Gene Ontology Analysis Highlights Biological Processes Influencing Non-Response to Anti-TNF Therapy in Rheumatoid Arthritis

Anti-TNF therapy has significantly improved disease control in rheumatoid arthritis, but a fraction of rheumatoid arthritis patients do not respond to anti-TNF therapy or lose response over time. Moreover, the mechanisms underlying non-response to anti-TNF therapy remain largely unknown. To date, many single biomarkers of response to anti-TNF therapy have been published but they have not yet been analyzed as a system of interacting nodes. The aim of our study is to systematically elucidate the biological processes underlying non-response to anti-TNF therapy in rheumatoid arthritis using the gene ontologies of previously published predictive biomarkers. Gene networks were constructed based on published biomarkers and then enriched gene ontology terms were elucidated in subgroups using gene ontology software tools. Our results highlight the novel role of proteasome-mediated protein catabolic processes (p = 2.91 × 10−15) and plasma lipoproteins (p = 4.55 × 10−11) in anti-TNF therapy response. The results of our gene ontology analysis help elucidate the biological processes underlying non-response to anti-TNF therapy in rheumatoid arthritis and encourage further study of the highlighted processes.


Introduction
Rheumatoid arthritis (RA) is a common complex autoimmune disease characterized by chronic and progressive joint inflammation. Currently, first-line therapeutic approaches in rheumatoid arthritis focus on minimizing disease activity using, primarily, corticosteroids with or without disease-modifying antirheumatic drugs (DMARDs). The development of biological drugs such as monoclonal antibodies against key inflammatory cytokines has significantly improved symptom control [1] in severe rheumatoid arthritis and chronic patients failing first-line therapy. Etanercept [2] and infliximab, inhibitors of proinflammatory cytokine tumor necrosis factor alpha (anti-TNF) [3], were the first anti-TNF biological drugs indicated for rheumatoid arthritis, and later more biological drugs against TNFα were developed, including adalimumab [4], certulizumab pegol [5] and golimumab [6]. In recent years, the emergence of biosimilars of anti-TNF biological drugs has also somewhat reduced the initially high cost of anti-TNF therapy while maintaining efficacy levels comparable to those of the originator biological drugs [7].
However, despite the immense therapeutic power of anti-TNF therapy, 10-30% of patients do not respond to anti-TNF biological drugs upon therapy initiation (i.e., primary non-response) and 23-46% of responders lose response to anti-TNF therapy over time (i.e., secondary non-response) [8]. Non-response to anti-TNF therapy usually represents loss of disease control in patients with severe rheumatoid arthritis, as well as unnecessary exposure to potentially severe adverse effects of anti-TNF drugs and inefficient use of expansive biological therapeutics. Patients who fail to respond to anti-TNF drugs may switch to a different biological drug, such as anakinra, rituximab or sarilumab [9]. Even so, other biological drugs face similar challenges to anti-TNF drugs in terms of non-response [1,10,11]. Therefore, disease-modifying antirheumatic drugs (DMARDs) remain the long-term therapy of choice alongside corticosteroids for disease flares, both of which are known to have significant long-term adverse effects [12].
Predicting non-response to anti-TNF therapy based on the patient's clinical and biological data would allow targeted therapy with higher efficacy and fewer adverse effects, as well as cost-efficient use of therapeutics. Physicians could determine if and when to switch anti-TNF therapeutics or whether it would be more effective to switch to biological drugs with different therapeutic targets. To date, response to anti-TNF therapy has been intensively studied and several DNA, RNA and protein response biomarkers with low to moderate predictive accuracy have been identified. However, despite the many published anti-TNF response biomarkers, the biological processes underlying non-response to anti-TNF therapy in RA remain largely unknown. Improving the understanding of the mechanisms underlying non-response to anti-TNF drugs on a molecular level would allow the development of novel therapeutic strategies to prevent non-response or the discovery of novel pharmaceutical targets for drug development. To this end, we reviewed already published genomic, transcriptomic and proteomic markers of response and non-response to anti-TNF biological drugs in rheumatoid arthritis and performed a gene ontology analysis to help elucidate biological processes linked to response and non-response to anti-TNF therapy.

Literature Search
To perform a comprehensive review of the literature on anti-TNF therapy response biomarkers, we searched the PubMed database using a combination of terms defining disease, drug, response, biomarker type and exclusion criteria. To prevent Mesh terms missing synonyms, we employed a combination of both Mesh terms and equivalent non-Mesh keywords. The final search query was defined as a combination of the following term groups:
Studies were included based on the following inclusion criteria: • Published between the years 2002 and 2022; • The study used well-defined response criteria (e.g., those included in the Disease Activity Score in 28 Joints, also known as ∆DAS28); • Biomarkers were analyzed prior to therapy initiation and, if applicable, after therapy (e.g., gene expression and serum protein levels); • Quantitative biomarkers were reported with a clearly defined direction of association (e.g., gene expression defined as up-regulated or down-regulated, not merely "associated").
In this gene ontology study, we did not make any additional distinctions based on the anti-TNF drugs used or on whether patients were anti-TNF naive or not.

Subset Definition
Subsets for gene ontology (GO) analysis were defined based on biomarker type. Preliminary subset analysis revealed no significant differences between the gene ontology terms of biomarkers measured in synovial fluid and those measured in sera. For this reason, we did not make any distinctions based on biomarker measurement locations.
Potential therapeutic targets can be either stimulated or blocked. In general, processes that are up-regulated in responders or down-regulated in non-responders could be stimulated to achieve better response or even restore response. Similarly, processes that are downregulated in responders or up-regulated in non-responders can be blocked. Following this reasoning, we created two additional separate groups for RNA and protein biomarkers. The first group (_UP_R_DO_N) contains biomarkers reported either as up-regulated in responders or down-regulated in non-responders; the second group (_DO_R_UP_N) contain biomarkers down-regulated in responders or up-regulated in non-responders.
To enhance biological process discovery with gene ontology analysis, gene networks were constructed. In this study, "gene network" refers to a set of interacting biomarkers produced from a list of biomarkers of interest (i.e., previously published anti-TNF response biomarkers). Biomarkers interacting with at least two biomarkers of interest were obtained from BIOGRID [13,14] using the biogridR package [15] for R (version 4.1.1, R Core Team, Vienna, Austria) [16].
Subset names are defined in Table 1. Moreover, comparative gene ontology analysis was employed to estimate GO term specificity between different subsets (e.g., _UP_RE_DO_NR vs. _UP_NR_DO_RE).
Gene ontology analysis results were visualized using default CytoScape settings and freely available style options.

Literature Search
Using the defined search query (see Materials and Methods-Literature Search), we obtained 185 results in the PubMed database. Based on the inclusion criteria, 125 studies were included in the gene ontology analysis. Among the 125 studies, 61 studies reported DNA biomarkers, 15 studies reported RNA biomarkers, 39 studies reported protein biomarkers, while 10 studies reported response biomarkers that could not be categorized as DNA, RNA or protein biomarkers as they were cell counts, nuclear magnetic resonance (NMR) spectra or metabolomic markers. In addition, five studies reported biomarkers at several molecular levels.
Use of technologies to comprehensively study the genome, transcriptome and proteome remains uncommon, but it has become more common in recent years. Among the 61 DNA biomarker studies, 8 employed next-generation sequencing (NGS) technology and 3 out of 15 RNA biomarker studies employed RNA sequencing (RNAseq). Similarly, 7 out of 39 protein biomarker studies used liquid chromatography with mass spectrometry (LC-MS/MS) for biomarker discovery.

Biomarker Collection
The biomarkers extracted from the studies gathered from the literature are shown in Table 2 (DNA biomarkers), Table 3 (RNA biomarkers) and Table 4 (protein biomarkers). For gene ontology (GO) analysis, only biomarkers indexed in GO datasets can be processed. To remove potential duplicate biomarkers and obsolete gene names, we used the g:Convert Gene ID Converter tool [19] to update the biomarker names to the most recent ones. Finally, biomarkers that could not be reliably assigned to a gene with GO definitions were excluded (e.g., intergenic genetic variants).    CCL4

BCL2A1
Up-regulated in responders

Up-regulated in non-responders
LGALS3BP Up-regulated in non-responders

MX2
Up-regulated in non-responders            Studies reporting biomarkers that could not be categorized as DNA, RNA or protein biomarkers are displayed below in Table 5.

Gene Ontology Analysis Results
The DNA subset has enriched GO terms related to the definition of non-response, while the DNA gene network only expanded upon the terms NF-κB signaling and TNF-α processes.
Gene ontology analysis of DNA biomarkers revealed terms already known to be associated with anti-TNF therapy non-response in rheumatoid arthritis, namely, terms connected to the definition of non-response or anti-TNF therapy, such as inflammation, tumor necrosis factor alpha, NF-κB signaling, IL-1, IL-2, IL-6 and IL-27. A subset of the terms related to NF-κB signaling is displayed in Figure 1.
Gene ontology analysis of DNA biomarkers revealed terms already known to be associated with anti-TNF therapy non-response in rheumatoid arthritis, namely, terms connected to the definition of non-response or anti-TNF therapy, such as inflammation, tumor necrosis factor alpha, NF-κB signaling, IL-1, IL-2, IL-6 and IL-27. A subset of the terms related to NF-κB signaling is displayed in Figure 1. RNA biomarker subsets revealed several enriched GO terms that were not previously directly associated with anti-TNF therapy response in rheumatoid arthritis. Such enriched terms in RNA subsets include prostaglandin synthesis, response to lipopolysaccharide (LPS), interferon gamma and macrophage chemotaxis. Gene networks based on RNA biomarkers and their BIOGRID interactors revealed novel significantly enriched GO terms related to the proteasome; the term proteasome-mediated ubiquitin-dependent protein catabolic process (p = 2.91 × 10 −15 ) is a significant novel hyponym. The gene ontology terms related to the proteasome and others identified in the BIOGRID RNA biomarker network are illustrated in Figure 2. RNA biomarker subsets revealed several enriched GO terms that were not previously directly associated with anti-TNF therapy response in rheumatoid arthritis. Such enriched terms in RNA subsets include prostaglandin synthesis, response to lipopolysaccharide (LPS), interferon gamma and macrophage chemotaxis. Gene networks based on RNA biomarkers and their BIOGRID interactors revealed novel significantly enriched GO terms related to the proteasome; the term proteasome-mediated ubiquitin-dependent protein catabolic process (p = 2.91 × 10 −15 ) is a significant novel hyponym. The gene ontology terms related to the proteasome and others identified in the BIOGRID RNA biomarker network are illustrated in Figure 2.
Similarly, protein subsets also revealed several enriched GO terms that were not previously directly associated with anti-TNF therapy response in rheumatoid arthritis. Gene ontology analysis revealed several enriched blood lipoprotein (HDL, VLDL and cholesterol) terms, illustrated in Figure 3.
The full results of the gene ontology subset analysis are available in Table S1. BIOGRID data gene networks based on DNA and protein biomarkers did not reveal any novel enriched GO terms but expanded the associated hyponyms of leading GO terms.
Comparative GO analysis of DNA, RNA and protein biomarkers showed no novel differences between analyzed subsets based on biomarker type. NF-κB signaling terms are specific to DNA, MHC protein complex terms are specific for RNA, while lipoprotein terms are specific to protein biomarkers. Similarly, protein subsets also revealed several enriched GO terms that were not previously directly associated with anti-TNF therapy response in rheumatoid arthritis. Gene ontology analysis revealed several enriched blood lipoprotein (HDL, VLDL and cholesterol) terms, illustrated in Figure 3.   Similarly, protein subsets also revealed several enriched GO terms that were not previously directly associated with anti-TNF therapy response in rheumatoid arthritis. Gene ontology analysis revealed several enriched blood lipoprotein (HDL, VLDL and cholesterol) terms, illustrated in Figure 3.

Discussion
The results of our study help to elucidate the mechanisms underlying response and non-response to anti-TNF therapy in rheumatoid arthritis. Biological markers linked to mechanisms associated with response and/or non-response to anti-TNF therapy have potential clinical applications as response predictors before or during anti-TNF therapy or even as potential novel therapeutic targets.
First, there was significant enrichment of protein metabolism terms in gene network subsets based on RNA biomarkers (specifically, RNA_UP_R_DO_N_BIO). The leading GO term was the hypernym positive regulation of protein metabolic process (p = 3.63 × 10 −37 ). Specifically, several enriched hyponyms under this leading term are associated with the proteasome, such as proteasome-mediated ubiquitin-dependent protein catabolic process (p = 2.91 × 10 −15 ). To our best knowledge, proteasome processes have not yet been implicated in anti-TNF therapy response in rheumatoid arthritis. In RA, the autophagy and proteasome protein degradation pathways are key processes for synovial fibroblast survival [141]. In response to TNFα, the autophagy pathway, but not the proteasome, is consistently stimulated, yet there is an increased dependence on the proteasome for cell viability [141]. If autophagy is blocked in the presence of TNFα, an increase in proteasome activity occurs in some RA synovial fibroblasts but decreases in healthy synovial fibroblasts [141]. Targeting the proteasome complex thus represents a therapeutic opportunity to decrease synovial fibroblast survival, pannus growth and inflammation in RA [142][143][144]. Bortezomib, a proteasome inhibitor indicated for hematological cancers, was shown to decrease bone loss in an animal model of RA [145] and inflammatory cytokine production in an ex vivo study of activated T cells of healthy controls and RA patients [146]. In a recent study, delanzomib, a novel proteasome inhibitor, was successfully used together with adalimumab in a rat model of rheumatoid arthritis [147]. Moreover, two case reports showed remission of rheumatoid arthritis complicated with multiple myeloma [148] or TEMPI syndrome [149] after administration of bortezomib.
Second, several terms related to lipoproteins were found to be significantly enriched in protein biomarker subsets. In the subset containing all protein biomarkers, the leading lipoprotein terms were lipoprotein particle receptor binding (p = 8.81 × 10 −12 ) and plasma lipoprotein particle (p = 4.55 × 10 −11 ). Interestingly, the hyponyms very-low-density lipoprotein particle (p = 1.83 × 10 −10 ) and spherical high-density lipoprotein particle (p = 5.22 × 10 −8 ) suggest the role of very-low-density lipoproteins (VLDLs) and high-density lipoproteins (HDLs) in response. Comparative GO analysis showed VLDL to be specific for protein biomarkers down-regulated in responders (or up-regulated in non-responders), and HDL was shown to be up-regulated in responders (or down-regulated in non-responders). These findings confirm clinical observations of increased HDL [150,151] as well as triglyceride and total cholesterol levels [152] after anti-TNF therapy initiation. Moreover, low baseline VLDL has been linked with a better response to anti-TNF therapy [153], which coincides with our finding of VLDLs being down-regulated in responders. Although blood lipid profiles may only reflect systemic inflammation and thus also disease severity, their role in anti-TNF therapy response is not yet understood. Blood lipid profiles are potential accessible and affordable anti-TNF response biomarkers that could be integrated into clinical routine.
Third, our results show a significant enrichment of GO terms related to leukocyte chemotaxis in RNA subsets, with the leading term being negative regulation of leukocyte chemotaxis (p = 3.26 × 10 −4 ). Hyponym investigation in a comparative analysis of RNA biomarkers up-regulated and down-regulated in responders showed the term negative regulation of macrophage chemotaxis (p = 3.00 × 10 −3 ) to be up-regulated in responders (or down-regulated in non-responders). This finding suggests that good responders have lower macrophage infiltration than non-responders. Macrophage chemotaxis thus represents both an opportunity for response biomarker discovery as well as a therapeutic target. An example of a leukocyte chemotaxis reducing drug is montelukast, a cysteinyl leukotriene receptor antagonist used to treat asthma and allergic rhinitis. Although montelukast is mainly used to block leukotriene-dependent human airway smooth muscle contractions, it also blocks up-regulation of vascular permeability and leukocyte chemotaxis. A study has shown that montelukast decreases inflammatory cytokine production in RA and thus represents a novel therapeutic strategy [154].
Finally, our review of anti-TNF therapy response biomarkers has revealed that many response biomarkers have been reported at several levels of biological data (DNA, RNA, proteins, etc.), but only 12 biomarkers were reported by more than one study. Biomarkers reported by more than one study include the DNA biomarkers CCL4 and IL1B; the RNA biomarkers FCGR2A, FCGR3A, IL10, IL6, PTPRC and TNF; and the protein biomarkers IL6, ITIH1, S100A8 and S100A9. Recently, a Japanese cohort has demonstrated the use of interferon signatures and their dynamics for use in long-term anti-TNF drug response prediction, which validates previously reported biomarkers related to interferon proteins [155]. Interestingly, results from another recent study showed that interferon-related chemokine levels (e.g., CXCL10) correlated with disease activity but not with short-term response to anti-TNF therapy (certolizumab pegol) in a Swedish cohort [156]. These studies highlight the difficulties of biomarker replication, especially with cohorts from different ethnic backgrounds and with different study designs.
Our GO analysis of anti-TNF therapy response biomarkers highlighted several biological processes as significantly enriched in response and/or non-response to anti-TNF therapy. Our results encourage targeted analysis of these biological processes for novel biomarker discovery but also the development of novel therapeutic strategies in the treatment of RA. The highlighted therapeutic targets could be useful either as alternatives for anti-TNF therapy non-responders, as co-therapies with anti-TNF treatment or as novel maintenance strategies. Moreover, our study's review of anti-TNF response biomarkers revealed that although response biomarkers have been extensively studied, there is a generally low rate of overlap and biomarker validation between studies.

Conclusions
Biological processes related to the proteasome and blood lipids could affect response to anti-TNF therapy according to gene ontology of existing anti-TNF therapy response biomarkers in RA. Our study encourages further investigation of proteasome and blood lipid processes in RA anti-TNF response.