Multitask Modeling with Confidence Using Matrix Factorization and Conformal Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Bioactivity prediction

Background:

• Multitask bioactivity prediction faces challenges with sparse data and imbalanced labels.
• Existing methods struggle to maintain high predictive performance under these conditions.

Purpose of the Study:

• To introduce a novel approach for large-scale bioactivity prediction using class conditional (Mondrian) conformal predictors with Macau models.
• To address data sparsity and label imbalance issues in bioactivity prediction.

Main Methods:

• Utilized class conditional (Mondrian) conformal predictors integrated with Macau models.
• Applied the approach to ten assay endpoints from PubChem for large-scale bioactivity prediction.

Main Results:

• Achieved valid model efficiency of 74.0-80.1% at an 80% confidence level.
• Demonstrated robust performance on imbalanced datasets, with similar efficiency for minority and majority classes.
• Showcased model robustness with progressive data deletion (0-80%), exhibiting only minor efficiency deterioration (5-10%).
• Significantly improved performance on imbalanced datasets compared to Macau models without conformal prediction.

Conclusions:

• The proposed class conditional conformal predictors with Macau models offer a powerful solution for large-scale bioactivity prediction.
• This method effectively handles data sparsity and label imbalance, ensuring high-quality predictive models.
• The approach provides reliable and robust predictions, even with substantial data missing or imbalanced class distributions.