The power of deep learning compromises when applied to biological problems with sparsely labeled data and a data distribution shift. We developed a highly data-efficient model-agnostic semi-supervised meta-learning framework DESSML to address these challenges, and applied it to investigate understudied interspecies metabolite-protein interactions (MPI). Knowledge of interspecies MPIs is crucial to understand microbiome-host interactions. However, our understanding of interspecies MPIs is extremely poor due to experimental limitations. The paucity of experimental data also hampers the application of machine learning. DESSML successfully explores unlabeled data and transfers the information of intraspecies chemical-protein interactions to the interspecies MPI predictions. It achieves three times improvement in the prediction-recall over the baseline model. Using DESSML, we reveal novel MPIs that are validated by bioactivity assays and fill in missing links in microbiome-human interactions. DESSML is a general framework to explore previously unrecognized biological domains beyond the reach of present experimental techniques.