Utilizing rat kidney gene co-expression networks to enhance safety assessment biomarker identification and human translation

Summary

Toxicogenomic data provide key insights into molecular mechanisms underlying drug-induced organ toxicities. To simplify transcriptomic data interpretation, we applied weighted gene co-expression network analysis (WGCNA) to rat kidney transcriptomics data from TG-GATEs (TG) and DrugMatrix (DM), covering time- and dose-response data for 180 compounds. A total of 347 gene modules were incorporated into the rat kidney TXG-MAPr web-tool, that interactively visualizes and quantifies module activity using eigengene scores (EGSs). Several modules annotated for cellular stress, injury, and inflammation were associated with renal pathologies and included established and candidate biomarker genes. Many rat kidney modules were preserved across transcriptome datasets, suggesting potential applicability to other kidney injury contexts. Cross-species preservation analysis using human kidney data further supported the translational potential of these rat-derived modules. The TXG-MAPr platform facilitates upload and analysis of gene expression data in the context of rat kidney co-expression networks, which could identify mechanisms and safety liabilities of chemical or drug exposures.

Graphical abstract

Highlights

• •Gene network analysis cluster biologically related genes for mechanistic interpretation
• •Module association with pathology can identify potential biomarker genes
• •Injury and inflammation responses are preserved between rat and human
• •Rat kidney TXG-MAPr tool can accelerate chemical hazard identification

Biocomputational method; Data processing in systems biology; Transcriptomics; Model organism

Introduction

Kidneys play a pivotal role in drug disposition and excretion, making them an important target organ for adverse reactions. Chemical- or drug-induced kidney injury (DIKI) is often observed in (proximal) tubular epithelial cells, but other nephron segments such as the glomerulus may also be targeted.¹ Initial injury can lead to subsequent cell death followed by repair and regeneration or, if injury is severe, to loss of nephrons and progression to renal failure.² The severity of the initial insult, the degree of cellular injury, and the potential for repair in these regions are determined by local extracellular drug concentration (blood and glomerular filtrate) and intracellular accumulation of drug driven by cellular transport/uptake. The balance between uptake and efflux transport, as well as drug metabolism dictate the effective exposure at cellular targets that yield on- or off-target pharmacology and toxicity.³ Activation of cellular stress response pathways, including oxidative stress, DNA-damage, and unfolded protein response, are early responses to injury and are critical in the adaptation versus progression of kidney injury, although the mechanistic understanding is still incomplete.^4,5

Nonclinical safety studies are designed to characterize target organ toxicity that can be monitored during clinical testing of drug candidates. Histopathology data are the gold standard for pre-clinical drug safety assessment but only informs the occurrence of injury and not the mechanism.⁶ Association of drug-induced transcriptomic changes with histopathology findings may provide key mechanistic insights into biological responses leading to adversities. However, despite two decades of application, transcriptomic analysis has not achieved wide-spread application in drug safety assessment, perhaps in part due to difficulties in achieving meaningful qualitative and quantitative interpretation and useful safety predictions from high content transcriptomic data.^7,8 Among the reasons for slow uptake is the difficulty in translating gene-level data into interpretable biological responses and the penalties incurred due to false discovery rates.^8,9 However, recent publications demonstrate that co-regulated gene network approaches can organize high dimensional toxicogenomic data into a smaller set of biological response networks, which has been applied to uncover mechanisms underlying drug-induced liver toxicities.^10,11,12,13

In this study, we aimed to develop co-regulated gene networks to study the mechanism of kidney injury using toxicogenomic data. We applied weighted gene co-expression network analysis (WGCNA) to identify rat kidney specific co-regulated gene networks (modules) using the publicly available TG-GATEs (TG) and DrugMatrix (DM) toxicogenomic dataset from rat kidney.^14,15,16,17 As done previously for primary human hepatocytes and rat liver,^10,11 we have developed an R-shiny-based interactive webtool, the rat kidney TXG-MAPr (https://txg-mapr.eu/login/). WGCNA gene modules could be quantitatively visualized by module eigengene scores (EGSs) and reduced to meaningful biological responses relevant to understanding mechanisms of toxicity. Specific module responses were statistically associated with pathology phenotypes (clinical chemistry and histopathology), thereby linking cellular mechanisms to renal injury. Several of these modules linked to kidney injury are preserved across datasets and species, including kidney transcriptomic data from human patients.¹⁸ Our results suggest that gene co-regulated modules in the rat kidney TXG-MAPr can be applied to other toxicogenomic datasets for chemical and drug safety assessment and to identify modules and associated biological responses that are relevant in the context of human translational safety assessments.

Results

Rat kidney TXG-MAPr tool to identify biological mechanisms of kidney injury

Gene co-expression networks (347 modules) were built from the TG and DM rat kidney datasets using WGCNA (Table S1). Gene expression changes were scored at the module level for each treatment condition by the EGS (see Table S2 for all treatment conditions). The EGS indicates the induction or repression of the entire module based on the weighted log2 fold change (log2 FC) of its component genes. Modules are arranged in a TXG-MAPr dendrogram, based on the hierarchical clustering across all treatments (Figure S1), and the size and color of each circle is proportional to the amount of induction or repression for a given treatment condition (Figure 1A, red to blue color scale, respectively). The rat kidney module and gene expression data were incorporated into an R-shiny framework (TXG-MAPr: https://txg-mapr.eu/login/) for interactive visualization and analysis, as previously described.¹⁰ Functional annotation of the modules was assessed using overrepresentation analysis (gene ontology [GO] term, pathway, and transcription factor [TF]) to provide cellular and mechanistic context of gene members in each module (Tables S3, S4, and S5). Not all modules showed strong enrichment for GO and pathway terms, but others were highly enriched for terms relevant to kidney injury and regeneration, including rKID:3m (metabolism, e−45), rKID:5m (cell cycle, e−67), rKID:6m (ribosome biogenesis, e−42), rKID:7m (immune system, e−47), and rKID:10m (extracellular matrix, e−34). From this functional enrichment analysis, annotation levels 1 through 3 were derived and summarized in a specific key annotation term for 170 modules (Table S5), i.e., when annotation terms were highly significant (p < 0.0001) and show a similar theme or could be linked to the enriched TF, which was included in further analyses (see following text).

Figure 1.

Overview of the rat kidney TXG-MAPr tool

(A) TXG-MAPr dendrograms showing WGCNA module eigengene scores (EGSs) of daily repeated 1 mg/kg/day cisplatin time course exposure from 4 to 29 days. The size of the circles is proportional to the module EGS, and the red/orange colors indicate induction of the module, while blue/green indicates repression of the module.

(B) Modules rKID:160, rKID:298, and rKID:2m (left) are strongly induced by high dose cisplatin exposure at 29 days, which is further displayed in cisplatin time response plots of the module EGS (middle) and the log2 FC (right) of the most significant genes, including the most hub-like gene (black).

(C) KEGG pathway and transcription factor (TF) enrichment (p value) for p53 signaling or TP53 projected on the TXG-MAP dendrogram, with rKID:160 and rKID:298 highlighted.

(D) Module EGS correlation plot of rKID:298 and rKID:160 for all treatment conditions (left) and correlation between 12 mg/kg/day puromycin aminonucleoside and 1 mg/kg/day cisplatin at 15 days for all modules (right).

To illustrate the utility of the rat kidney TXG-MAPr tool as an approach to extract mechanistic information from gene expression data, we selected compounds in the TG dataset for which various pathologies were noted in the single and repeated dose studies (Table S6). Repeated daily dosing for up to 29 days with the genotoxic chemotherapeutic drug cisplatin (CSP) caused progressive perturbation of gene expression in rat kidneys, represented by the module induction or repression visualized on the TXG-MAPr dendrograms (Figures 1A and S2). Histology findings were also progressive over time, including tubular necrosis, dilatation, and regeneration, and were accompanied by increases in blood urea nitrogen (BUN) and serum creatinine (CRE) (Table S6). We looked for activation of modules associated with the DNA-damage response, a primary mechanism of CSP inducing kidney injury.¹⁹ Rat kidney module 160 (i.e., rKID:160) was among the top induced modules after CSP treatment at various time points and dose levels (Table S2) and contained many well-known p53-target genes (Mdm2, Cdkn1a, Phlda3, Ccng1, Plk2, Aen, Zmat3, Fas, Pias3, and Gas6) involved in the DNA-damage response (Figure 1B, top), and had the highest TF enrichment score for Tp53 (p value of 2e-10) among all modules (Figure 1C; Table S4). Induction of module rKID:160 and its component genes was already evident early and increased over time (Figure 1B, top). Another module, rKID:298, also enriched for terms consistent with p53 signaling (Table S3) and contained genes involved in apoptosis (Bax, Ei24, and Thyn1) and DNA repair (Mgmt), showed a similar pattern of induction compared to rKID:160 (Figure 1B, middle). Although both rKID:160 and rKID:298 are enriched for p53 signaling terms, only rKID:160 had strong TF enrichment for Tp53 (Figure 1C; Table S4). Nonetheless, module rKID:160 showed high correlation (Pearson’s r) with rKID:298 across all TG treatments (Figure 1D, left), indicating that both modules are strongly induced by the same treatment conditions (e.g., especially genotoxic compounds CSP and doxorubicin), as expected given their proximity in the dendrogram and involvement in similar processes.

To determine if the modular gene expression responses also reflected the progressive injury noted with CSP treatment, we also investigated module membership for injury processes and kidney specific biomarkers. Module rKID:2m, contained well-known kidney biomarker genes, including kidney injury molecule-1 (Kim-1/Havcr1), and was strongly associated with toxicity (discussed in the next sections). After only modest early induction, the rKID:2m EGS, and log 2FC for component genes, increased markedly after day 4 (Figure 1B, bottom). Other modules that annotated for biological processes important for kidney injury and regeneration also showed dose and time dependent induction with CSP and other nephrotoxicants allopurinol (APL) and puromycin aminonucleoside (PAN) (Figure S3). Several modules involved in extracellular matrix (rKID:10m and 31m), immune response (rKID:56 and 63m), cytoskeleton (rKID:42m and 142m), cell cycle (rKID:5m), and transcription (rKID:29) were induced after 29 days CSP exposure and were also observed with other nephrotoxic compounds, including APL and PAN (Figure S3), both of which caused substantial kidney injury and regeneration (Table S6). Notably, the high module EGS correlation (Pearson’s r), between 15-day repeated treatment with 1 mg/kg/day CSP and 12 mg/kg/day PAN suggested that these responses reflect general kidney responses to injury, while also reflecting distinct mechanisms (Figure 1D, right). For example, the inflammatory (rKID:56, 63m, and 327), cytoskeleton/injury (rKID:2m and 142m), and cell cycle (rKID:5m) responses are comparable between CSP and PAN at day 15, but the p53 responses (rKID:160 and 298) are more pronounced with CSP, because of the genotoxic mechanism of action.

Taken together, the TXG-MAPr tool demonstrated that the co-expression modules capture meaningful biological responses after nephrotoxic insults, where early events, such as activation of p53 by CSP, are followed by later events associated with cellular and tissue injury, including modules containing inducible biomarker genes.

Modules associate biological response networks with kidney pathology

Having demonstrated that modular gene expression captures valuable information on mechanism associated with kidney injury, we wanted to investigate the statistical association of module changes, and the associated biological responses, with kidney injury (histopathology or clinical chemistry; see Tables S6 and S7 for frequency and occurrence for each treatment). Therefore, for each relevant kidney toxicity phenotype (Table S7), we measured effect sizes (Cohen’s D) of module EGS for treatments resulting in the presence (positives) or absence (negatives) of renal injury (Table S8). To control for the association of injury with overall gene expression, we also applied logistic regression to account for differences in module EGS while controlling for average absolute EGS (AvAbsEGS) as covariates in the analysis, as described previously.¹³ Estimation of the odds of toxicity was quantified by the (adjusted) p value from the logistic regression, which was signed (signed log10 p-adjust) for positive (induction) or negative (repression) module responses associated with toxicity. Both concurrent and predictive (i.e., early module changes that precede emergence of toxicity at a later time point) module associations with pathology were calculated, although concurrent associations showed higher statistical significance as expected (Table S8).

The effect sizes of many modules associated with concurrent toxicity phenotypes showed strong Pearson’s r correlation, suggesting that several biological response-networks are associated with multiple toxicity phenotypes (Figure 2A). This is likely an effect of either the high transcriptional activity, measured as AvAbsEGS, concurrent with multiple pathologies, and/or activation of generalized tissue responses to injury. When we repeated the clustering using the signed p-adjust values, which accounts for the AvAbsEGS as a covariate in the logistic regression, there was somewhat better separation of the toxicity phenotypes into two distinct clusters (Figure 2B). Cluster 1 modules were more associated with tubular injury phenotypes (increases in tubular necrosis, degeneration, BUN, and CRE), and cluster 2 with regeneration and inflammation phenotypes (increases in tubular regeneration, dilatation, cellular infiltration, and fibrosis). The overlap and differences in module association could be appreciated when comparing the TXG-MAP dendrogram plots overlaid with the signed log10 p value per module for the most prevalent renal injuries (Figures 2C and S4). Modules in branches (see branches in Figure S1) F1 and F2 (stress responses and cell cycle) were strongly associated with tubular injury phenotypes (red colors). In contrast, modules in branches A2 and B2 (immune response and extracellular matrix/adhesion) were strongly associated with regeneration and inflammation. Cellular stress, transcription, and cytokine-related modules in branch H1 showed strong association with all pathologies (red colors), as well as modules in branch D2a (RNA splicing). Modules in branch I (mitochondrion/respiration) showed negative association with all toxicity phenotypes, as well as some metabolism modules in branch H1 (blue colors). Other (non-tubular) injuries did not show strong association with transcriptional module changes, likely due to the low number of positive treatments in the TG data that cause these injuries or the limited transcriptional perturbation (Figure S4).

Figure 2.

Module association with pathology

(A and B) Strong Pearson’s r correlation between different pathologies based on Cohen’s D effect size (A) and signed log10 adjusted p value (B). This indicates that the same modules are associated with multiple pathologies, which could be the results of co-occurring pathologies (cluster 1 in A). Cluster 1 from (A) could be separated in 2 clusters (B), when controlling for module EGS using signed log10 adjusted p values. Only the most occurring pathologies were taken along, with a pathology score ≥0.67 (and ≥1.33 for some pathologies).

(C) TXG-MAPr dendrograms showing module association with different toxicity phenotypes based on the signed log10 adjusted p value. Red color indicates that the module EGS positively correlates with the selected pathology (module induced with pathology present), while blue means a negative correlation (module repressed with pathology present). The size of the circles is proportional to the −log10 adjusted p value. For the branches with the strongest pathology associations, the general enrichment score is provided based on ORA of all genes in the branch.

(D and E) Heatmap of signed log10 p-adjust of concurrent (D) and 4-day predictive (E) module associations with pathology. Red color indicates that the module EGS positively correlates with the selected pathology, while blue means a negative correlation (i.e., module is repressed when the pathology is present). Hierarchical clustering was applied on rows and columns to cluster both modules (rows) and toxicity phenotypes (columns) with similar association scores together. For concurrent associations, modules from clusters 1 and 2 were selected for strongest negative and positive correlation with toxicity, respectively. Module shows highest statistical association with concurrent toxicity phenotypes (see bottom pathology cluster II), which were selected to look at the mean effect sizes and p values (see also Table S9). For predictive associations, the 4 days show the highest statistical significance, where modules from clusters 2 and 7 were selected for strongest negative and positive correlation with toxicity, respectively. Heatmaps are clustered by Euclidean distance, using the complete method from pheatmap package.

Hierarchical clustering of the modules for concurrent association with toxicity phenotypes revealed a subset of 44 strong positively (cluster 2) and 81 strong negatively (cluster 1) associated modules with injury and repair phenotypes (Figures 2D and S5A; Table S9), including serum biomarkers increases (BUN or serum CRE), several types of tubular injury, regeneration, inflammation, or fibrosis (pathology cluster II in Figures 2D and S5A; Table S7). Notably, module rKID:2m, which contained genes for known and inducible biomarkers of renal injury, is in cluster 2 (Figure 2D) and ranked highest, on average, for concurrent association with the most prevalent toxicities on both effect size and adjusted p value (Table S9). Other positively associated and high-ranked modules (cluster 2, Figures 2D and S5A; Table S9) were annotated for immune response (rKID:56, 63m, 71, 212m, and 327), cytoskeleton (rKID:42m, 112, and 142m), extracellular matrix organization (rKID:10m and 31m), proteasome (rKID:35m), cell cycle (rKID:5m), and transcription (rKID:64m), all of which are processes associated with progressive renal injury and/or repair,^20,21 and were induced by compounds such as CSP, APL, and PAN that caused progressive renal injury (Figure S3). Negatively associated and high-ranked modules (cluster 1, Figures 2D and S5A; Table S9) were annotated for metabolism (rKID:3m, 9m, 14m, 25m, and 292), transport (rKID:19, 171, and 329), AMPK signaling (rKID:383), mitochondrion (rKID:26m, 28m, 96m, and 264), and renal or glomerular development (rKID:196) among others. Interestingly, the latter module rKID:196 showed various genes involved in glomerular development or podocyte function, including Nphs1, Podxl, Ptpro, and Wt1.^22,23

Transcriptional changes at day 4 showed on average the highest statistical power (effect size and log10 adjusted p value) to predict renal injury on a later time point, followed by day 1 (Figure S6). Modules for predictive association with toxicity phenotypes at day 4 revealed a subset of 30 positive (cluster 7) and 46 negative (cluster 2) associations (Figures 2E and S5B; Table S9). Predictive association at day 1 revealed a subset of 44 positive (cluster 3) and 17 negative (cluster 1) associations (Figure S5B; Table S9). Several modules that showed strong positive predictive association with pathology were also strongly associated with concurrent pathology, including cytoskeleton (rKID:2m, 112, and 142m) and immune response (rKID:7m, 56, 63m, 113, 212m, and 327), while mitochondrial (rKID:26m, 28m, and 96m) and metabolism (rKID:3m) showed negative association. Besides the overlap between modules that showed both concurrent and predictive association with pathology, there were also some unique modules in the predictive association selection. These include modules involved in ribosome biogenesis (rKID:6m, 8m, 40m, and 130m), RNA processing (rKID:34m, 278, and 387), protein processing (rKID:173 and 208), immune response (rKID:16m, 138, and 351), transcription (rKID:29), p53 pathway (rKID:160 and 298), Nrf2 pathway (rKID:53), Atf4 pathway (rKID:74), and Atf6 pathway (rKID:81), which are early responses after cellular stress or injury.

Overall, this led to a subset of 169 modules (75 positive and 94 negative) that showed strongly perturbed transcriptional activity and could be correlated with induction of renal injury (Table S9, column “selected strong association”). Thus, module changes reflect biological responses that may be associated with higher likelihood of concurrent or future development of kidney adverse outcomes (i.e., pathologies), while these modules may also contain potential biomarkers.

Pathology associated modules facilitate biomarker discovery

In the previous section, we have described several modules that were most strongly associated with pathogenesis (Table S9), including rKID:2m, which contained well-known kidney injury biomarkers, including kidney injury molecule-1 (Kim-1/Havcr1), lipocalin 2 (Lcn2/Ngal), osteopontin (Spp1), clusterin (Clu), and tissue inhibitor of metalloproteinases (Timp1), all of which had high corEG (i.e., hubness) (Table 1). Interestingly, the primary enrichment terms for rKID:2m included cell adhesion and actin cytoskeletal organization, suggesting a more complex set of biological responses is associated with activation and interpretation of the inducible biomarker genes in rKID:2m. Other known kidney injury biomarkers were in metabolism (rKID:3m), immune response (rKID:7m, 15, 56, 63m, and 281) and extracellular matrix (rKID:10m) modules, and most were strongly associated with renal pathogenesis (Tables 1 and S9).

Table 1.

Kidney injury biomarkers in co-expression modules

Gene	Biomarker name	Injury marker	Module	corEG	References
Havcr1	KIM-1	proximal tubular	rKID:2m	0.8156	Han et al.24; Dieterle et al.25; Wen et al.26; Teo and Endre27; Zhang and Parikh28; Srisawat and Kellum29; Vaidya et al.30
Lcn2	lipocalin2, NGAL	proximal/distal tubular	rKID:2m	0.7698	Wen et al.26; Teo and Endre27; Zhang and Parikh28; Srisawat and Kellum29; Vaidya et al.30
Clu	clusterin	proximal/distal tubular	rKID:2m	0.8282	Dieterle et al.25; Srisawat and Kellum29; Vaidya et al.30
Spp1	osteopontin, OPN	inflammation/distal tubular	rKID:2m	0.7815	Wen et al.26; Vaidya et al.30
Timp1	timp-1	proximal tubular	rKID:2m	0.8948	Sieber et al.31; Huang et al.32
Umod	uromodulin	loop of henle	rKID:3m	0.7673	Wen et al.26; Zhang and Parikh28
Tff3	trefoil factor 3	proximal tubular	rKID:3m	0.5565	Dieterle et al.25
Cst3	cystatin c	glomerular/proximal tubular	rKID:7m	0.3903	Dieterle et al.25; Teo and Endre27; Zhang and Parikh28; Srisawat and Kellum29; Vaidya et al.30
B2m	β2-microglobulin	glomerular/proximal tubular	rKID:7m	0.6849	Dieterle et al.25; Srisawat and Kellum29
Timp2	timp-2	proximal tubular	rKID:10m	0.7746	Wen et al.26; Zhang and Parikh28
Igfbp7	IGFBP-7, IBP-7	proximal tubular	rKID:10m	0.5295	Wen et al.26; Zhang and Parikh28
Ambp	α1-microglobulin	glomerular/proximal tubular	rKID:15	0.7985	Vaidya et al.30
Alb	albumin	glomerular/proximal tubular	rKID:15	0.7157	Dieterle et al.25; Teo and Endre27; Srisawat and Kellum29; Vaidya et al.30
Fabp1	L-FABP	glomerular/proximal tubular	rKID:15	0.8531	Wen et al.26; Zhang and Parikh28; Srisawat and Kellum29
Tnfrsf1a	TNFR1	inflammation	rKID:56	0.9015	Tummalapalli et al.33
Tnfrsf1b	TNFR2	inflammation	rKID:63m	0.6241	Tummalapalli et al.33
Il18	interleukin-18	inflammation	rKID:63m	0.7694	Wen et al.26; Teo and Endre27; Zhang and Parikh28; Srisawat and Kellum29; Vaidya et al.30
Ccl2	MCP-1	inflammation	rKID:281	0.7258	Wen et al.26; Zhang and Parikh28

Well-known and established kidney injury biomarkers are in cytoskeleton/injury (rKID:2m), metabolism (rKID:3m), immune response (rKID:7m, 15, 56, 63m, and 281), and extracellular matrix (rKID:10m) modules.

To discover inducible kidney injury biomarker genes, we investigated modules with strong pathology association and clear biological annotation. The top 10 modules showing the strongest positive association with each of the eight most severe kidney pathologies were selected, which resulted in a set of 29 unique modules (Tables 2 and S9). The selected modules showed very good overlap with modules in cluster 2 of Figures 2D and S5A (concurrent), as well as with cluster 6 and 7 in Figures 2E and S5B (4-day predictive). In addition to the known biomarker genes found in these modules (Table 1), many modules with strong pathology association contain hub(-like) genes of which literature evidence indicates that their respective proteins could be potential kidney injury or disease biomarkers (Table 2). Modules with potential biomarkers are rKID:7m (Cst3 = cystatin C, B2m = β2-microglobulin, and Cd68), rKID:10 (Cdh11 and Smoc2), rKID:31m (Il24), rKID:38m (Tcp1), rKID:56 (Icam1 and Tnfrsf1a), rKID:63m (C1qb, Il18, and Tnfrsf1b), rKID:71 (Mmp7), rKID:138 (Cd3g), rKID:142m (14-3-3 proteins, i.e., Ywhah, Ywhag), rKID:212m (Cp = ceruloplasmin, C1s, C1r, and Dcn), and rKID:327 (fibrinogen, i.e., Fga, Fgb, and Fgg). Many genes in these modules are strongly induced by nephrotoxic treatments, including the potential biomarker genes (Figure 3A). Gene expression responses of these modules mimicked the time dependence of kidney pathology (Figure 3B), suggesting that these modules could be relevant for characterization of previously identified biomarkers and identification of candidate biomarkers of kidney injury. In contrast, traditional serum biomarkers (BUN and CRE) increased much later (Figure 3C), when there is loss of renal function and hence these do not recapitulate the renal injury already present at earlier time points. These results suggest that induction of putative biomarker genes and their modules is strongly associated with renal pathology and capture the progression of kidney injury much better than serum biomarkers BUN or CRE.

Table 2.

Most strongly associated modules with renal pathology containing candidate biomarkers

Module	Module annotation	Hub gene	Other hub genes	Branch	Strongest pathology association	# Toxicity phenotypes	Rank	Max signed log10 p-adjust	References (DOI)
rKID:31m	extracellular matrix (ECM)	Il24	Cthrc1	A2aIIβ	regeneration/basophilic change, tubule ≥ 0.67	3	1	27.68	https://doi.org/10.1007/s00795-019-00239-0
rKID:212m	immune response; complement	Cp	ceruloplasmin (Cp), C1s, C1r, and Dcn	B2aII	regeneration/basophilic change, tubule ≥ 0.67	1	2	23.94	https://doi.org/10.1007/s10157-019-01734-5 https://doi.org/10.1681/ASN.2008111145
rKID:358	–	Fkbp10	–	B2aII	regeneration/basophilic change, tubule ≥ 0.67	2	3	22.49	https://doi.org/10.1159/000448338
rKID:148	–	Sept10	–	D2a	regeneration/basophilic change, tubule ≥ 0.67	1	5	19.54	–
rKID:10m	extracellular matrix (ECM)	Fbn1	Cdh11 and Smoc2	B2aII	regeneration/basophilic change, tubule ≥ 0.67	1	8	18.68	https://doi.org/10.1016/j.kint.2021.04.037
rKID:184	other	Rexo2	–	F2b	necrosis/degeneration, tubule ≥ 0.67	1	2	18.83	–
rKID:142m	cytoskeleton	Ywhah	14-3-3 protein (Ywhah, Ywhag)	H1bI	necrosis/degeneration, tubule ≥ 0.67	2	5	17.09	https://doi.org/10.1111/jcmm.13691
rKID:38m	protein processing; folding	Cct5	Tcp1	F1aI	necrosis/degeneration, tubule ≥ 0.67	2	7	15.80	https://doi.org/10.1093/toxsci/kfz242
rKID:35m	proteasome	Eef1d	–	F2b	necrosis/degeneration, tubule ≥ 0.67	1	10	15.51	–
rKID:112	cytoskeleton	Arpc3	–	F2aI	hyaline cast, tubule ≥ 0.67	1	6	11.71	–
rKID:5m	cell cycle; S/M phase	Mad2l1	Aurkb	F2b	hyaline cast, tubule ≥ 0.67	1	7	11.70	https://doi.org/10.2147/IJGM.S336799
rKID:220	signaling	Htra3	–	A2aIIβ	fibrosis, interstitium ≥ 0.67	1	5	9.33	–
rKID:272	–	Cdk4	–	F2b	fibrosis, interstitium ≥ 0.67	1	10	8.75	–
rKID:2m	cytoskeleton; cell adhesion/injury biomarkers	Tes	Havcr1, Clu, Spp1, Lcn2, Timp1, and Plaur	H1aIIαii	dilatation, tubule ≥ 0.67	4	2	25.13	https://doi.org/10.1097/MNH.0000000000000275
rKID:51m	–	Trpv6	–	A2aIIα	dilatation, tubule ≥ 0.67	3	3	25.06	–
rKID:327	immune response; complement and coagulation	Fgb	fibrinogen (Fga, Fgb, and Fgg)	H1aIIαii	dilatation, tubule ≥ 0.67	4	4	24.50	https://doi.org/10.1016/j.ajpath.2012.06.004
rKID:235	–	Tspan8	–	H2aI	dilatation, Tubule ≥ 0.67	1	5	24.49	–
rKID:144	–	Pla1a	–	H1aIIαii	dilatation, tubule ≥ 0.67	2	6	24.25	–
rKID:332	signaling; calcium	Map3k1	–	H2aII	dilatation, tubule ≥ 0.67	1	7	24.05	–
rKID:56	immune response; TNF/NFkB signaling	Icam1	Tnfrsf1a	H1bII	dilatation, tubule ≥ 0.67	2	9	22.68	https://doi.org/10.1159/000068920 https://doi.org/10.1097/MNH.0000000000000275
rKID:71	immune response	Mmp7	Reg3b and Reg3g	A2aIIβ	dilatation, tubule ≥ 0.67	4	10	22.36	https://doi.org/10.1681/ASN.2016030354
rKID:63m	immune response	C1qb	Il18 and Tnfrsf1b	A2b	cellular infiltration ≥ 0.67	2	1	27.36	https://doi.org/10.1111/petr.13885
rKID:7m	immune response; leukocyte activation	Ptprc	B2m, Cd68, Coro1a, and Cst3	A2b	cellular infiltration ≥ 0.67	1	4	21.23	https://doi.org/10.1111/his.13768
rKID:138	immune response; T cell receptor	Cd3g	Cd3 (Cd3d, Cd3e, and Cd3g)	A2b	cellular infiltration ≥ 0.67	1	5	21.23	https://doi.org/10.1111/cei.12461
rKID:233	–	Tmsb4x	–	B2aII	cellular infiltration ≥ 0.67	2	8	17.86	–
rKID:305	immune response; T cell receptor	Zap70	Cd247 and Klrk1	A2b	cellular infiltration ≥ 0.67	1	9	17.81	–
rKID:59m	–	Akip1	–	B2aII	cellular infiltration ≥ 0.67	1	10	17.76	https://doi.org/10.1097/MD.0000000000020742
rKID:29	transcription	Btg2	Atf3	H1bI	BUN change >100%	1	9	15.51	https://doi.org/10.21037/atm-21-5231
rKID:64m	transcription	Klf6	–	H1bI	BUN change >100%	1	10	15.24	https://doi.org/10.1080/0886022X.2020.1793353

Top 10 ranked modules for positive association with concurrent kidney pathologies were selected for each phenotype. For each module, the hub gene and other highly connected genes are shown with a focus on identification of the corresponding proteins as putative biomarkers of renal injury (see references). The module annotations provide additional mechanistic context for these candidate biomarkers. The strongest pathology association, the rank and the signed log10 p-adjust value is provided for each module. Some of the modules are in the top 10 rank for more than one toxicity phenotype (column: “# Toxicity phenotypes”), including rKID:2m, 31m, 71, and 327.

Figure 3.

Gene expression responses during nephrotoxic treatments

(A) Expression of genes of selected modules with high pathology association is induced during nephrotoxic treatment conditions. Significant gene log2 FC are depicted in blue color (p-adjust < 0.05).

(B) The most prevalent renal pathologies induced by nephrotoxic treatments.

(C) Serum biomarker increases during nephrotoxic treatments versus controls (percentage increase, red line indicates 100% or 2-fold increase). APL, allopurinol; CSA, cyclosporine A; CSP, cisplatin; GMC, gentamicin; LS, lomustine; NPAA, phenylanthranilic acid; PAN, puromycin aminonucleoside; TRI, triamterene.

Module correlation analysis reveals both mechanistic and generalized tissue responses to kidney injury

Results already discussed allude to the fact that kidney injury and progression occur in distinct phases. Early molecular initiating events may be mechanistically distinct for different nephrotoxicants, but lead to a common event, such as death of proximal tubule epithelial cells, which results in a regeneration response. Therefore, we further explored how treatment correlation analysis, using module EGS, might reflect on early and later events in the progression of kidney injury. Besides CSP, other potentially nephrotoxic compounds show strong perturbations of gene expression and pathology at the highest tested dose levels and later time points (Figures 3, S3, S7, and S8). As exemplified earlier, there was a good correlation (R > 0.7) among module EGS after 15 days of CSP and PAN treatment (Figure 1D, right), as well as after 8 and 29 days of treatment (Figure S9A). Both drug treatments also induced comparable pathologies at these later time points, including tubular necrosis, dilatation, and regeneration (Figures 3B and S8B; Table S6). At the 3-h time point there was also a modest correlation (R = 0.55) between CSP and PAN, mainly driven by transcriptional stress (rKID:29 and 64m) responses (Figure S9A). However, from 6 h through 4 days post treatment, there was poor correlation of module EGS between CSP and PAN and little indication of pathology (Figures S8B and S9A). Induction of the p53 response by CSP is apparent at day 1 and later (rKID:160 and 298), while PAN induced modules associated with an early inflammatory response (rKID:15, 16m, 63m, 113, and 351), the anticipated mechanisms to induce glomerular nephropathy.³⁴

Correlation analysis based on module EGS was further leveraged to investigate similarities between nephrotoxic treatment conditions given their overlap in activated and deactivated modules (Figures 4, S8A, and S9). Treatment with nephrotoxic compounds at high dose levels and late time points showed strong correlation and clustering, when there was clearly pathology present in the kidneys (Figure 4A, cluster II, and S9B). This suggests that gene expression patterns are similar when there are co-occurring pathologies, likely an effect of upregulation of injury induced repair mechanisms of the tissue (e.g., regeneration, dedifferentiation, and inflammation responses), and downregulation of metabolism and mitochondrial function, that may be common after acute DIKI.³⁵ However, mechanistic information of compound exposures can still be extracted by focusing on outlier modules (rKID:160 and 298 in Figure 1D) or earlier time points when module activities are more diverse, see distinct compound specific clusters (Figure 4A, cluster I = APL, III = CSP, and IV = PAN). The DM data also showed clustering of stress, immune, injury and regeneration responses with high module EGS during nephrotoxic exposures that clearly showed strong pathology association (Figure S10, clusters 6 and 7, see row annotation). Mitochondrial and metabolism modules that were negatively associated with pathology were indeed downregulated during these exposures with nephrotoxicants in the DM data (Figure S10, clusters 1, 2, and 5). Two estrogen receptor agonists primarily induced steroid hormone metabolism modules among others, showing their primary mechanism of action (Figure S10, cluster 9), while exposure with diethylstilbestrol showed clear activation of stress and injury response modules (clusters 6 and 7), which leaded to mild renal pathology on day 5 (Figure S10B, see column annotation).

Figure 4.

Correlation of ischemic injury and PTC dedifferentiation with nephrotoxic treatment conditions

(A) Cluster correlation plot for transcriptomic responses to nephrotoxic compounds, ischemic injury, and primary PTC cultures, which shows strong Pearson’s r correlation between module EGS at time points when there is concurrent pathology (cluster II in black box), while early allopurinol (cluster I), cisplatin (cluster III), and puromycin aminonucleoside (cluster IV) responses are more distinct. Left heatmap displays the percent increase of serum injury biomarkers (BUN and CRE) compared to control (purple) and histopathology grades of tubule necrosis and regeneration (blue). Heatmap is clustered by Euclidean distance, using the complete method from pheatmap package.

(B) Examples of correlation plots of responses during renal injury, showing high correlation between nephrotoxic treatments (APL, allopurinol), ischemic injury, and PTC dedifferentiation.

To see if observations with nephrotoxicants can be extended to other form of renal injury, we analyzed transcriptomic data from a previous rat renal ischemia reperfusion injury (IRI) study, a drug-independent model of acute kidney injury (AKI).³⁶ Gene expression data were uploaded into the TXG-MAPr environment and module EGS calculated (Tables S11 and S12). The IRI study showed high similarity with early APL treatment, mainly triggered by an early inflammatory response by both treatments (Figures 4A cluster I, 4B1, and S11A). In contrast, the ischemic injury response at days 1 and 5 correlated well with late nephrotoxic TG treatments with co-occurring pathologies, suggesting that this is a general tissue response to injury, since there is no drug involved in vehicle (NaCl) treated ischemic kidneys (Figures 4A cluster II, 4B2, 4B3, and S11A). Valproic acid pre-exposure was able to accelerate the regeneration response at 1 day post ischemic injury compared to NaCl, an effect associated with a stronger activation of cell cycle module rKID:5m (Figure S11B).

Regenerating or injured proximal tubule epithelial cells (PTCs) in vivo and in primary culture undergo dedifferentiation and regenerative responses, including a partial epithelial-to-mesenchymal transition, based on investigation of specific signaling pathways and expression of the mesenchymal intermediate filament, vimentin.^37,38 To determine if the global transcriptional response to dedifferentiation of cultured PTC bore any relationship to the dedifferentiation response associated with injury and regeneration in vivo, we compared rat PTC at various time in culture (i.e., 1–7 days after plating on Transwell filters) with TG treatments (Tables S11 and S12). Notably, there was high similarity between the modular gene expression responses of cultured PTC and late nephrotoxic compound treatments (Figures 4A cluster II, S11C, and S12), as well as to ischemic injury response at day 1 and 5, a time at which regeneration of proximal tubule epithelium is expected (Figures 4A, 4B4, 4B5, and 4B6). Thus, the dedifferentiation response of primary PTC appears similar, at the transcriptional level, to that of the regenerative PTC phenotype associated with the generalized tissue injury/regeneration response in vivo. Notably, activation of modules involved in cell cycle (rKID:5m), cytoskeleton/cell adhesion (rKID:2m and 42m), transcription (rKID:29 and 64m), and RNA and protein processing peaked in the first two days of culture compared to freshly isolated PTC (Figure S12). In addition, there was clear downregulation of metabolism, mitochondrial function, and transport in cultured primary PTC (Figure S12), which was also seen after kidney injury in vivo.

In conclusion, high similarities between cellular and tissue responses during nephrotoxic treatments, ischemic injury, and cultured primary PTC suggest that these are generalizable compound agnostic tissue responses to injury, dedifferentiation, and repair. Responses to nephrotoxic compounds at early time points are more distinct and likely indicate the unique compound specific mode of action that could lead to cellular stress and later renal injury.

Preservation analysis of rat kidney modules reveals common biological processes, stress responses, and kidney-specific networks

One of the most important safety decisions when compounds transition from nonclinical to clinical testing is whether hazards identified in nonclinical studies will translate to human safety concerns. Ideally, quantitative metrics could be used to determine the likelihood that biological events (e.g., molecular initiating events [MIE] or key events [KE] in an adverse outcome pathway (AOP) framework) will translate across species. A unique aspect of network models, such as WGCNA, is that preservation statistics can be applied to determine if (node-edge relationships in) biological networks are preserved across datasets or systems.³⁹ We first tested robustness of network preservation statistics by comparing rat kidney modules based on all data versus the individual TG or DM rat kidney transcriptomic data separately using the Z-summary module preservation statistic.³⁹ Notably, all modules are moderately preserved (274 modules = 79% with Z-summary >2) or highly preserved (73 modules = 21% with Z-summary >10) in the DM data (Figure 5A; Table S10). Self-preservation (testing for preservation in the dataset from which modules were derived) in DM is only slightly lower compared to the self-preservation of the larger TG dataset. In addition, there is good correlation between Z-summary scores for preserved modules in DM or TG self-preservation, including a subset of 67 modules that are highly preserved using both comparisons (Figure 5B, green line = Z-summary >10). This preservation analysis suggests that both TG and DM rat kidney datasets capture similar gene co-expression patterns and biological responses.

Figure 5.

Module preservation

(A) Number of preserved rat kidney modules across datasets (TG-GATEs [TG] rat kidney, DrugMatrix [DM] rat kidney, TG rat liver and human kidney). TG and DM kidney datasets were used as a control to check if all modules are preserved. Z-summary >2 indicate moderate module preservation and Z-summary >10 indicate high module preservation.

(B) Strong correlation between TG and DM preserved modules indicated by Z-summary scores for both comparisons, displayed on the axis as a log scale. Specific modules are highlighted that are strongly preserved in both datasets.

(C) Comparison of rat kidney and liver Z-summary scores displayed on the axis as a log scale. Several modules are strongly preserved in both kidney and liver datasets, indicating that these processed have similar co-expression networks between the organs. Various kidney specific processes are only preserved in the kidney, but not in the liver (left of the red dashed line), including modules rKID:23m (kidney development), steroid metabolism (rKID:18m and 95), and Golgi function (rKID:109).

(D) Comparison of rat kidney and human kidney Z-summary scores displayed on the axis as a log scale. Correlation plot of rat kidney modules preserved in DM kidney versus human kidney data, indicating that several modules are also preserved in human (rKID:2m, 3m, 5m, 7m, 10m, 16m, and 42m among others), but others lack preservation in human (rKID:15 among others). Dashed red line indicates low/moderate module preservation threshold (Zsum > 2) and dashed green line indicates strong module preservation threshold (Zsum > 10). See also Table S10 for more details.

To determine if highly preserved modules reflect common biological processes, we investigated their biological annotations. Highly preserved module annotate for processes important for maintaining cellular homeostasis as well as adaptive responses to chemical injury, including: cell cycle (rKID:5m), metabolism (rKID:3m, 9m, 14m, 21, 22, 25m, and 116), mitochondrial function (rKID:26m, 28m, 37m, 43m, and 96m), immune responses (rKID:7m, 15, 16m, 56, 63m, 113, 138, and 212m), injury biomarker (rKID:2m), cytoskeleton (rKID:42m, 112, and 142m), extracellular matrix (ECM) remodeling (rKID:10m and 31m), transport (rKID:13, 19, 78, and 171), development (rKID:12m, 23m, 45m, 57m, 65, and 70), ribosomal biogenesis/RNA processing (rKID:6m, 8m, 34m, 37m, and 40m), proteasome (rKID:35m), circadian regulation (rKID:33), and responses to oxidative (rKID:53 = Nrf2) or endoplasmic reticulum (ER) stress (rKID:81 = Atf6; rKID:74 = Atf4). Interestingly, there was a large group of 255 kidney modules also preserved in rat liver TG-GATEs dataset (Figures 5A and 5C; Table S10), including the aforementioned modules involved in regulation of cellular stress or homeostasis. In contrast, some kidney-specific modules were not preserved in liver, including modules involved in (renal) development (rKID:23m, 45m, 57m, 70, 186, 196, and 290), steroid hormone metabolism (rKID:18m and 95) and transmembrane transport (rKID:171, 211, and 329) among others (Figure 5C). Thus, preservation analysis illustrates that modules are highly preserved and robust reflections of cellular responses to drug- or chemical-induced renal injury, suggesting that the rat kidney TXG-MAPr analysis framework can be applied to other kidney toxicogenomic data to investigate DIKI mechanisms.

Preserved injury and repair-related networks in human patients

Having established that preservation of rat kidney modules is robust when comparing different large transcriptomics datasets of the same organ and species, we tested preservation between rat and human using available human kidney transcriptomic data. We evaluated preservation of rat kidney modules in a large human dataset (GEO: GSE36059) consisting of over 400 renal transplant samples from human patients at various times after transplantation with or without evidence of graft rejection as well as “control” nephrectomy biopsies.¹⁸ Although the human dataset is limited to renal transplant samples, among the 347 rat kidney modules, 5 modules had strong preservation and 125 had moderate preservation in human kidneys (Figure 5A; Table S10). Modules involved in immune response (rKID:7m and rKID:16m), cell cycle (rKID:5m), and metabolism (rKID:3m) were strongly preserved (Figure 5D), while several other modules representing other immune responses, metabolism, mitochondrial function, RNA processing, cell cycle, and ECM remodeling/fibrosis were moderately preserved (Table S10). These observations are in line with the fact that a subset of the human samples were patients experiencing immune response-mediated graft rejection.^18,40 Strikingly, the cytoskeleton or injury biomarker module rKID:2m was also strongly preserved in human, showing the relevance of the biological context of the renal biomarkers that cluster in similar co-expressed modules (Figure 5D; Table S10). Other potential biomarker modules with strong pathology association were also preserved in human (Tables 1, 2, and S10). By contrast, some of the modules were not (or hardly) preserved (Zsum ≤ 2) in the available human dataset, including modules reflecting cellular stress responses, such as ATF4 (rKID:74), oxidative stress (rKID:53, 105, and 106), as well as modules annotated for ribosomal biogenesis (rKID:6m and 8m), renal development (rKID:23m, 45m), circadian regulation (rKID:33), immune response (rKID:15, 56), steroid hormone metabolism (rKID:18m and 95), and several other metabolic processes (Table S10).

Next, the human dataset was uploaded into the rat kidney TXG-MAPr tool (after gene ortholog conversion) to see how the human responses were captured using only the preserved modules (Figures 6 and S13; Table S13). As expected, immune responses (rKID:7m, 16m, 63m, 113, 138, 212m, and 305), extracellular matrix (rKID:10m, 31m, and 104), and cell migration (rKID:127) modules were most strongly induced in human transplant samples and clustered together in the heatmap (Figure S13, cluster 3). More inter-individual variation was seen in modules involved in mitochondrial function, metabolism and transport, which grouped together in cluster 2 (Figure S13). Immune response and injury biomarker modules showed on average stronger activation in patients with immune-mediated renal rejection despite some inter-individual variation, while metabolism was stronger downregulated in these patients (Figures 6A and S14). Similar patterns could be seen for the log2 FC of hub genes and most significant genes of each of these inflammation modules (Figures 6A and S14). In addition, inflammation modules rKID:7m and rKID:63m showed a very strong correlation (Pearson’s r = 0.97) across all patient samples (Figure 6B, left), which was similar for most of the inflammation and cell migration modules (Figure 6C, cluster 3). Also, the biomarker module rKID:2m showed strong correlation (Pearson’s r = 0.72) with inflammation module rKID:7m (Figure 6B, center). In contrast, metabolism module rKID:3m showed inverse correlation (Pearson’s r = −0.62) with inflammation module rKID:7m (Figure 6B, right) and rKID:2m, similar to responses in rat kidney. This bi-directional response was also seen for other metabolism, mitochondrial and transport modules compared to inflammation (Figure 6C, cluster 1 vs. 3). This indicates that metabolism, mitochondrial function, and transport are down regulated in patients’ kidneys undergoing renal inflammation, which was also noted in rats. A large group of modules with various annotations for RNA and protein processing showed very high correlation (Figure 6, cluster 4), which was mainly due to downregulation of these modules in a group of patients with highest AvAbsEGS. With the presented gene co-expression approach in pre-clinical models, we could identify several modules that were preserved in human, indicating the relevance for human translation. Moreover, we showed differential gene network dynamics in human patients with rejected renal grafts that could be linked to immune responses, injury, and biomarker expression, indicating that human and rat have similar generalized tissue responses to kidney injury and regeneration.

Figure 6.

Human preservation and network dynamics

(A) Boxplot for the different human transplant groups showing the module EGS (left) and the log2 fold change (log2 FC) of the hub gene (second column) and most significant genes of inflammation module rKID:7m. The boxplots show the 25%, 50% (median), and 75% quartiles as well as the individual points of each sample. The lines on the boxes represent the minimal and maximal values except for the outlier samples. The groups are indicated by rejection type: non-rejecting; ABMR, antibody-mediated rejection; TCMR, T cell-mediated rejection; MIXED, both antibody and T cell-mediated rejection. Significant differences between groups were calculated using a Wilcoxon test (∗∗∗p < 0.001). Significant gene log2 FC of individual patients were indicated in blue (p-adjust < 0.05).

(B) Module EGS correlation plot of inflammation modules rKID:7m vs. rKID:63m (left), inflammation module rKID:7m vs. injury biomarker module rKID:2m (center), and inflammation rKID:7m vs. metabolism rKID:3m modules. Strong correlation is shown between the two inflammation and injury biomarker modules, while there is inverse correlation between the inflammation module rKID:7m and the metabolism module rKID:3m.

(C) Cluster correlation heatmap (based on Pearson correlation of module EGS) of the preserved modules in human with an EGS >2. There is strong clustering of the inflammation, ECM and cell migration modules (cluster 3), and cellular stress or injury related modules (cluster 4). The mitochondrial function, metabolism, and transport modules cluster together as well (cluster 1) but show anti correlation with the inflammation and stress/injury modules. Heatmap is clustered by Euclidean distance, using the complete method from pheatmap package.

Discussion

In pre-clinical research, a wealth of toxicogenomic data has been generated in the past decades, with the aim to investigate biological responses to drugs and to understand the mechanisms of toxicity.⁷ Unfortunately, while standardized reporting approaches for bioinformatics have been proposed,^41,42,43 a comprehensive qualitative and quantitative mechanistic interpretation framework for toxicogenomics underpinning a mechanism-based safety assessment approach is needed.^7,8 A major hurdle has been the data complexity and the inability to generalize conclusions from transcriptomics responses across different cell types, tissues, and species. Methods that reduce complex transcriptomic data to networks of co-expressed genes that retain biological meaning are an attractive approach to reducing the dimensionality of transcriptomic data and facilitating interpretation,^15,44 as demonstrated previously for rat liver toxicity.^10,11,12,13 Besides liver injury, kidney toxicity is also a common adverse response in both drug and chemical toxicity.¹ Therefore, methods that facilitate the extraction of mechanistic information from kidney transcriptional responses would be useful in hazard identification and risk assessment.

Herein, we used WGCNA to construct a set of 347 co-regulated gene networks (modules) from the two largest, publicly available in vivo rat kidney toxicogenomics datasets (TG and DM) and provide biological annotations for the modules allowing direct translation of gene expression data into meaningful biological conclusions. The modules are embedded in an interactive R-shiny application, called the rat kidney TXG-MAPr, that allows dose- and time-response analyses, treatment correlation, and functional annotation (GO terms, pathways, and TF enrichment). We demonstrated that this gene co-expression approach and analysis framework has the advantage that modules and their quantitative EGS can assist in (1) data complexity reduction, (2) biological annotation for mechanistic interpretation, (3) pathology association, (4) biomarker discovery, and (5) translation to human relevant responses.

First, we have demonstrated that the co-expression modules could identify meaningful biological responses relevant to mechanistic investigation using cisplatin exposure as the example. Early activation of DNA-damage and p53 signaling (rKID:160) was followed by later occurrence of cellular injury and biomarker expression (rKID:2M), regeneration (rKID:5m), and inflammation (rKID:56, 63m, 212m, and 327). These later responses, including activation of biomarker-containing, regeneration, and inflammatory modules, were mirrored with other nephrotoxicants that had concurrent renal histopathology and clinical chemistry changes. This suggests that similarity in gene expression patterns reflects common injury-induced repair mechanisms (e.g., regeneration, dedifferentiation, and inflammation responses). The common injury and regeneration processes across different nephrotoxicants, as well as during ischemic injury, are compound agnostic and general tissue responses to injury and repair, where a complex interplay between renal inflammation and cellular dedifferentiation or partial EMT are contributing to renal repair.⁴⁵ However, at early stages after nephrotoxic compounds exposure, there was a distinct activation of biological response networks relevant to the expected different mechanisms of injury. For example, the induction of rKID:160 and rKID:298 (the modules annotating for p53 responses) is much stronger and unique for CSP, carboplatin, and doxorubicin, due to their genotoxic mechanism of action, compared to other nephrotoxicants, such as PAN. In contrast, the primary key event of puromycin is an early inflammatory response, which the anticipated mechanisms of action that can induce glomerular nephropathy.³⁴ In addition, estrogen receptor agonists are the main inducers of steroid hormone metabolism modules, which is expected based on their primary mechanism of action. Thus, the rat kidney TXG-MAPr tool also allows evaluation of molecular mechanisms and early mode-of-action of drugs and chemicals with respect to onset of nephrotoxicity. The module annotations that we provide in this study enhance a qualitative and quantitative mechanistic interpretation of renal toxicogenomics data and capture most of the well-known biological processes. In the future, AI-driven annotation could further enhance and accelerate the biological annotation of the module, although expert curated interpretation will still be needed.

Various studies have applied gene co-expression network analysis on rat and human kidney injury samples.^{46,47,48,49,50,51} A previous study applied gene co-expression analysis on the rat kidney DM data and also identified modules that correlated with AKI, which contained Havcr1 and Clu genes,⁴⁶ which are in module rKID:2m with the strongest pathology association in our study as well. Most other studies identified a small number of modules due to the limited number of samples used to build the gene co-expression networks, but the captured modules include stress and immune responses, which were also the most relevant modules for pathology in our study. However, these previous studies did not integrate their kidney gene co-expression networks into a publicly accessible online tool, such as the rat kidney TXG-MAPr, underpinning its value. Though there are co-expression tools built from data of various species and tissues that are publicly available,^{52,53,54,55,56} these tools mainly show gene interactions, pathway analysis, or conservation across species. In contrast, the rat kidney TXG-MAPr tool allows the users to interactively visualize quantitative changes in gene networks by chemical or biological perturbations. In addition, users can upload log2 FC gene expression data for comparison to the reference toxicogenomics dataset in the TXG-MAPr tool (Figure S15), which is highly valuable to identify common mechanisms of action for chemical hazard identification.

Inflammation accompanied by metabolic re-wiring are hallmarks of acute tubular injury and the balance between injury and repair.³⁵ Thus, the juxtaposition of inflammation module induction and repression of metabolism and mitochondrial modules is worth noting. Repression of module rKID:3m (highly annotated for small molecular metabolism and mitochondrion [Table S3]) ranked among the top ten modules (based on effect size and p-adjust) for multiple pathologies (Tables S8 and S9). Conversely, modules rKID:7m, 63m, and 327 are annotated for immune response and were among the top five ranked inducible modules associated with the presence of cellular infiltrates, i.e., inflammation (Tables 2 and S8). Thus, strong activation in regeneration and inflammation responses impacted on cellular energy production (metabolism and mitochondrial function) in the kidney, which was also concluded previously for liver diseases.⁵⁷ The bi-directional response of these processes was also demonstrated in human patients with renal transplants (Figure 6C), suggesting that these injury and repair responses are also conserved between species.

In this study, we have identified modules with strong statistical association with histopathology findings and biomarkers of tissue injuries. Module rKID:2m was most strongly associated with concurrent and predictive renal pathology and contained many well-known kidney injury biomarkers, including Havcr1 (Kim-1), Lcn2 (lipocalin 2), Spp1 (osteopontin), and Clu (clusterin). In addition, several immune response modules ranked highly for association with concurrent and predictive kidney injury (Tables 2 and S9), including modules rKID:327, rKID:56, and rKID:212m, containing hub genes that are candidate biomarkers such as fibrinogen (Fga, Fgb, and Fgg), Icam1, and ceruloplasmin (CP), respectively, which were also reported as novel renal biomarkers.^58,59,60,61 This suggests that candidate injury biomarkers could be discovered using gene co-expression networks that are strongly associated with pathology, as was recently reported for liver and kidney.^11,62 This approach was validated in a recently published study where early activation of injury modules after CSP exposure will lead to renal pathology in rats by 28 days, including necrosis, inflammation, and fibrosis.⁶³ It was shown that urinary clusterin correlated well with Clu gene expression, which is an important biomarker in rKID:2m that showed the strongest pathology association. The subset of both concurrent and predictive kidney injury-associated modules is important for chemical or drug development since aberrant transcriptional activity could indicate higher risk of developing adversity. Moreover, association of (early) WGCNA module changes that are predictive of late-stage kidney pathology is expected to impact drug candidate optimization and enable prioritization of safer compounds.

Preservation analysis is a unique feature of network-based models allowing not only the robustness and reproducibility of modules to be tested using datasets from the same experimental species but also the likelihood of translation across species, which is an important consideration in drug or chemical hazard assessment. All rat kidney modules had at least moderate self-preservation in the TG and DM rat kidney transcriptomic data, indicating that the modules were robust and reproducible representations of biological response networks in rat kidneys. However, only 130 (37%) of the 347 rat modules were preserved (translational modules) when compared to a human renal transplantation transcriptomic dataset. Among the translational modules were those that reflect immune response, cell cycle, metabolism, mitochondrial function, RNA processing, and ECM remodeling/fibrosis modules. These biological programs are deterministic of end-stage renal disease, involving inflammatory and regenerative programs that are part of renal transplant rejection.³⁵ By contrast, various modules that reflect early cellular stress response programs, including ER stress activation of Atf4, Nrf2-mediated oxidative stress, as well as modules annotated for ribosomal biogenesis, renal development, circadian regulation, steroid hormone metabolism, were not preserved in human. These differences are likely due to the nature of the human dataset, which only contained human transplant samples, where early injury or disease responses, including adaptive cellular stress responses, are not captured in these late-stage human transplant disease settings.¹⁸ Nonetheless, the fact that important toxicity-related rat kidney modules are preserved in human kidney samples suggests that the preservation approach is an attractive and valid strategy to determine the likelihood that mechanisms of renal injury observed in rat kidney will translate to human. Future studies should focus on collection of more extensive and diverse human renal pathology transcriptomics datasets representing a broader spectrum of early and late-stage disease phenotypes, which may improve the preservation statistics and confidence that rat modules will translate to human, which is significant knowledge for chemical safety assessment. Currently, we could already identify multiple hub genes from modules that were strongly associated with kidney pathology and did translate to human, of which several were already described in the literature as candidate injury biomarkers (Table 2). Newly identified hub genes from these modules could be validated in follow-up studies as human renal biomarker, which was beyond the scope of this study. Those potential biomarkers should be stability expressed proteins and should be excreted or released during events of injury or cell death to be picked up in urine.

In conclusion, the rat kidney TXG-MAPr tool provides a user-friendly interface that enables visualization of gene expression data in the context of co-expressed modules that can be mined for the association of important biological processes in acute and chronic renal injury. The tool can be utilized for identifying possible safety liabilities and/or mechanisms that can lead to adversity, useful for chemical safety assessment. Association of early WGCNA module changes that are predictive of later kidney pathology is expected to impact lead optimization and enable prioritization of better compounds that are less likely to induce pathology. We foresee that quantitative gene co-expression modules, which are strongly associated to kidney pathology, and which translate from rat to human, could potentially be implemented in translational quantitative systems toxicology models for mechanistic evaluation and prediction of kidney injury of chemical or drug candidates.⁶⁴

Limitations of the study

In this study, module annotations were based on expert judgment of the enriched terms. Currently, AI models could be used to summarize enrichment terms in common annotations, although expert judgment remain essential to verify annotation calls. Several modules were identified that showed strong association with pathology and did contain established and candidate renal biomarker genes. While some candidate biomarkers from modules have been assessed in other studies, further validation of these kidney injury biomarkers was beyond the scope of this study. We have provided a proof of concept that several rat kidney modules translate to human, but the diversity of human data was limited to renal transplant samples. In the future, the confidence of cross species translation could be improved by including larger panels of human biopsy samples, covering different types and stages of human kidney diseases.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Prof. Bob van de Water (water_b@lacdr.leidenuniv.nl).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Raw microarray data have been deposited at EMBL-EBI ArrayExpress as E-MTAB-13633 and are publicly available as of the date of publication. This paper also analyzes existing, publicly available data, accessible at Open TG-GATEs, DrugMatrix, or Gene Expression Omnibus (GEO). The accession numbers for the datasets are listed in the key resources table.
• •All original code has been deposited at Zenodo and is publicly available at https://doi.org/10.5281/zenodo.14926143 as of the date of publication.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work has received funding from the EU-EFPIA Innovative Medicines Initiative 2 (IMI2) Joint Undertaking TransQST project (grant number 116030) and eTRANSAFE project (grant number 777365). This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation program and EFPIA, the EC Horizon2020 EU-ToxRisk project (grant number 681002), the EC Horizon2020 RISK-HUNT3R project (grant number 964537), the Horizon Europe PARC programme (The European Partnership for the Assessment of Risks from Chemicals, grant number 101057014), and the Virtual Human Platform for safety assessment (VHP4SAFETY) project (grant number 1292.19.272, which is part of the NWA research program “Research along Routes by Consortia [ORC]”, which is funded by the Netherlands Organisation for Scientific Research [NWO]).

Authors declare that this work reflects only the author’s view, and that the European Commission and IMI-JU are not responsible for any use that may be made of the information it contains.

Author contributions

S.J.K., G. Callegaro, J.J.S., and P.T. performed data analysis. S.J.K., G. Callegaro, and H.W.v.K. built the R-shiny TXG-MAPr tool. S.J.K., L.S.W., G. Chung, K.P., C.B., K.M.G., C.R.T., J.L.S., and B.v.d.W. designed experiments. S.J.K., G. Chung, K.P., and K.M.G. performed experiments. G. Callegaro, J.J.S., P.T., C.P.F., J.S.-R., C.B., S.A.E., and K.M.H.-T. provided expertise and feedback. S.J.K., J.L.S., and B.v.d.W. conceptualized, wrote, and edited the manuscript. J.L.S. and B.v.d.W. acquired research funding and supervised the study. All authors approved the manuscript.

Declaration of interests

J.S.-R. reports funding from GSK, Pfizer, and Sanofi, and fees/honoraria from Travere Therapeutics, Stadapharm, Astex, Owkin, Pfizer, and Grunenthal. P.T. is a Sanofi employee and may hold shares and/or stock options in the company.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Transcriptomics data of rat primary proximal tubule cultures (PTC) during in vitro differentiation	This paper	EMBL-EBI ArrayExpress: E-MTAB-13633
Open TG-GATEs rat in vivo kidney transcriptomics data (single and repeated exposure) https://dbarchive.biosciencedbc.jp/data/open-tggates/LATEST/Rat/in_vivo/Kidney/	Igarashi et al.16	EMBL-EBI ArrayExpress: E-MTAB-799, E-MTAB-800
DrugMatrix rat in vivo kidney transcriptomics data	Svoboda et al.17	GEO series: GSE57811
Rat renal ischemia reperfusion injury (IRI) transcriptomics data	Speir et al.36	GEO series: GSE58438
Transcriptomics data from human kidney transplant biopsies	Reeve et al.18,40	GEO series: GSE36059
Experimental models: cell lines
Rat primary proximal tubule cultures (PTC)	NewCells Biotech	N/A
Software and algorithms
TXG-MAPr tool (shiny app)	This paper, Callegaro et al.10	https://txg-mapr.eu/login/
WGCNA and preservation analysis script	This paper	https://doi.org/10.5281/zenodo.14926143
R version > 4.0.0	R Core Team65	https://cran.rstudio.com/
stats R package	R Core Team65	https://cran.rstudio.com/
WGCNA R package	Langfelder et al.15	https://cran.rstudio.com/
affy R package	Gautier et al.66	https://cran.rstudio.com/
limma R package	Ritchie et al.67	https://cran.rstudio.com/
topGO R package	Alexa et al.68	https://cran.rstudio.com/
Consensus Pathway DB	Kamburov et al.69	http://cpdb.molgen.mpg.de/CPDB
DoRothEA R package	Garcia-Alonso et al.70	https://bioconductor.org/
viper R package	Alvarez et al.71	https://bioconductor.org/
shiny R package	Chang et al.72	https://cran.rstudio.com/
ggplot2 R package	Wickham73	https://cran.rstudio.com/
pheatmap R package	Kolde74	https://cran.rstudio.com/
BrainArray chip description files (CDF)	Dai et al.75	http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/CDF_download.asp

Experimental model and study participant details

Differentiation of primary rat proximal tubular cells

For isolation of primary rat proximal tubular cells (PTC) 3 male Sprague Dawley rats (10 weeks old) were sacrificed and kidneys were harvested and decapsulated. The rats were supplied by Charles River, UK, and were housed at Newcastle University before being euthanized by cervical dislocation. Ethical approval for the isolation and culture of primary proximal tubules cells from rats was obtained from the Animal Welfare and Ethical Review Body at Newcastle University and all procedures were carried out in accordance with Animals (Scientific Procedures) Act 1986. Newcastle University holds OLAW Animal Welfare Assurance certification for animal welfare. At least 1 mg of whole kidney biopsy sample was soaked in 1 mL of RNAlater. Cortical slices were taken through the kidney samples from each animal and rat PTCs were isolated from the minced cortical slices as previously described.^76,77 Approximately 1 million freshly isolated rat PTCs were soaked in 1 mL of RNAlater serving as the day 0 sample. The remaining PTCs were seeded onto 12-well Transwell inserts at density of 200,000 cells per insert. Medium was changed on day 1, 3 and 5 after seeding. PTCs were harvested day 1, 2, 5 and 7 post-seeding. On the day of RNA sampling, media was gently removed from the wells, membranes were cut out and put into 10 mL of RNAlater (Thermo Fisher) per timepoint. Samples were stored at 4°C overnight, then at -20°C.

Method details

Sample collection and gene expression analysis

Toxicogenomic data from the TG-GATEs (TG) rat kidney repository (Affymetrix Rat Genome 230 2.0 array) were downloaded¹⁶ and jointly normalized using the Robust Multi-array Average (RMA) method within the affy package using R.^65,66 The BrainArray chip description file (CDF) version 19 ( http://brainarray.mbni.med.umich.edu/Brainarray/Database/CustomCDF/CDF_download.asp, Rat230_2 array version 19) was used to map Affymetrix probesets to NCBI gene Entrez_IDs,⁷⁵ which resulted in 13877 unique probe sets, each mapped to a single gene. The TG kidney repository contains 975 rat treatments, where each treatment is defined as a combination of compound, time and concentration. Vehicle treated control samples are available for each individual time point (time-matched controls). The limma R package was used to calculate log2 fold change (log2FC) values, by building a linear model fit and computing differential expression by empirical Bayes moderation for each treatment condition.⁶⁷ Toxicogenomic data from the DrugMatrix (DM) rat kidney repository and a renal ischemia reperfusion injury (IRI) study were downloaded from Gene Expression Omnibus (GEO https://www.ncbi.nlm.nih.gov/geo/; GSE57811 and GSE58438) and analysed following the same procedure as the TG data.^17,36 For the DM data the time-matched vehicle control samples were indicated per treatment condition. The IRI study had one group of naïve treated rat controls (n=5), which was used for differential gene expression at the different timepoints after ischemia reperfusion. Human Affymetrix transcriptomics data (Human Genome U133 Plus 2.0 Array) from human patients after kidney transplantation were collected from GEO (GSE36059),^18,40 annotated using BrainArray CDF version 25 (HGU133Plus2) and normalized using RMA in the affy R package. One CEL file sample was removed because it was deprecated. Differential gene expression of each patient kidney transplant sample was contrasted with the eight nephrectomy controls samples in the dataset using the limma package.

RNA was isolated from rat kidney samples and PTC cultures using Trizol + Direct-zol columns according to the referenced protocol (Zymo Research). For PTC cultures, RNA was isolated from Transwell filters (multiple filters were pooled per timepoint) and concentrated over RNA Clean & Concentrate (RCC) columns (Zymo Research). RNA quality was assessed by Agilent Bioanalyzer and all samples had a RIN > 6.3. For microarray analysis, 1 μg of total RNA was used for preparation of biotin-labeled cRNA. Samples were hybridized to Affymetrix Clariom-D rat arrays for transcriptome analysis and scanned on an Affymetrix GeneChip Scanner. Raw CEL files were annotated using BrainArray CDF version 25 (ClariomDRat) and normalized using RMA in the affy R package. Differential gene expression for the PTC cultures was contrasted with the whole kidney samples using the limma package to investigate the transcriptional response of cells taken into culture. Raw and processed data of the PTC cultures is deposited in EBML-EBI ArrayExpress database (E-MTAB-13633).

Rat kidney gene co-expression network analysis

Kidney specific co-regulated gene networks (modules) were obtained from the TG and DM rat kidney log2FC data matrix,^16,17 using the WGCNA package in R.¹⁵ The rat kidney TG data contains transcriptomic data of single (3, 6, 9 and 24 hours) and daily repeated (4, 8, 15 and 29 days) exposures for 41 (drug) compounds at various dose levels (i.e., 975 compound, dose and time combinations (conditions) in total). The rat kidney DM data contains transcriptomics data from daily repeated exposures with 139 (drug) compounds at 1 or 2 dose levels and for up to 14 days (i.e., 360 treatment conditions). We applied unsigned WGCNA to the log2FC matrix of all TG and DM data to group co-induced and co-repressed genes together. The optimal soft power threshold of 8 was selected based on the standard power-law plotting within the WGCNA package and by maximizing the difference in basal gene expression of genes showing co-expression (module genes) versus genes not showing co-expression (excluded genes) based on a t-test with the assumption that low expressed genes are not correlating. From the 13877 unique genes on the Rat230_2 array, we obtained 399 co-expression modules containing 11,244 genes, while 2633 excluded genes did not meet the co-expression criteria. Modules were merged when having a pair-wise Pearson’s r correlation ≥ 0.8 of the module eigengene scores (EGS) based on all treatments, which resulted in a final list of 347 merged modules (Table S1). Merged modules contain the suffix ‘m’, i.e., module rat KIDNEY:2m (i.e., rKID:2m) was merged from child modules 2 and 11. For each treatment condition a module EGS was computed as the first principal component of variation for all genes after Z-scoring the log2FC matrix (Table S2), as described previously.^10,12 This module EGS represents a quantitative measure of activation or repression of all genes (based on log2FC) in the module. An EGS > 2 or < -2 was considered as a large perturbation in expression of the underlying genes.¹³ The correlation between the log2FC of each gene with the module EGS was also calculated (called the corEG = correlation eigengene) to estimate the intramodular connectivity of the genes and their network (note that genes can also have opposite (negative) weighting because of the unsigned co-expression). The gene with the highest absolute corEG within a module is the hub gene and is the most representative gene of the module. The circular rat kidney TXG-MAPr dendrogram was constructed using the ape package based on Ward’s hierarchical clustering of pair-wise Pearson correlation for every module across all treatment conditions (Figure S1). Each main branch is indicated by letters A-I, followed by subbranches (i.e., A1aIαi) as indicated in Figure S1 and Table S1. To improve visualization of the dendrogram, some module edges and nodes were manually adjusted to separate module clusters. The module EGS are displayed on the TXG-MAPr dendrogram as circles, with the size and colour proportional to the amount of induction or repression for a given treatment condition (Figure 1A, red to blue colour scale, respectively). Compound or treatment correlation was calculated for all treatment conditions using pair-wise Pearson or Spearman correlation of the module EGS. Similarly, module correlation was calculated across all modules (347 x 347 matrix). Preservation of the TG and DM rat kidney WGCNA module structure with other datasets (TG rat kidney only, DM rat kidney only, TG rat liver and GEO human kidney) was performed using the modulePreservation function within the WGCNA package in R.³⁹ Modules with a Z-summary > 2 were considered moderately preserved and Z-summary > 10 were highly preserved. The R-script for WGCNA module generation and preservation analysis can be found in an online GitHub repository linked to persistent identifier in Zenodo: https://doi.org/10.5281/zenodo.14926143.

Functional enrichment analyses and annotation of modules

Modules were annotated by functional enrichment (GO-terms, pathways and transcription factors) using hypergeometric tests (Tables S3, S4, and S5) as described previously.¹⁰ Briefly, GO-term enrichment was performed with the topGO package in R using the algorithm = "classic" and statistic = "fisher".⁶⁸ Over Representation Analysis (ORA) was performed on the gene members of each module using Consensus Pathway DB (cpdb version 34), including the following databases: BioCarta, EHMN, HumanCyc, INOH, KEGG, NetPath, Reactome, PharmGKB, PID, Signalink, SMPDB, Wikipathways, UniProt, InterPro.⁶⁹ For both resources, we included enriched terms with a hypergeometric test p-value < 0.01 (Table S3). To identify regulatory transcription factors (TFs) of each module, a hypergeometric test was performed on module gene members using the function phyper within the stats package in R (Table S4). The gene set of TFs and their regulated genes (regulons) are derived from DoRothEA⁷⁰ with two sets of confidence levels: the “high confidence” level comprises categories A, B and C, while the “high coverage” level comprises categories A, B, C and D. The enriched TFs with p-value less than 0.01 were included in the study. In parallel, TFs’ activities were estimated as normalized enrichment scores using the viper function from the viper package⁷¹ with two confidence sets of TF-regulon from DoRothEA as described. All parameter settings were assigned as in the original DoRothEA study.⁷⁰ The most significant module enrichment terms (gene overrepresentation analysis) were reviewed for a common theme, which were summarized in annotation levels 1 through 3 and a key annotation (Table S5). Only modules with an enrichment p-value < 10^-4 (0.0001) were reviewed to find a common annotation. Annotation level 1 describes the general cellular process or function (e.g., metabolism, immune response, RNA/protein processing, stress response), level 2 describes the sub-process or function, and level 3 provides the most specific process or transcription factor involved. These 3 levels of annotation were summarized in the Key_annotation, which describes the function of the genes in the modules based on the most significant enrichment terms (Table S5).

Module association with pathology

The TG database also provides clinical chemistry (biomarker data) and histopathology data for all treatment conditions. Histopathology data was reviewed by pathologists for 4 nephrotoxic compounds (cisplatin, carboplatin, cyclosporine A and allopurinol) and most calls were in agreement with the pathology score provided by TG. Only the most occurring pathologies were included for the statistical association, including necrosis, degeneration, regeneration, basophilic change (cellular staining indication a regenerative process), cellular infiltration (indicating an inflammatory/ immune response), fibrosis, vacuolization, dilatation (widening of the renal tubules), cysts and hyaline casts (protein aggregates in the tubular lumen). Overlapping pathology terms were combined (e.g., different types of cellular infiltration or different regions of tubular injury), resulting in a final list of 19 unique histopathology’s (Tables S6 and S7). Necrosis and degeneration were also considered as a combined term because these pathologies have overlapping biological meaning, as well as regeneration and basophilic change. For these overlapping pathologies, the maximum grade was used for each treatment condition. Histopathology severity was converted in a numerical scale (normal = 0, minimal = 1, slight = 2, moderate = 3, severe = 4) and the average score was calculated per treatment group of maximum 3 animals that were also used for the microarray analysis. Histopathology presence in a treatment group was indicated by a severity score >= 0.67 for all pathologies, but a higher threshold (>=1.33) was also applied for a few pathologies (Table S7). Clinical chemistry measurements were calculated as average percent change from the time-matched control group. The KW-BW ratio (kidney weight - body weight ratio) was calculated by dividing the total kidney weight by the body weight. The threshold for a significant increase in clinical measurements is defined in Table S7. The kidney toxicity phenotypes (histopathology or clinical chemistry) were used to evaluate statistical association between module EGS and the occurrence of kidney pathology, as described previously.¹³ Both concurrent and predictive module associations with pathology (i.e., early module changes that are preceding the toxicity at a later timepoint) were calculated. Effect sizes (Cohen’s D) of each module were calculated by comparing mean module EGS in the presence (positives) or absence (negatives) of renal injury, with thresholds specified in Table S7, for each of the 29 toxicity phenotypes (Table S8).

Cohen’s D = (mean module EGS positives – mean module EGS negatives) / pooled SD module EGS.

To control for overall gene expression, we applied logistic regression to account for differences in both module EGS and average absolute EGS (AvAbsEGS) as covariates in the analysis, as described previously.¹³ Briefly, the glm() function of the stats package was applied with single linear models:

Module_single = glm(phenotype ∼ module EGS, family = “binomial”)

AvAbsEGS_single = glm(phenotype ∼ AvAbsEGS, family = “binomial”)

The Module_single model predicted the outcome of using only module EGS as variable and provided p_single (Table S8). In addition, the full linear model was calculated:

Module_Av_full = glm(phenotype ∼ AvAbsEGS + module EGS, family = “binomial”)

This determined the coefficients, which is the natural log odds ratio for observing a phenotype for a single-unit increase in module EGS. Finally, the significance of module EGS as an additional variable was tested, which provided the p-adj (Table S8):

add1(AvAbsEGS_single, scope = ∼ AvAbsEGS + module EGS, test = “Chisq”)

The p-adj value was converted to a signed log10 p-adj, by using the sign of the effect sizes. This signed log10 p-adj and the Cohen’s D effect size were used as measure/rank for module associations with toxicity phenotypes in several plots and tables.

Rat kidney TXG-MAPr webtool

An interactive visualization of rat kidney WGCNA module and gene expression data has been implemented in the rat kidney TXG-MAPr (Toxicogenomics-MAPr) webtool using the shiny package in R to create a user-interface,⁷² as described previously.¹⁰ The rat kidney TXG-MAPr tool is available at https://txg-mapr.eu/login/. The tool allows evaluation of dose- and time-response data, compound correlation plots and module association with pathology, which is also available in tabulated format. New or external gene expression data (log2FC) can be uploaded in the TXG-MAPr tool, as described previously.¹⁰ Briefly, users need to calculate gene log2FC values for their treatment of interest and upload these in the TXG-MAPr tool (Figure S15). This enables calculation of new EGS for each module from external dataset based on the gene log2FC and weighted by the gene corEG. The new module EGS will be overlaid onto the rat kidney TXG-MAPr dendrogram and will be fully integrated into the web application for that session. Uploaded data will be removed when the session is closed. A guide how to use the TXG-MAPr tool can be found in the manual tab of the tool.

Quantification and statistical analysis

All (sub)figures were plotted in R version 4.0.0⁶⁵ or higher using mainly ggplot2, pheatmap, shiny, igraph and ape packages. Main and supplemental figures were combined in Adobe Illustrator and graphical abstract created in BioRender (https://biorender.com/0gk7uzm). Statistical analyses were carried out using functions from R packages for differential gene expression analysis (affy, limma), correlation analysis and logistic regression (stats), enrichment analysis (topGO, stats, Consensus Pathway DB, DoRothEA), gene co-expression and preservation analysis (WGCNA). Multiple testing correction was applied on p-values derived for gene log2FC using default settings of limma. Gene log2FC values with a p-adjust < 0.05 were considered statistically significant. Modules with EGS > 2 or EGS < -2 indicate a large perturbation in expression of the co-expressed genes and were considered significant.¹³

Additional resources

The rat kidney TXG-MAPr tool is available at https://txg-mapr.eu/login/.

Published: June 21, 2025

Supplemental information

Table S1. Rat kidney WGCNA module definitions, related to Figure 1

For each gene the official gene symbol, Entrez_ID, module, correlation eigengene, and branch number are provided.

Table S5. Rat kidney TXG-MAPr module annotation, related to Figure 1

For every module the following characteristics are provided: module number, number of genes in the module, hub-gene, NCBI Entrez_ID of the hub-gene, correlation eigengene score (corEG), key annotation, annotation lvl1, annotation lvl2, annotation lvl3, branch number, preservation in human, top 100 genes ranked on highest corEG (or less for smaller modules), top 10 highest GO-terms (−log10 p value), top 10 highest pathway enrichment (−log10 p value), top 5 highest transcription factor enrichment with confidence levels A-C or A-D (from DoRothEA), top 10 compounds showing strongest module perturbation. For modules with a clear GO-term, pathway or transcription factor enrichment (threshold: p < 10–4) an annotation is provided.

Table S6. TG-GATEs rat kidney pathology scores and clinical chemistry biomarker changes, related to Figures 2 and 3

Mean histopathology scores and clinical chemistry biomarker changes are shown for the three replicates that were used for transcriptomic analysis. BUN, blood urea nitrogen; CRE, serum creatinine; IP, inorganic phosphate; BW.KW ratio, kidney weight to body weight ratio.

Table S7. Rat kidney toxicity phenotypes and their prevalence in the TG-GATEs dataset, related to Figures 2 and 3

The toxicity threshold is identified by a significant percentage change from the control group or by the average pathology score of the three animals. Two pathology clusters could be identified in Figures 2D and S5A: (I) mild pathologies with only weak association to gene expression changes and (II) severe renal pathologies with strong association to gene expression changes. BUN, blood urea nitrogen; CRE, serum creatinine; IP, inorganic phosphate; KB.BW.ratio, kidney weight to body weight ratio.

Table S8. Rat kidney TXG-MAPr (WGCNA) module association with pathology, related to Figures 2, S4, and S5

For concurrent expression, establish whether a given pathology finding was positive at the same time point. For expression ≤1 day, establish whether a given finding was positive at any time point ≥4 days. For expression at 4 days, establish whether a given finding was positive at any time point ≥8 days.

Table S9. Rat kidney TXG-MAPr module association with pathology clusters from Figure S5, related to Figures 2, S4, and S5

The mean adjusted p values and effect sizes were calculated from the 17 most strongly associated toxicity phenotypes (see also Table S7, and Figures 2D and S5A cluster II).

Table S10. Preservation of rat kidney WGCNA modules across datasets (TG_RAT_KIDNEY, DM_RAT_KIDNEY, TG_RAT_LIVER, GEO_HUMAN_KIDNEY), related to Figure 5

Preservation is calculated by a Z-summary score (Zsum), where Zsum >2 indicates low/moderate module preservation and Zsum >10 indicates strong module preservation. Module coverage for human data was calculated after gene orthology conversion to rat using the RGD database. Modules were selected when they showed human preservation (Zsum > 2), strong concurrent pathology association (see also Figures 2D and S5A, and Tables 2 and S9) or strong predictive association (see also Table S9).