HGAlign：異種グラフアラインメントによるメタボロミクスにおける生物学的バッチ補正と分類の維持

Matrix-Assisted Laser Desorption/Ionization Mass Spectrometry (MALDI-MS) is a powerful tool for profiling complex biological samples. However, large-scale metabolomics experiments often suffer from substantial batch effects caused by variations in sample processing, instrument conditions, and acquisition protocols. These non-biological variations obscure true biological signals, reduce reproducibility, and compromise the generalizability of downstream models. Existing correction methods either rely on oversimplified linear assumptions or risk over-correcting and removing meaningful biological differences. To address this challenge, we propose HGAlign(Heterogeneous Graph Alignment Model), a neural network model that corrects batch effects in large-scale MALDI-MS experiments while preserving important biological differences. Our approach uses heterogeneous graph convolutional networks to learn relationships between samples and metabolic features, enabling effective batch correction without losing disease-related information. Extensive experiments on CyTOF public datasets and clinical MALDI-MS serum data from systemic lupus erythematosus (SLE) patients demonstrate that HGAlign significantly reduces inter-batch discrepancies while maintaining or improving classification accuracy. Quantitative evaluation shows that our method achieves the lowest MMD values among state-of-the-art methods, and consistently improves classification metrics. Moreover, HGAlign avoids over-correction, enabling stable identification of cross-batch differential metabolites that retain biological interpretability. HGAlign offers a principled framework for balancing batch effect removal and biological signal preservation in high-throughput metabolomics. By introducing heterogeneous graph representation learning, it achieves superior performance in both batch correction and disease classification tasks, showing strong potential for large-scale clinical applications.