Batch Integration

The integration results are interpreted as preserving immunophenotypic boundaries while handling nonlinear batch structure. On the Stuart dataset, scpFormer achieved AvgBIO 0.795, higher than ComBat and Harmony. On the Zheng dataset, scpFormer achieved the best reported overall balance with AvgBIO 0.724 and AvgBATCH 0.911. The article says Harmony overcorrected biological structure on the Zheng dataset. scpFormer adds adversarial domain adaptation and an Elastic Cell Similarity contrastive loss for integration-specific training. For clustering and batch integration, scpFormer extracts the CLS representation with bidirectional attention over shared proteins.