A novel adaptive ensemble classification framework for ADME prediction.

Area of Science:

• Computational Chemistry and Cheminformatics
• Pharmacology and Drug Discovery
• Machine Learning in Bioinformatics

Background:

• Accurate in silico prediction of ADME characteristics is crucial for efficient drug discovery.
• Existing in silico ADME prediction models face challenges with unbalanced datasets and high dimensionality.
• There is a need for robust computational frameworks to handle these complexities in ADME profiling.

Purpose of the Study:

• To introduce a novel adaptive ensemble classification framework (AECF) for in silico ADME prediction.
• To address the simultaneous challenges of data imbalance and high dimensionality in ADME modeling.
• To improve the performance and generality of ADME property predictions.

Main Methods:

• Developed AECF, a four-component framework including data balancing, individual model generation, model combination, and ensemble optimization.
• Utilized a choice pool of five sampling methods, seven base modeling techniques, and ten ensemble rules.
• Implemented an adaptive approach where model construction routes are automatically determined by the imbalance ratio (IR).

Main Results:

• AECF demonstrated superior performance across five ADME datasets (CacoP, HIA, OB, PS, PI) with average AUC values ranging from 0.7821 to 0.9182.
• The framework achieved better performance and generality compared to individual models and traditional ensemble methods like bagging and boosting.
• Investigated ensemble member complementarity, revealing that AECF effectively selects diverse members through auto-adaptive optimization.

Conclusions:

• AECF offers a powerful and flexible solution for in silico ADME prediction, adeptly handling data imbalance and high dimensionality.
• The adaptive nature of AECF allows for its application to diverse ADME datasets, with balanced data being a specific case.
• The framework's success is attributed to the effective selection of complementary ensemble members, enhancing predictive accuracy and generalizability.