Search-based health status detection and disease classification using species-level profiles of metagenomes

Abstract

Microbiome biomarker-based modeling has been widely used in classifying health states. However, many diseases do not have explicit biomarkers, or exhibit shortages in detection accuracy using specific species. Based on microbiome big data and cutting-edge computing engine, here we report the search-based strategy of health status detection for shotgun metagenomes. Comparing the species-level profiles against large-scale metagenomes, outlier samples are screened out as unhealthy, and their detailed disease types can be identified by top matches. Benchmarking on a multi-cohort dataset with over 3,000 metagenomes, the search-based approach achieved a promising overall accuracy that was superior to marker-based models constructed by random forest (RF), supporting vector machine (SVM) and extreme gradient boosting (XGBoost). More importantly, the search-based method also featured a balanced performance on different diseases. Hence, this case study further demonstrates the potential and capability of metagenome big data in human health, as well as moves one-step forward of search-based approach in microbiome research and application.