integration.