A novel gene-expression-signature-based model for prediction of response to Tripterysium glycosides tablet for rheumatoid arthritis patients

Tripterygium wilfordii Hook F (TwHF), a traditional Chinese medicine, has been used in the treatment of rheumatoid arthritis (RA) for hundreds of years in China [1]. TwHF has been considered as a potential source for developing new drugs to treat RA, due to its considerable improvement in the outlook for patients suffering from RA [2]. Tripterysium glycosides (TG) tablets, as the main effective ingredients of TwHF, are the most commonly used TwHF-based therapy and display better therapeutic effects than several first-line disease-modifying anti-rheumatic drugs according to recent clinical observations [3]. However, approximately 30% of RA patients treated with TG tablets fail to achieve clinical improvement, depicting significant inter-individual variations caused by epigenetic, physiologic, environmental and especially genetic factors in combination or alone [4, 5]. The heterogeneity of RA patients in pathological manifestations, disease progression and treatment response implies the existence of several disease subtypes at the molecular level [6]. Researchers are trying to figure out various biomarkers of therapeutic response in rheumatic diseases for further benefiting individualized treatment and curative outcomes, including microRNAs, genes, proteins and etc. [7]. However, existent biomarkers that reflect bone and cartilage turnover are insufficient surrogates for predicting therapeutic outcomes of certain drugs [8]. Thus, it is of great clinical significance to identify biomarkers and to develop tools, which are predictive of RA patients’ response to TG tablets.

In recent years, high-capacity genomics and transcriptomics technologies, such as gene microarrays that detect expression profiles of numerous genes simultaneously and comprehensively, play important roles in identifying biomarkers for disease behavior and therapeutic response prediction [9, 10]. Although gene expression microarray is characterized with high sensitivity and high-throughput, it is insufficient to comprehensively clarify the entire biological regulatory processes in RA. Moreover, variations in sample size and quality may often result in the inconsistencies among numerous differentially expressed genes identified by microarrays [11]. To address these problems, we herein designed a hypothesis-generating study combining the high-throughput merits of gene expression profiling and the comprehensive illustration of drug-disease interactions by molecular network analysis, to assess the differentially expressed genes between responsive and non-responsive RA patients to TG tablets and to identify the candidate gene biomarkers according to both the differential expression patterns and the network topological features, as well as to construct a partial-least-squares (PLS) model based on the expression levels of the candidate gene biomarkers in peripheral blood for stratification and prediction of RA patients’ response to TG tablets (Fig. 1).

Since the clinical observations have indicated that not all RA patients benefit to the same extent from the treatment of TG tablets, the identification of drug response biomarkers and the development of predictive tools based on the biomarkers may be of great significance to determine patients with a low probability of response to TG tablets, allowing clinicians to choose alternative drugs at an earlier stage of the disease without any delay of efficacious treatment. In the current study, we integrated the drug response-related gene expression profile and gene–gene interactions to identify the six candidate gene biomarkers that are predictive of response to TG tablets. Then, we applied the PLS algorithm to construct the prediction model for the treatment outcome based on the expression levels of these candidate gene biomarkers. Both 5-fold cross-validation and the independent dataset validations showed the high predictive accuracy (87.50–100.00%) and area under ROC curve (0.875–1.000) of this model. More importantly, our data demonstrated a distinguished improvement of the PLS-model based on the expression levels of the six candidate gene biomarkers in combination over the commonly used clinical and inflammatory parameters, as well as the gene biomarkers alone, in predicting RA patients’ response to TG tablets. To the best of our knowledge, this is the first study that identified the gene biomarkers of RA patients’ response to TG tablets and also evaluated their utility in the prediction model of the clinical treatment outcome.

To facilitate the implementation of personalized medicine and the improvement of therapeutic outcomes of patients with RA, recent studies have been trying to identify various predictors of response to therapy, including single nucleotide polymorphisms, genes, proteins and microRNAs, based on genomics, transcriptomics, proteomics and miRNomics, respectively [2, 24‐26]. However, the studies only based on the “omics” data analysis may fail to fully explain all of the alterations occurring during the RA progression. The need for a deeper understanding of how molecular regulatory networks function in response to therapy may lead to increased efforts to model multiple “omics” dimensions simultaneously. In the context, we here performed a systematic integration of the differential expression analysis on microarray data and topological features of gene signal transduction network, which offers us two main advantages: first, it enables us to sufficiently utilize the gene co-expression information provided by the microarray data, which is believed to be more informative than expression changes of individual genes for target gene identification. Second, network analysis is a powerful tool to understand disease occurrence and progression, as well as therapies and drug responses. By integrating the topological features of biological network, some information lost in the differential expression analysis may be added to the identification of the candidate gene biomarkers predictive of response to TG tablets. Functionally, as shown in Table 3, these candidate gene biomarkers were associated with RA pathogenesis and drug metabolism. Especially, MX1 and OASL both function as interferon (IFN-I)-inducible genes and are predominantly involved into signal pathways in the immune system. Recent studies have indicated that the IFN-I signature in RA may display clinical relevance in relation to disease onset and therapeutic response [27, 28]. Sanayama et al. [29] identified MX2 (a member in the same family with MX1), and OASL as biomarkers for predicting the therapeutic response to tocilizumab in patients with RA. RNF2 (the RING domain-containing E3 ubiquitin-protein ligases RING finger protein 2) is involved in the maintenance of histone H2A levels and impacts transcriptional activity [30]. RING E3 ligases have been revealed to be involved into the control of multiple cellular processes and also regarded as candidate therapeutic target of many human diseases, including RA [31]. The ubiquitin/proteasome protein degradation pathways were found to offer a contribution to prolonging the survival of synovial fibroblasts in RA tissue [32]. Interestingly, Torre et al. [33] found that a E3 ubiquitin ligase might positively regulate type I interferon responses and promote pathogenesis during neuroinflammation, implying several possible associations of RNF2 with MX1 and OASL. Crk family adaptors have been reported to be widely expressed and mediate the timely formation of signal transduction protein complexes upon a variety of extracellular stimuli. Hisakawa et al. [34] found that tyrosine phosphorylation of Crk-associated substrate lymphocyte-type (Cas-L) was markedly enhanced in synovial fluid T cells from patients with RA. Miyake-Nishijima et al. [35] demonstrated the overexpression of Cas-L protein and the increase of its tyrosine phosphorylation in RA mouse model, as well as observed a large number of Cas-L-positive lymphocytes infiltrating to the inflammatory lesions of RA patients, implying the important roles of Cas-L in the pathophysiology of RA. The above literature reports support the evidence that the candidate gene biomarkers identified in the current study may be associated with disease progression and treatment outcome of RA.

To determine the clinical utility of the candidate gene biomarkers, we established a PLS-based prediction model for the treatment of TG tablets using the expression levels of the six candidate gene biomarkers in patients with RA. PLS is a method for modeling a relationship between two sets of multivariate data via a latent space, and of performing least squares regression in that space. It maps input and output variables to low-dimensional spaces so that the covariance of data in the latent spaces is maximized. The algorithm can efficiently differentiate the two datasets by extracting effective information from a large number of features [36‐38]. Notably, both the cross-validation and the independent clinical cohort validation suggested that this model may be capable to predict the therapeutic effectiveness in RA patients treated with TG tablets, and also confirmed the essentiality of the model construction by comparing the performance of commonly used clinical and inflammatory factors, each gene biomarker alone and the model with gene combination.

This hypothesis-generating study identified MX1, OASL, SPINK1, CRK, GRAPL and RNF2 as novel targets for RA therapeutic intervention, and the PLS model based on their expression levels may have the potential prognostic value in RA patients who are treated with TG tablets. Our PLS model may not only warrant evaluation in prospective studies, but also potentially benefits individualized therapy of RA in a daily clinical setting. However, the clinical cohort using in this study is the relatively small for the generation and validation of the predictive model, which may lead to some model over-fitting. Therefore, this is a hypothesis-generating study and future studies based on large clinical cohorts to verify the utility of the six gene biomarkers and the PLS model in predicting and monitoring TwHF-based treatment outcome are required.