Fig 1.
Data flow for (A) predicting Association Scores and (B) analyzing disease-specific cytokine profiles. A. Network features of the high confidence STRING pairs were used to embed 14,707 human genes. The predicted associations between the 14,707 genes were validated using medium and low confidence STRING pairs. B. For each gene associated with a given disease, we calculated Association Scores with each of the 126 cytokines. The Association Scores were averaged and normalized to NAAS that represent the cytokine profile of the given disease. These profiles were further analyzed by (1) calculating Immune Scores and (2) analyzing subnetworks formed between pathogenesis and inflammation by employing network visualization, spectrum partition, and estimation of connection density.
Fig 2.
The predicted Association Score between two genes measures the confidence of their associations.
We calculated Association Scores for all possible pairs (108,140,571 pairs) of 14,707 human genes. A total of 9,250,034 of these pairs have a known STRING confidence index. The confidence indexes of known STRING pairs are shown in the four boxplots below, grouped by their predicted Association Scores. As the predicted Association Score increases (left to right), the average STRING confidence index also increases (low to high).
Fig 3.
The predicted Association Score between two genes measures the confidence of their associations.
The percentage of known STRING pairs above a certain confidence index cutoff. (Note: STRING confidence indexes are discrete scores).
Fig 4.
The 126 cytokines form six clusters based on their Association Scores with the 14,581 non-cytokine genes.
The six clusters are named after their most enriched types of cytokines: TGF-CLU (growth factors), Chemokine-CLU (chemokines), TNF-CLU (TNFs), IFN-CLU (interferons), IL-CLU (interleukins), and Unclassified-CLU. Based on the dendrogram of the hierarchical cluster tree, we identified six gene sets (SIG1-SIG6) that associate with the six individual clusters and two gene sets (BLU1, BLU2) that do not interact with the three major clusters (Chemokine-CLU, IFU-CLU, IL-CLU). Details of cytokines and signatures are in Tables 1–4.
Table 1.
The associations between the six gene sets (SIG1-SIG6) and specific clusters (CLU).
Table 2.
Specific cytokines in each of the six clusters.
Table 3.
The functional annotations of the six genes sets (SIG1-SIG6).
Table 4.
The two gene sets (BLU1, BLU2) that do not interact with the three major clusters Chemokine-CLU, IFU-CLU, and IL-CLU.
Fig 5.
Predicted cytokine profiles for 171 well-studied diseases correlate with cytokine sampling in literature.
The Spearman correlation coefficients between each disease’s NAAS and known literature sampling frequency are plotted against the P-values. Of the 171 diseases, we were able to predict the profiles for 95 diseases with p-value<0.0003 (corrected cutoff by multiple testing), suggesting the accuracy of the predicted profiles.
Fig 6.
The NAAS between aneurysm and each of the 79 cytokines for which the literature sampling frequency in disease is known in ImmuneXpresso.
Known associations (frequency cutoff of 0.005 in ImmuneXpresso) are marked in solid blue squares.
Fig 7.
Cytokine features for the 171 well-studied diseases.
The 171 diseases formed three clusters based on their NASS with different types of cytokines. Immune disorders are enriched in cluster-1. Infections are split into two clusters (cluster-1 and cluster-2). Note that of the twenty neoplasms in cluster-1, nineteen are hematic and lymphatic diseases (C15/C04). Most metabolic diseases (11/13) and cardiovascular disorders (11/17) are enriched in cluster-3. Note that diseases of other classes are not counted in the labels. Cluster details are in Table D in S1 Text.
Fig 8.
Immune Scores of five disease classes (23 immune disorders, 48 infections, seventeen cardiovascular, thirteen metabolic, 55 neoplasms).
For each class, the average NAAS between its diseases and the cytokines within four categories are plotted: 47 inflammation related cytokines, 37 chemokines, thirteen growth factors, and 29 other cytokines. The chemokine scores for immune disorders are spread in a wide range. Growth factors have the highest scores in infections. Cardiovascular diseases have higher scores than metabolic diseases over the three groups of cytokines. Neoplasms show the lowest scores for all four categories.
Table 5.
Analysis of subnetworks formed by high confidence associations between the known disease associated genes and the predicted cytokines of five diseases.
The known DisGeNET genes (column #2) of a given disease often contain cytokine receptors. The number of cytokine receptors and other disease genes captured by high-confidence associations (column #3) is listed in column #4 and column #5, respectively. The number of predicted essential cytokines that interact with receptors and disease genes from DisGeNET is listed in column #6. The Immune Connection Density (ICD) estimated on the subnetworks formed by receptors, disease genes, and essential cytokines is shown in column #7.
Table 6.
The recall rate (column #4) of these cytokines being recognized by the predicted subnetworks formed by high-confidence associations ranges from 50% to 88%.
Fig 9.
Disease-specific cytokine profiles of five immune disorders.
The Y-axis shows the Probability of Association between each cytokine and the five immune disorders: rheumatoid arthritis (RA), psoriasis (PS), ulcerative colitis (UC), Crohn’s disease (CD) and systemic lupus erythematosus (SLE). A conserved association pattern is observed in inflammation-related cytokines, while differential patterns are observed in other types of cytokines.
Fig 10.
SLE subnetwork formed between pathogenesis genes (green and purple) and inflammatory responses (orange, red, blue). The graph was plotted using a force-directed layout that uses attractive forces between adjacent nodes and repulsive forces between distant nodes. The distances between two vertices are roughly proportional to the length of the shortest path between them. Six genes (ACKR3, HRH4, HTR1, GAL, GRM3, S1PR1) in Box-C are making high degree contacts with the chemokine core (red box), with ANXA1 interacting with 16 chemokines. Interactions with the inflammation core (orange box) appear in multiple directions. Seven pathogenesis genes (Box-I-1) interact with the inflammation core (orange box) directly or through receptors (green nodes). Nine disease genes in Box-I-3 form a small core with four cytokines (IL18, IL22, IL1A and IL1B). Two other groups of genes (Box-I-2 and Box-I-4) appear distant from the cytokine core but are linked to the TNFs, as they cannot overcome the repulsive force to association with the center of inflammation responses.
Table 7.
Pathogenesis genes in the highly connected modules that were identified by spectrum partition on the subnetworks formed by pathogenesis genes, receptors, and cytokines, in the context of five immune disorders: rheumatoid arthritis (RA), psoriasis (PS), ulcerative colitis (UC), Crohn’s disease (CD) and systemic lupus erythematosus (SLE).
The table shows that 36 disease genes were identified out of 1,340 disease-associated genes for RA (2.7%), four disease genes from the 542 genes for PS (0.7%), 35 disease genes from the 793 genes for SLE (4.4%), eight disease genes from the 654 genes for UC (1.2%), and 18 disease genes from the 622 genes for CD (3%). Note that many of these disease-associated genes are related to immune responses.