BRAID: Batch-effect Reconciliation and Anomaly-aware Imputation for Data

Abstract Motivation High-throughput biomedical data often contain missing values, which, if mismanaged, can distort downstream analysis. While various imputation methods exist, few account for systematic bias from batch effects. Moreover, many batch effect correction algorithms require complete data, making imputation a prerequisite. This sequence inadvertently creates confounding issues, as batch effects can hinder imputation performance. Despite this, standard imputation evaluation practices rarely consider imputation-batch interactions. Results To address this, we propose BRAID (Batch-effect Reconciliation and Anomaly-aware Imputation for Data), an evaluation framework that systematically assesses imputation techniques under assumed batch-free and batch-affected conditions, using ComBat as the primary correction algorithm and limma to assess generalization. Via controlled batch and missing value simulations applied to diverse real-world datasets, we find that many popular imputation methods exhibit non-negligible performance degradation or instability when batch effects are present, and that by incrementing batch effect intensities, some degrade with greater sensitivity than others. A case study involving K-Nearest Neighbours applied to samples further highlights how parameter choice can induce batch bias, underscoring the need for batch-conscious imputation. BRAID equips researchers with an evaluation framework to assess and compare imputation methods under controlled or simulated conditions of interest, informing reliability and reproducibility. Availability and implementation BRAID is available at https://github.com/HarvardHui/BRAID.