Global Prioritization of Disease Candidate Metabolites Based on a Multi-omics Composite Network

Abstract

The identification of disease-related metabolites is important for a better understanding of metabolite pathological processes in order to improve human medicine. Metabolites, which are the terminal products of cellular regulatory process, can be affected by multi-omic processes. In this work, we propose a powerful method, MetPriCNet, to predict and prioritize disease candidate metabolites based on integrated multi-omics information. MetPriCNet prioritized candidate metabolites based on their global distance similarity with seed nodes in a composite network, which integrated multi-omics information from the genome, phenome, metabolome and interactome. After performing cross-validation on 87 phenotypes with a total of 602 metabolites, MetPriCNet achieved a high AUC value of up to 0.918. We also assessed the performance of MetPriCNet on 18 disease classes and found that 4 disease classes achieved an AUC value over 0.95. Notably, MetPriCNet can also predict disease metabolites without known disease metabolite knowledge. Some new high-risk metabolites of breast cancer were predicted, although there is a lack of known disease metabolite information. A predicted disease metabolic landscape was constructed and analyzed based on the results of MetPriCNet for 87 phenotypes to help us understand the genetic and metabolic mechanism of disease from a global view.

Figure 1. The flow chart of MetPriCNet.

(A) Construction of the multi-omics composite network. The multi-omics composite network is composed of six sub-networks. White circles indicate metabolites, white squares indicate genes, and white triangles indicate phenotypes. The thickness of an edge indicates the weight score. (B) The flow chart of MetPriCNet to optimize the candidate metabolite. First, the interested candidate metabolite and seed nodes are mapped to the multi-omics composite network. Then, a global extended RWR method is used to score the candidate metabolites according to their proximity to the seed nodes. Finally, the candidate metabolites are ranked by the scores. Orange circles represent the candidate metabolites of interest. Red triangles indicate the disease phenotype of interest (phenotype seed) from the OMIM data base, red squares represent known disease genes (gene seeds) from the OMIM database, and red circles indicate known disease metabolites (metabolite seeds) from the HMDB database.

Figure 2. The performance of MetPriCNet method.

(A)The ROC curve of the overall performance of MetPriCNet in whole-metabolome candidates. (B) The ROC curve of the overall performance of MetPriCNet in random candidates. (C) The ROC curve of the overall performance of PROFANCY method in whole-metabolome candidates. (D) The ROC curves of MetPriCNet method and PROFANCY method in 30 phenotypes with only two known metabolites. Red line indicates MetPriCNet method, blue line represents PROFANCY method. (E) The ROC curves of MetPriCNet method in 62 phenotypes with only one known metabolite. (F)The ROC curves of MetPriCNet method when only phenotype and corresponding disease genes were used as the seed nodes.

Figure 3. The comparison of performance between MetPriCNet method and PROFANCY method in various disease classes.

Red bar indicates MetPriCNet method and blue bar indicates PROFANCY method.

Figure 4

(A) The subnetwork of top-ranked risk metabolites, seeds and their first neighbors of breast cancer. Red indicates seed nodes, blue indicates the top-ranked risk metabolite glycerol, and pink indicates the neighbor nodes. Squares indicate genes, triangles indicate phenotypes and circles indicate metabolites. Only interaction scores above 0.6 in the whole composite network are retained. (B) The direct and indirect interaction (only two-step neighbors are considered) between the top 5 ranked risk metabolites and seeds in breast cancer. Only interactions with confidence scores above 0.6 are considered. Blue circles represent the top 5 risk metabolites identified by MetPriCNet. Red rectangles represent the seed nodes. Red lines indicate direct interactions, and blue lines indicate indirect interactions.

Figure 5. Global view of the predicted landscape of human diseases.

(A) Hierarchical clustering of the MetPriCNet scores between 87 phenotypes and 3,764 metabolites. The color of each cell represents the MetPriCNet score of a phenotype (column) and a metabolite (row), where red/blue indicates high/low MetPriCNet scores. Phenotype clusters are annotated with enriched disease categories (bottom), and metabolite clusters are annotated with the most enriched pathways of KEGG (right). The purple circled region indicates a module composed of the metabolite set of arginine/proline metabolism involved in a set of metabolic diseases. (B) Zoomed-in plot of the purple circled region, involving 3 metabolic diseases and 54 highly related metabolites. (C) The sub-network composite of the 3 phenotypes (blue triangles) and 54 highly related metabolites (red circles).