MicroDIG: Estimating directed interactions in large-scale microbiome longitudinal data based on self-supervised learning

Species within microbial communities influence and depend on each other through complex interactions, thereby maintaining a dynamic balance. Investigating these interactions is crucial for understanding the functions and underlying mechanisms of microbial communities. In this paper, we introduce a self-supervised learning framework, the MicroDIG (microbiome directed interaction on GLV), which utilizes longitudinal abundance data to explore species interactions within large-scale microbial communities. The MicroDIG combines mathematical modeling with self-supervised learning to estimate and validate directed interactions between species. It leverages the GLV (generalized Lotka-Volterra) equations for data simulation and augmentation, and employs an autoencoder to infer the strengths of directed interactions between species. Additionally, it implements a statistical testing strategy to validate these interaction strengths, identifying high-confidence directed interactions. Through comprehensive evaluation on both simulated data and real human gut microbiome data, the MicroDIG has been shown to outperform five baseline methods. Specifically, the MicroDIG effectively reduces false positives in its estimations, accurately identifies the directionality of species interactions, and rejects interactions that arise from transient relationships. The MicroDIG is applicable to communities of various sizes, particularly large-scale ones that traditional methods struggle to handle. The self-supervised learning framework proposed in the MicroDIG does not require training data with ground truth, offers clear interpretability, and is highly flexible and scalable. It can be extended to explore more intricate interactions in real-world, large-scale, and complex microbial communities.