The gut microbiome is of utmost importance to human health. While a healthy microbiome can be represented by a variety of structures, its functional capacity appears to be more important. Gene content of the community can be assessed by “shotgun” metagenomics, but this approach is still too expensive. High-throughput amplicon-based surveys are a method of choice for large-scale surveys of links between microbiome, diseases, and diet, but the algorithms for predicting functional composition need to be improved to achieve good precision. Here we show how feature engineering based on microbial phenotypes, an advanced method for functional prediction from 16S rRNA sequencing data, improves identification of alterations of the gut microbiome linked to the disease. We processed a large collection of published gut microbial datasets of inflammatory bowel disease (IBD) patients to derive their community phenotype indices (CPI)—high-precision semiquantitative profiles aggregating metabolic potential of the community members based on genome-wide metabolic reconstructions. The list of selected metabolic functions included metabolism of short-chain fatty acids, vitamins, and carbohydrates. The machine-learning approach based on microbial phenotypes allows us to distinguish the microbiome profiles of healthy controls from patients with Crohn's disease and from ones with ulcerative colitis. The classifiers were comparable in quality to conventional taxonomy-based classifiers but provided new findings giving insights into possible mechanisms of pathogenesis. Feature-wise partial dependence plot (PDP) analysis of contribution to the classification result revealed a diversity of patterns. These observations suggest a constructive basis for defining functional homeostasis of the healthy human gut microbiome. The developed features are promising interpretable candidate biomarkers for assessing microbiome contribution to disease risk for the purposes of personalized medicine and clinical trials.