This study proposes a machine learning approach, LOCATE, to predict metabolite concentrations from microbiome composition and to produce a latent representation of the microbiome-metabolome interaction. The latent representation is then used to predict host condition. LOCATE outperforms existing methods in predicting metabolite concentrations and host condition. The latent representation is strongly correlated with host demographics and improves prediction accuracy, even with a small number of metabolite samples. The study shows that the microbiome-metabolome interaction is not linear and is dominated by a few taxa. The latent representation of the microbiome-metabolome interaction is more accurate in predicting host condition than either the microbiome or the metabolome alone. The study also shows that microbiome-metabolome relations are dataset-specific and that LOCATE performs better than existing methods in cross-dataset and cross-condition predictions. The latent representation is associated with host demographics and improves host condition prediction compared to microbiome or metabolome alone. LOCATE is available as a GitHub and PyPI package and can be used for both 16S and WGS datasets. The study highlights the importance of considering the complex interaction between the microbiome and metabolome in predicting host condition.This study proposes a machine learning approach, LOCATE, to predict metabolite concentrations from microbiome composition and to produce a latent representation of the microbiome-metabolome interaction. The latent representation is then used to predict host condition. LOCATE outperforms existing methods in predicting metabolite concentrations and host condition. The latent representation is strongly correlated with host demographics and improves prediction accuracy, even with a small number of metabolite samples. The study shows that the microbiome-metabolome interaction is not linear and is dominated by a few taxa. The latent representation of the microbiome-metabolome interaction is more accurate in predicting host condition than either the microbiome or the metabolome alone. The study also shows that microbiome-metabolome relations are dataset-specific and that LOCATE performs better than existing methods in cross-dataset and cross-condition predictions. The latent representation is associated with host demographics and improves host condition prediction compared to microbiome or metabolome alone. LOCATE is available as a GitHub and PyPI package and can be used for both 16S and WGS datasets. The study highlights the importance of considering the complex interaction between the microbiome and metabolome in predicting host condition.