Multiblock Integration and Modeling of Localized Microbiome, Metabolome, and Clinical Metadata to Identify Biomarkers Predictive of Outcome in Veterans with Non-Healing Wounds
Abstract
Background
Type 2 Diabetes affects more than 37 million people in the United States and is the number one cause of lower-limb amputation in adults due to diabetic foot ulcers (DFU). The chronic wound microenvironment consists of a complex milieu of host cells, microbial species, and metabolites. While much is known about the wound microbiome, our knowledge of the metabolic landscape and its influence on microbial diversity and wound healing is limited. Furthermore, the integration of these complex datasets into a predictive model with relevance to clinical outcome is almost non-existent. Here, we present a multiomics data analysis coupled with machine-learning cross validation of microbiome and metabolome profiles from human chronic wounds. The model was integrated with patient metadata to determine predictive correlation to clinical outcome under standard of care.
Methods
Microbial ribosomal RNA (rRNA) and total metabolites were extracted from 45 DFU debridement samples collected from 13 patients at the Boise VA Medical Center. Of 45 samples analyzed, 25 samples were isolated from wounds that failed to respond to standard treatment while the remaining 20 samples were taken from wounds that progressed to healing and remained closed for >30 days. 16S rRNA sequencing and global metabolomics were performed and clinical metadata was collected from patient records. Healing outcome was modeled as a function of three blocks of features (N = 21 clinical, 634 microbiome, and 865 metabolome) using DIABLO (Data Integration Analysis for Biomarker Discovery using Latent Components) based on multiblock sparse partial least squares discriminant analysis (sPLS-DA) which performs feature selection using LASSO regularization. Seven-fold cross-validation with 100 repeats was used to find the amount of regularization associated with the smallest predictive error.
Results
The final model selected a total of 176 features (N = 15 clinical, 8 microbiome, and 153 metabolome) and was able to predict the clinical outcome with an overall error rate of 6.44%.
Conclusion
These results indicate that the integration of wound microbiome and metabolomics data with patient clinical metadata can be utilized to predict clinical outcomes regarding wound healing and with low error rates. Furthermore, the biomarkers selected within the model may offer novel insights into wound microenvironment composition, reveal innovative therapeutic approaches, and improve treatment efficacy in difficult to heal wounds.
Related articles
Related articles are currently not available for this article.