Revisiting Batch Effect Correction: A Structured Problem Formulation and Novel Solutions

Existing methods for batch effect removal often suffer from an ill-defined problem formulation and rely on unverified assumptions, limiting their robustness and generalizability. This study proposes a principled approach to systematically address these limitations.

We present two novel methodologies for removing systematic batch effects: an unsupervised approach based on independent component analysis and a supervised approach using a domain adaptation technique. Both methods are grounded in rigorous mathematical principles. Our approaches outperforms existing methods across multiple datasets.

The implementation of our methods, including all code, datasets, experimental workflows, and supplementary material, is available in Python and has been made publicly accessible on GitHub: https://github.com/shaham-lab/revisiting-batch-effect-removal, Supplementary Material: https://github.com/shaham-lab/revisiting-batch-effect-removal/blob/main/proofs/suplementery_matirial.pdf.