Background
Lupus nephritis (LN) is a severe complication of systemic lupus erythematosus (SLE). LN is a form of glomerulonephritis and is typically classified into six distinct histological classes depending on the manifestation and severity of renal involvement [
1]. Even though there have been accumulating knowledge and effective therapeutic options in recent decades, LN remains a clinical challenge [
2‐
4]. With current treatments such as glucocorticoids, cyclophosphamide (CYC), or mycophenolate mofetil (MMF), less than 50% of patients achieve a complete clinical response after 1 year [
5]. Even with clinical remission, 44.4% of patients show residual histological activity, and 27–66% develop renal flare (RF) [
6,
7]. Once RF occurs, the risk of progressive kidney disease is dramatically increased, leading to poor outcomes and greater economic burden [
8‐
10]. Thus, the prevention of RF with appropriate maintenance immunosuppressive therapy is vital and may decrease long-term morbidity and mortality [
7]. However, 5–30% of patients develop end-stage kidney disease (ESKD) within 10 years [
5,
11]. This discrepancy in therapeutic response among patients with LN indicates that the effect of conventional immunosuppressive drugs is not uniform, and the identification of candidates responsive to therapies is necessary for personalized medicine.
As histological changes are limited in LN, depending on the histological classification for therapeutic assessment is unreliable. In fact, the response to therapy might be potentially affected by intrarenal molecular mechanisms that drive disease through specific pathogenic pathways [
12]. Currently, biomarkers for predicting treatment response in LN are accumulating but are mainly focused on serum and urinary analysis [
13‐
15]. According to a recent systematic review, there was vast heterogeneity across studies, limiting their use in clinical settings [
16]. Further, as the exploration of renal tissue is the gold standard for LN diagnosis, obtaining genomic information to identify the contributing disease pathways may provide the best value in predicting treatment response.
To the best of our knowledge, few studies have assessed large-scale transcriptomic profiles of LN under varying conditions. Mejia-Vilet et al. extracted RNA from kidney biopsies and found that intrarenal immune gene expression differed between LN at diagnosis and at flare [
17]. Recently, Parikh et al. adopted serial renal biopsies and conducted extensive transcriptomic analyses to dissect the immune pathways responsible for determining drug responses after RF, providing insights into this clinical scenario [
12]. In fact, their findings could be extended if the biological pathways responsible for drug response and genes involved in LN were connected. This approach of incorporating knowledge relevant to disease could help eliminate false-positive markers and enhance the signal-to-noise ratio in large-scale omics data [
18,
19].
In this study, we aimed to construct prediction models for treatment response in patients with LN after the first RF. We attempted to identify disease-defining genes (DDGs) for LN and incorporated these genes into the feature selection process for model establishment.
Discussion
We designed a pilot study to evaluate the feasibility of using subsets of nCounter immunology genes to predict treatment response in LN after the first RF. The subsets were defined according to the clustering results, linking the regulatory activity of the hub genes to the immunology gene panel. Overall, through extensive bioinformatics analyses, we identified 45 hub genes that potentially discriminated LN samples from healthy controls and other chronic renal diseases. After considering the systemic nature of SLE, the ssGSEA scores of DDGs calculated in both tissue and blood samples divided the immunology gene panel into eight clusters. With this biologically pre-processed information, a subset of immunology genes was found to be of higher predictive importance and was incorporated into model training. The model performance was high and was validated in an independent dataset. This study demonstrated that our machine learning (ML) model was interpretable and could potentially be used in the clinical setting if more validation data were used.
The initial analysis involved two tissue compartments (glomerulus and tubulointerstitium) in GSE32591. Although they had different DEGs profiles, the two compartments shared a large number of genes involved in interferon signaling. The interferon signaling, especially the type 1 interferon pathway, was well established in the pathogenesis of SLE and LN [
35,
36]. This shared pathogenesis was confirmed by Parikh et al., who identified this common dysregulation between the two compartments using renal biopsy from patients with LN [
12]. The top hub genes selected via the five Cytohubba algorithms were also supported by current evidence for LN. For example,
STAT1 is involved in the JAK/STAT signaling pathway in response to interferons [
37]. Its activation up-regulates
IFI16, triggering a positive feedback loop that promotes
APOL1’s expression [
38].
APOL1 overexpression is toxic to podocytes and increases the risk of ESKD in LN. In addition, it has been shown that the protein levels of another interferon-inducible gene,
MX1, are significantly higher in both the peripheral blood and renal tissues of patients with LN before immunosuppressive treatment, confirming our findings [
39]. Down-regulated genes, including
EGR1 and
DUSP1, were also supported by independent human renal biopsies in LN [
20]. Additionally, a permutation test for the ssGSEA scores of the DDGs demonstrated their significance compared with randomly chosen gene sets. Taken together, the DDGs identified in this study represent a valid gene signature.
Based on the 527 immunology genes, evaluation of enriched biological and molecular functions could be redundant, as this panel covers specific genes to address general immune-related gene families, such as the major cytokines, chemokines, and their receptors. Moreover, it does not include all the possible genes for a phenotype to be characterized. For example, infiltration of exhausted CD8 T cells, which is related to prolonged remission in SLE after treatment, was defined by the expression of
LAG3,
CD244, and
EOMES [
40,
41]. However, this panel includes only
EOMES, making it difficult to define the presence of this T cell subgroup. Using our approach, however, several enriched top pathways were identified; and according to literature, the enriched pathways for DDC-2 and DDC-6, TNF signaling pathway, and Th1 and Th2 cell differentiation were most relevant to LN [
33,
34]. The benefit of biological knowledge was successfully translated into model performance using LASSO-DDC-6. This indicates that important features (i.e., genes) for determining responders were captured.
ML algorithms are excellent for making successful predictions based on learning the input/output data. In many cases, predictions are accurate even though there is no prior knowledge that directly reflects the underlying physical interactions [
42]. However, the interpretation of the models is trivial if the model becomes complex. This is especially true for models such as neural networks, in which interpreting the hidden nodes is challenging, and each node corresponds to a complex nonlinear function of the input data [
43]. Furthermore, it is difficult to select the optimal gene sets responsible for classification purposes for bioinformatics analyses that often include DNA microarray datasets. This is due to the small sample size compared to a large number of genes. By accounting for irrelevant and noisy genes, the risk of overfitting increases, which may reduce the generalization of the prediction model [
44,
45]. To avoid overfitting, Xiong et al. used several microarray datasets and divided the initial gene pool into clusters based on their structure, followed by LASSO and binary particle swarm optimization [
46]. Using this double-filter approach, they could select optimal gene subsets with higher interpretability. For multi-omics data, Xu et al. found LASSO outperformed support vector machine and random forest algorithms [
47]. To provide a cost-effective approach for breast cancer detection and patient stratification, they modified the ‘dfmax’ parameter of the
glmnet function, limiting the maximum number of features in the LASSO model. Instead of incorporating DNA methylation profile and copy number data, they also found using transcriptomic data alone leads to sparse and accurate signatures. In our study, as there were only approximately 500 genes in the immunology panel to determine the responders, we adopted a different approach by projecting the clustering results within the microarray discovery datasets onto the immunology gene panel. The extent of co-regulation of a gene set was quantified by calculating the ssGSEA score. We then linked the LN-specific regulatory activity with the immunology gene expression in the same individual by Spearman’s correlation analysis to estimate the connection their connection. This estimation was indirect, as we used information from microarray data on the nCounter platform. This inherent limitation might be resolved when more features or complete transcriptomic profiles are available. In this way, we could apply novel mix-LASSO model to predict drug response by identifying a smaller number of tissue-specific features, while maintaining the model interpretability and stability for various purposes [
48]. Nevertheless, our study suggests that the approach of incorporating knowledge from one platform into another is feasible. However, we have to ensure that platforms being evaluated are sufficiently comparable. Despite the poor correlation between lowly and highly expressed genes in microarray and control-gene-normalized nCounter measurements, the relative expression levels were preserved for most genes [
49]. Therefore, we believe that the cross-platform estimation is feasible. Moreover, clustering with the K-means method was applied to the optimal group immunology genes, generating separate gene sets that could be further filtered, reducing the final dimensionality for model construction. In this study, each DDC comprised a maximum of 98 genes (DDC-4) and a minimum of 19 genes (DDC-1), improving the original variable-to-sample ratio. This approach could also be used in samples with complete transcriptomic information to derive interpretable gene sets. Feature scores can then be obtained from gene sets by building simple linear models or feeding them into neural networks. This model transparency could help explain the link between genotypes and phenotypes or assist in discovering novel biomarkers [
50,
51]. In the present study, the gene set selected in DDC-6 by LASSO could be a potential gene signature and predictive biomarker.
We also identified that
LCK was differentially expressed and able to regulate other top-ranked genes in the sub-networks.
LCK is an Src kinase lymphocyte-specific protein tyrosine kinase that phosphorylates the immunoreceptor tyrosine-based activation motif of CD3ζ after T-cell receptor, which in turn recruits ZAP-70 and causes calcium influx in T cells [
52]. In addition to its vital roles in the development, function, and differentiation of T cells,
LCK is involved in many cellular diseases, such as cell cycle control, proliferation, and differentiation [
53]. The activation of T cells, including CD8 T cells, CD4 T follicular helper cells, and subsets of Th17 cells, has been recognized as a key contributor to the pathogenesis of SLE and LN [
54]. Ko et al. examined kidney tissues in LN and found increased immunohistochemical staining for CD4 + , CD8 + , and CD68 + in the renal periglomerular area [
21]. In our study,
LCK was up-regulated in non-responders and indirectly correlated with the infiltration of various T cell subtypes. Based on its therapeutic significance in various inflammatory diseases, we hypothesized that
LCK could be a therapeutic target for LN at RF.
Our study had several limitations. First, the estimation of DDCs was based on different platforms, which could reduce the generalizability of the clustering results, as some genes might not have the same expression pattern or regulatory structure detected by different techniques, which should be carefully addressed by comparing the DDG profiles on the microarray and nCounter platforms. Second, K-means clustering is sensitive to initial conditions. Even though we have evaluated the clustering results for many times to identify the optimal clustering results, there was still slight difference for each random start. However, one of the advantages of K-means clustering is its efficiency and the ability of handling larger datasets. In the future work, bootstrap sampling may be helpful to improve its problem. Third, only one ML algorithm was used in this pilot study. Other models suitable for transparency and interpretability will be of interest in the future. Nevertheless, the performance of our model was comparable to that of a recent ML-based prediction model [
15]. In their study, they trained a variety of ML algorithms with 246 subjects to develop prediction models for 1-year proteinuria and estimated glomerular filtration rate (eGFR) in LN. They found the combined model with traditional clinical data and novel urine biomarkers for eGFR had the best performance in training and validation datasets and the AUC was near 0.7, which was slightly lower than our model. Fourth, the validation cohort had a small sample size, and replicates of patients were present due to inclusion of different tissue compartments. Finally, we did not investigate the gene expression for cell junctions as they are important factors related to proteinuria, which is a clinical determination of RF [
55]. Further experiment and the inclusion of larger validation samples are required.
Treatment of LN has been faced with many challenges. One of the challenges is the poor target tissue distribution for immunosuppressive drugs [
56]. However, with the emergence of nanotechnology, potential nanomaterials with special physiochemical properties are being developed and applied for treating various diseases including glomerulonephritis [
57]. With better penetration of loaded drugs, clinical outcomes could be more likely associated with their therapeutic effects in the absence of tissue barrier, which may lead to better prediction model performance due to stronger connection between drugs and outcomes.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.