DeepGCN-FP: An Explainable Transfer Learning Approach to Predict Adduct Ion Formation for Enhanced Metabolite Annotation in Spatial Metabolomics

Accurate metabolite annotation remains a major bottleneck in spatial metabolomics because MALDI mass spectrometry imaging (MALDI-MSI) lacks retention time and other orthogonal descriptors. We investigated whether reproducible adduct ion formation could provide complementary evidence for annotation. Profiling 753 authentic standards under fixed matrix conditions identified 10 prevalent adduct types in negative mode and 8 in positive mode, including often-overlooked matrix-derived ions involving 1,5-diaminonaphthalene (DAN). We developed DeepGCN-FP, a transfer-learning framework that extracts graph convolutional network fingerprints (GCN-FPs) from a pretrained DeepGCN model and combines them with machine learning classifiers to predict adduct formation. GCN-FP models with gradient-boosted decision trees outperformed extended-connectivity fingerprints, achieving cross-validated AUCs of 0.86−0.95 for all 10 negative-mode adducts and 0.80−0.86 for 6 of 8 positive-mode adducts. Atom-level interpretation highlighted functional groups associated with deprotonation, while on-slice experiments showed that adduct profiles were preserved on tissue, and titration experiments found that low-concentration adduct repertoires were subsets of those at higher concentrations. On 19 independent standards, predictions achieved 87% precision. Embedded in an adduct-aware scoring framework for metabolite annotation, adduct prediction routinely elevated the rank of correct metabolites in both standard mixtures and real biological samples and yielded pathway-level enrichments in mouse skeletal muscle that match the known biochemical differences between slow-twitch and fast-twitch myofibers, including carnosine- and pyruvate-related pathways that mass-only annotation failed to recover. Adduct prediction thus provides a practical, chemically interpretable aid for metabolite annotation in spatial metabolomics. Our code is available at: https://github.com/dadazhetang/DeepGCN-FP.