All statistical results and figures generated for the manuscript.

Batch effects, referring to unwanted technical variations caused by differences in labs, pipelines or batches, are notorious in MS-based proteomics data, wherein protein-level quantities are inferred from precursor- and peptide-level intensities. However, the choice of which data level for effective batch-effect correction is crucial but remains elusive. Leveraging real-world multi-batch data from the Quartet protein reference materials as well as simulated data, we benchmarked batch-effect correction at precursor, peptide and protein levels combined with two designed scenarios, three quantification methods and seven batch-effect correction algorithms. Our findings reveal that protein-level batch-effect corrections were more robust in both feature-based and sample-based quality assessment, and that the quantification process interacts with the batch-effect correction. Furthermore, we extended to the large-scale data from 1,431 plasma samples of type 2 diabetes patients in phase 3 clinical trials and demonstrated the best prediction performance of Ratio-MaxLFQ combination. This study provides comprehensive insights supporting the batch-effect correction at the protein level to enhance proteomic data integration.