Transformational machine learning: Learning how to learn from many related scientific problems.

Area of Science:

• Machine Learning
• Computational Biology
• Cheminformatics

Background:

• Traditional machine learning relies on intrinsic features for representing data.
• Related machine learning tasks offer potential for feature enhancement.
• Transformational ML (TML) leverages predictions from other tasks to create novel representations.

Purpose of the Study:

• To introduce and evaluate Transformational Machine Learning (TML).
• To assess TML's impact on predictive performance across various machine learning methods and scientific domains.
• To explore TML's contribution to scientific understanding via explainable ML.

Main Methods:

• TML was applied to nonlinear machine learning models including random forests, gradient boosting machines, support vector machines, k-nearest neighbors, and neural networks.
• Thousands of machine learning problems were utilized across drug design, gene expression prediction, and ML algorithm selection.
• Extrinsic features generated by TML were compared against intrinsic features.

Main Results:

• TML significantly improved predictive performance for all tested machine learning methods, with average gains ranging from 4% to 50%.
• The extrinsic features generated by TML generally outperformed intrinsic features.
• TML provided valuable insights in drug design, including drug target specificity and relationships between drugs and proteins.

Conclusions:

• TML offers a synergistic approach to machine learning, enhancing model performance and interpretability.
• TML fosters an ecosystem where tasks and predictions interact to improve overall ML capabilities.
• Openly published data, code, and models support the TML ecosystem and its principles.