Pairwise Difference Regression: A Machine Learning Meta-algorithm for Improved Prediction and Uncertainty...

Area of Science:

• Computational chemistry
• Machine learning applications
• Chemical informatics

Background:

• Machine learning (ML) is increasingly used in chemical design, but faces challenges in generalizing from small datasets and quantifying uncertainty.
• Current ML models often struggle with vast chemical spaces, necessitating better methods for exploring unexplored regions.
• Traditional chemical intuition leverages differences between conditions for reliable predictions, a principle underexplored in ML.

Purpose of the Study:

• To develop a novel ML regression approach inspired by comparison-based chemical intuition.
• To enhance generalization and uncertainty quantification in ML models for chemical discovery.
• To create a method applicable to various ML algorithms and competitive with existing techniques.

Main Methods:

• Developed pairwise difference regression (PADRE), a comparison-based ML regression technique.
• PADRE learns to predict differences between pairs of input data points during training.
• Predictions are generated by pairing test points with training points, yielding a distribution for mean prediction and uncertainty.

Main Results:

• PADRE significantly improved the performance of the random forest algorithm across five chemical ML tasks.
• The uncertainty measure derived from PADRE (dispersion) correlated well with model error.
• PADRE demonstrated strong performance in active learning scenarios.
• The method proved competitive with state-of-the-art neural network approaches.

Conclusions:

• Pairwise difference regression (PADRE) offers a robust and versatile approach for ML in chemical discovery.
• PADRE enhances prediction accuracy and provides reliable uncertainty quantification.
• This method shows significant promise for improving candidate selection algorithms in drug and materials discovery.