A novel batch effect correction framework for robust integration of high-variance data via a global-information virtual reference batch.

The integration of multi-batch high-variance datasets is increasingly important in studies of complex biological systems. In application domains such as microbial symbiosis, host-microbe interactions, and ecosystem robustness, this places a stringent demand on batch correction methods which must reduce technical batch effects while preserving the biologically meaningful cross-sample structure required for downstream interpretation. Here, we present GIR-Combat, a novel batch correction framework that constructs a global-information virtual reference batch from shared cross-batch structure, thereby enabling more consistent and objective correction across datasets. GIR-Combat identifies mutually nearest neighbors across batches, leverages their shared information to define a virtual reference, and incorporates this reference into a linear modeling framework for correction. By transforming reference-batch specification from a subjective choice into a modeling step, GIR-Combat provides a more objective and robust solution for correcting high-variance and compositionally imbalanced datasets in which conventional methods frequently underperform. We evaluated GIR-Combat on simulated datasets and multiple public benchmark datasets. The results show that GIR-Combat improves batch correction performance relative to existing methods, achieving better batch correction while preserving biologically meaningful structure. Quantitative and visual evaluation metrics further demonstrate its robustness and scalability in challenging integration scenarios. Overall, GIR-Combat provides a practical and methodologically grounded framework for high-variance multi-batch data integration, with potential value in applications where reliable integrated representations are required for interpreting complex biological interactions.