Predicting individual responses to weight loss interventions is a significant challenge. While energy balance, macronutrient composition, anthropometrics, and glycemic status have been explored, individual variability remains substantial. Multi-omics data offers a promising avenue to capture the complex interplay between host genetics, gut microbiome, metabolome, and diet. Previous research suggests associations between obesity and gut microbiome composition, plasma metabolome, and host genome. Integrating these omics datasets holds potential for improved prediction, although computational challenges related to data heterogeneity, high dimensionality, small sample sizes, and missing data exist. Machine learning, particularly random forests and ensemble methods, offer robust approaches for handling these complexities and improving predictive accuracy. This study builds upon previous research involving two Danish randomized crossover trials examining the impact of whole grain-rich and low-gluten diets on metabolic health. These trials showed varying weight loss responses despite overall significant weight reduction in both intervention arms compared to a refined grain control diet. This study aims to leverage machine learning to predict individual weight loss response based on baseline biomarkers before any dietary intervention. By integrating anthropometry, blood serum markers, gut microbiome markers, urine metabolomics, and host genomics, the study seeks to improve prediction accuracy and understand individual predisposition to weight loss.

Existing research on weight loss prediction has primarily focused on factors such as energy intake and expenditure, macronutrient balance, anthropometrics, glycemic and insulinemic status, and gut microbiome profiles. Studies have shown correlations between specific gut microbial compositions (e.g., *Prevotella*-to-*Bacteroides* ratio) and weight loss outcomes. However, these approaches often fail to capture the full complexity of individual variation. The application of multi-omics data integration has shown promise in improving our understanding of complex phenotypes like metabolic health. Studies have identified associations between obesity and various omics datasets; however, the translation of these findings into robust individual-level predictions remains a challenge. The integration of multi-omics data is hampered by data heterogeneity, the high dimensionality of data, the limitations imposed by small sample sizes, and the presence of missing data. Computational methods, particularly machine learning approaches, are necessary to effectively analyze and model this complex data.

This study utilized data from two previously conducted randomized controlled dietary trials in Denmark, one involving a whole grain-rich diet and the other a low-gluten diet, both compared to a refined grain control diet. A total of 203 participants completed the trials. Weight loss responders (N=106) and non-responders (N=97) were defined based on changes in body weight over an 8-week intervention period. Data included anthropometry, physiology (blood pressure, cytokines, gut permeability, blood serum markers), urine metabolome (GC-MS and LC-MS), gut microbiome (16S rRNA amplicon sequencing and shotgun metagenomics), and host genomics (CoreExome-24 BeadChip). Random forest models were trained using 50 shuffle-split fivefold cross-validation to ensure robustness. Feature engineering involved both prior knowledge-driven and data-driven selection methods. Prior knowledge was incorporated by prioritizing biomarkers related to metabolic pathways, inflammation, and gut microbiome composition. Data-driven feature selection used ReliefF, forward selection, and pair/triplet combinations of metabolites to identify informative features. Several models were trained with different combinations of data types, including diet, clinical features, SNPs, microbiome markers, metabolomic data, and postprandial response. Model performance was evaluated using ROC-AUC, sensitivity, specificity, and MCC. An ensemble model was created to integrate the predictions from multiple high-performing models. To improve model robustness and handle missing data, an ensemble of models was created, integrating predictions based on various confident score ranges.

A diet-only model achieved a ROC-AUC of 0.62, indicating limited predictive power. However, models integrating gut microbiome and urinary metabolome data significantly improved performance (ROC-AUC: 0.84–0.90). Models incorporating features from the 16S rRNA-based OTUs and urine metabolites (Diet.16S_B.LC-MS) reached a ROC-AUC of 0.86, while models using butyrate-producing species from the MGmapper Bacteria draft database and urine metabolites (Diet.MGm_B1.LC-MS) reached a ROC-AUC of 0.90. The most important features included bacterial taxa from the family Ruminococcaceae and the genus Streptococcus. For MGmapped gut microbiome species, *F. prausnitzii*, *E. ramulus*, and *R. faecis* were important predictors. The ensemble model, which integrated multiple models with high predictive power, achieved a ROC-AUC of 0.86. When using only confident predictions (s ≤ 0.25 or s ≥ 0.75), the ensemble model correctly classified 64% of non-responders with only 17% false negatives. Using a threshold of s=0.70 for weight loss responders, the model correctly identified 61% of responders with 26% false positives. Excluding microbiome and metabolome data reduced the ensemble model’s performance (ROC-AUC: 0.72), highlighting the crucial role of these features in accurate prediction.

This study demonstrates the feasibility of predicting weight loss response to dietary interventions using a multi-omics approach and machine learning. The significant improvement in predictive performance by integrating microbiome and metabolome data, compared to diet alone, highlights the complex interplay of host factors and diet in weight management. The identification of specific bacterial taxa and urine metabolites as strong predictors offers valuable insights into potential biological mechanisms driving weight loss variability. The ensemble model's robust performance and resilience to missing data suggests its practical applicability for personalizing weight loss strategies. However, the study's findings should be interpreted within the context of the study's limitations. The identification of 64% of non-responders with 80% confidence suggests the potential for AI models to assist in tailoring weight loss strategies to individual needs. The identified biomarkers may serve as valuable targets for future research to further elucidate the mechanisms underlying individual weight loss response.

This study successfully demonstrated that integrating multi-omics data with machine learning can significantly improve the prediction of weight loss response to dietary interventions. Gut microbiome and urine metabolome features were particularly important, improving prediction accuracy compared to diet alone. The ensemble model provides a robust and resilient approach for personalized prediction, especially when confident predictions are considered. Future research should focus on validating these findings in larger, independent cohorts and further investigating the biological mechanisms underlying the identified associations to refine the predictive model and inform the development of truly personalized weight management strategies. Incorporating additional factors like exercise and detailed dietary information will enhance the model’s predictive power. Longitudinal body weight measurements to account for normal fluctuations should be incorporated in future studies.

Several limitations should be considered. The sample size, while deeply phenotyped, is relatively small for complex data integration. The definition of weight loss responders and non-responders based on a single point cut-off might not adequately capture clinically significant weight loss in all individuals. The lack of data on exercise habits and specific dietary fiber composition could influence the results. Finally, the study lacks external validation, and future studies should replicate these findings in independent cohorts. Many important urine metabolites lacked detailed annotation, limiting the interpretation of their predictive value.