Prototypical transfer learning framework for small-sample ¹H NMR metabolomics: Accurate and interpretable...

Short stature (SS) is a common pediatric endocrine disorder, primarily including growth hormone deficiency (GHD) and idiopathic SS (ISS). Nuclear magnetic resonance (NMR)-based metabolomics can reveal metabolic signatures associated with SS. However, the limited availability of clinical data results in small sample sizes and high-dimensional features, which constrain the stability and generalizability of traditional methods. In this study, a prototypical transfer learning approach for small-sample metabolomics data is proposed. By integrating a shared feature encoder, cross-domain feature alignment, and a prototype learning strategy, knowledge from the source domain (child malnutrition data) is effectively transferred to the target domain (SS data), mitigating domain shift and enhancing classification performance. Experimental results show that the proposed method achieves classification accuracies of 96.76% and 95.12% on the preadolescent and adolescent SS datasets, respectively, substantially outperforming traditional machine learning and deep learning models trained solely on the target domain. Interpretability analysis using Integrated Gradients indicates that the model primarily focuses on chemical shift regions corresponding to disease-related metabolites, confirming the biological relevance of the discriminative features. This study provides a novel and interpretable framework for small-sample metabolomics analysis and clinical diagnosis of SS.