Spectral divergence prioritizes key classes, genes, and pathways shared between substance use disorders and cardiovascular disease

Abstract

Introduction: Substance use disorders (SUDs) are heterogeneous diseases with overlapping biological mechanisms and often present with co-occurring disease, such as cardiovascular disease (CVD). Gene networks associated with SUDs also implicate additional biological pathways and may be used to stratify disease subtypes. Node and edge arrangements within gene networks impact comparisons between classes of disease, and connectivity metrics, such as those focused on degrees, betweenness, and centrality, do not yield sufficient discernment of disease network classification. Comparatively, the graph spectrum's use of comprehensive information facilitates hypothesis testing and inter-disease clustering by using a larger range of graph characteristics. By adding a connectivity-based method, network rankings of similarity and relationships are explored between classes of SUDs and CVD.

Methods: Graph spectral clustering's utility is evaluated relative to commonly used network algorithms for discernment between two distinct co-occurring disorders and capacity to rank pathways based on their distinctiveness. A collection of graphs' structures and connectivity to functionally identify the relationship between CVD and each of four classes of SUDs, namely alcohol use disorder (AUD), cocaine use disorder (CUD), nicotine use disorder (NUD), and opioid use disorder (OUD) is evaluated. Moreover, a Kullback-Leibler (KL) divergence is implemented to identify maximally distinctive genes (D ^g ). The emphasis of genes with high D ^g enables a Jaccard similarity ranking of pathway distinctiveness, creating a functional "network fingerprint".

Results: Spectral graph outperforms other connectivity-based approaches and reveals interesting observations about the relationship among SUDs. Between CUD and CVD, the gamma-aminobutyric acidergic and arginine metabolism pathways are distinctive. The neurodegenerative prion disease and tyrosine metabolism are emphasized between OUD and CVD. The graph spectrum between AUD and NUD to CVD is not significantly divergent.

Conclusion: Graph spectral clustering with KL divergence illustrates differences among SUDs with respect to their relationship to CVD, suggesting that despite a high-level co-occurring diagnosis or comorbidity, the nature of the relationship between SUD and CVD varies depending on the substance involved. The graph clustering method simultaneously provides insight into the specific biological pathways underlying these distinctions and may reveal future basic and clinical research avenues into addressing the cardiovascular sequelae of SUD.

Keywords: cardiovascular disease; disease-associated prioritization; functional fingerprint; graph spectrum; substance use disorder.

Figure 1

The framework of prioritizing pathways and genes using spectral clustering and a KL divergence. (A) Disorder-associated genes sets derived from humans, acquired from DisGeNET and published sources, are prioritized from experimental analyses or from database mining and are subsequently enriched for KEGG pathways. (B) Enriched pathways' KGML files are then parsed in KNeXT as functional gene-gene networks. (C) KNeXT-generated gene networks are hierarchically clustered through spectral clustering. (D) Post clustering, individual genes are assessed through KL divergence against an opposing cluster, dotted green lines. (E) Genes in cluster k are compared to genes in cluster l and then the comparison is reversed where genes in cluster l are compared to genes in cluster k. All genes with high D^g are compared to all pathways within its origin cluster. (F) The results of this framework are Jaccard scores for all pathways in each cluster. KEGG pathways with a high Jaccard score have an abundance of top D^g genes, which in turn, is driving distinction between clusters.

Figure 2

Comparison of spectral clustering to commonly used algorithms. DisGeNET-derived pathway groups include cocaine use disorder (CUD) and opioid use disorder (OUD) compared to cardiovascular disease (CVD). Benchmarked groups (BEN) are groups derived from BRITE terms for nervous system classes and surveyed CVD pathways. Spectral clustering outperformed all other algorithms.

Figure 3

Agglomerative hierarchical clustering for CUD vs. CVD. The top driving genes are genes that have a high D^g and are listed according to their cluster, which is color and shape coordinated. Furthermore, pathways with a high Jaccard index are bolded. Pathway hsa00220 is arginine biosynthesis and hsa04727 is GABAergic synapse.

Figure 4

The results of the agglomerative hierarchical clustering for OUD vs. CVD. The top driving genes are genes which have a high D^g and are listed according to their cluster, which is color coordinated. Furthermore, pathways with a high Jaccard index are bolded. Pathway hsa03050 is tyrosine metabolism and pathway hsa05020 is prion disease.

Figure 5

All similarity results sorted by KEGG class. Pathways that cluster separately are highlighted in yellow. (A) Similarity results for CUD. As shown, two metabolism pathways diverged compared to the rest of the results. (B) Similarity results for OUD. For OUD, metabolism and two human disease pathways drove the cluster separation. (C) The differences in magnitude of Jaccard similarity between CUD and OUD. As shown, metabolism plays the largest role in the differences between both CUD and OUD.