Background: Resting metabolic rate (RMR) is modulated by a variety of factors. Accurate prediction of RMR is essential for planning energy requirements but remains challenging due to inter-individual variability. Objective: This study aimed to develop and evaluate machine learning models for predicting RMR using comprehensive data from the cross-sectional enable study and to identify the most predictive and stable features across different study populations. Methods: RMR was predicted using data from 454 participants of the enable phenotyping platform (Freising and Nuremberg cohort). We systematically compared linear and nonlinear machine learning models trained on either the full set of 94 predictors or a reduced set of routinely accessible variables, including sex, age, body weight, fat mass, and fat-free mass. Model performance was assessed by cross-validation. The best-performing model (Lasso) was further evaluated on independent test datasets from other cohorts. Feature importance and stability were assessed using repeated cross-validation and marginal variance decomposition. Results: Lasso regression consistently outperformed other models, particularly when trained on the enable cohort feature set. The final model explained 76.8% of RMR variance in the Freising cohort. Key predictive features included fat-free mass, body weight, and mean outdoor temperature. Blood-based features contributed marginally, whereas microbiota and fecal short chain fatty acids variables did not contribute to explaining RMR. Conclusion: This novel prediction model for RMR shows improved accuracy in comparison to traditional models. While microbiota composition did not contribute to explain the residual variation in RMR, the inclusion of clinical blood parameters and outdoor temperature improved predictive performance.