Enteromatch: A Sparse MoE Model for FMT Matching

Fecal Microbiota Transplantation (FMT) has emerged as a promising therapeutic approach for various gastrointestinal and systemic diseases. However, optimizing do-nor-recipient matching remains a critical challenge that constrains its clinical efficacy. In this study, we propose EnteroMatch, a deep learning-based Sparse Mixture of Experts (Sparse MoE) model designed for FMT donor-recipient matching. By integrating a dynamic routing mechanism, EnteroMatch effectively captures the complex ecological characteristics of the gut microbiota. Furthermore, we employ k -means clustering to partition both donors and recipients into two distinct enterotypes, allowing the model to adaptively adjust to their unique microbial profiles and better reflect the influence of microbiome diversity on FMT outcomes. Extensive experimental evaluations on large-scale datasets demonstrate that EnteroMatch outperforms other state-of-the-art deep learning architectures in terms of matching accuracy, generalization, and robustness. This work not only provides a novel computational framework for personalized FMT strategies but also lays a solid foundation for future research in microbiome-based therapies.