Universal Toxicity Gene Signatures for Early Identification of Drug-Induced Tissue Injuries in Rats

Abstract

A new safety testing paradigm that relies on gene expression biomarker panels was developed to easily and quickly identify drug-induced injuries across tissues in rats prior to drug candidate selection. Here, we describe the development, qualification, and implementation of gene expression signatures that diagnose tissue degeneration/necrosis for use in early rat safety studies. Approximately 400 differentially expressed genes were first identified that were consistently regulated across 4 prioritized tissues (liver, kidney, heart, and skeletal muscle), following injuries induced by known toxicants. Hundred of these “universal” genes were chosen for quantitative PCR, and the most consistent and robustly responding transcripts selected, resulting in a final 22-gene set from which unique sets of 12 genes were chosen as optimal for each tissue. The approach was extended across 4 additional tissues (pancreas, gastrointestinal tract, bladder, and testes) where toxicities are less common. Mathematical algorithms were generated to convert each tissue’s 12-gene expression values to a single metric, scaled between 0 and 1, and a positive threshold set. For liver, kidney, heart, and skeletal muscle, this was established using a training set of 22 compounds and performance determined by testing a set of approximately 100 additional compounds, resulting in 74%–94% sensitivity and 94%–100% specificity for liver, kidney, and skeletal muscle, and 54%–62% sensitivity and 95%–98% specificity for heart. Similar performance was observed across a set of 15 studies for pancreas, gastrointestinal tract, bladder, and testes. Bundled together, we have incorporated these tissue signatures into a 4-day rat study, providing a rapid assessment of commonly seen compound liabilities to guide selection of lead candidates without the necessity to perform time-consuming histopathologic analyses.

Drug candidate attrition at the stage of nonclinical development due to drug-induced tissue injury seen in animals can have significant impact on timelines to early clinical trials. The most common target organs of drug-induced tissue injury observed historically in our internal rat drug development toxicology studies are liver and kidney, followed by skeletal muscle and heart (Sistare et al., 2018). These tissue injuries, defined as degeneration/necrosis, account for approximately 85% of the tissue injury seen in rat acute toxicity studies (Sistare et al., 2018). Similar experience has been reported by others who have noted that 1-month studies are often sufficient to detect most drug candidates destined to present with tissue toxicities in rat (Olson et al., 2000). For discovery teams, the selection of lead candidates typically relies on efficacy screening and study data from animal models. Following lead selection, the teams are then faced with the additional cost and time to scale up production of chemical quantities needed for safety assessment. We sought to enable early discovery teams to also assess promising molecules for tissue injury liabilities in a short-term 4–7-day rat study using novel tissue transcriptional data in a rapid and resource-sparing manner that might be integrated opportunistically into early animal efficacy studies. The approach was designed to streamline drug development timelines by eliminating molecules with high toxicity potential that may otherwise first present in a 1-month GLP toxicity study, and to open the opportunity for designing additional transcriptional signatures that could inform mechanisms responsible for other longer study duration safety liabilities.

Many researchers have proposed panels of prodromal or predictive gene signatures that may provide a forward looking “risk assessment” following drug treatment (Afshari et al., 2011; Bercu et al., 2010; Qin et al., 2016; Ryan et al., 2008), suggesting mechanistic signals or pathway induction that may ultimately result in downstream toxicity. However, associating toxicological outcome with these prodromal signatures has proven challenging with varying degrees of predictivity by mechanism, perhaps due to the lack of an endpoint to anchor the gene expression changes during identification of the signatures. In contrast, here we propose the use of diagnostic gene expression signatures to directly inform rat tissue injury (degeneration and/or necrosis). The development of diagnostic signatures allows direct correlation of gene expression changes to an observed toxicity endpoint within the same animal and within the same tissue, increasing their probability of success and higher potential performance. These novel diagnostic signatures have the ability to serve as surrogates to detect the presence of tissue lesions early, negate costly and time-intensive histopathologic tissue assessment, and select lead candidates with higher probability of success for later development.

Here, we present the identification, evaluation, and qualification of gene expression toxicity biomarker panels that diagnose tissue degeneration and/or necrosis in rats. These gene expression signatures identify tissue toxicity through common transcriptional regulation pathways induced during tissue damage. These molecular signals of the tissue response to that injury include indicators of tissue repair and regeneration, cell adhesion, and replication, as well as acute inflammatory responses to remove damaged cells. Because these pathways induced by tissue damage are independent of a specific tissue, it is logical to consider common toxicity signatures as more “universal” signatures to diagnose degeneration/necrosis among diverse toxicants across a variety of tissues. A systematic informatic analysis using microarray data from liver, kidney, skeletal muscle, and heart from rats treated with known target tissue toxicants and positive histopathologic outcome resulted in panels of genes that readily diagnose tissue degeneration/necrosis. The performance of these tissue toxicity gene sets is presented demonstrating the high sensitivity and specificity of their response. Additionally, the ultimate utility of these signatures to provide early de-risking of compounds for tissue toxicities with safety lead optimization in mind, in an early rat tolerability study is also presented.

MATERIALS AND METHODS

In vivo animal studies

Rat toxicity studies were designed and conducted with a variety of compounds selected to induce tissue toxicity specific to either liver, kidney, skeletal muscle, heart, pancreas, gastrointestinal tract (GI), bladder, or testes. All studies were reviewed and approved by the Institutional Animal Use and Care Committee of Merck & Co., Inc, Kenilworth, New Jersey, and conducted in an Association for Assessment and Accreditation of Laboratory Animal Care International–accredited facility in compliance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and the Animal Welfare Act.

The compounds dose levels, rat strain, and time points collected are presented in Table 1. In general, the study details were consistent between protocols and were as follows: male and/or female Sprague Dawley CD (SD) or Wistar Han WI (HAN) rats were obtained from Charles River Laboratories, Inc (Raleigh, North Carolina). Initial studies used to identify the gene expression signatures were conducted in SD rats. Following a departmental decision to switch to WI(HAN) for all routine rat study work, bridging studies were conducted to confirm that rat strain difference did not have an impact on the gene expression biomarkers reported here. All subsequent studies were conducted in WI (HAN). The rats were approximately 6–8 weeks of age and weighed from 125 to 325 g at the start of the study. The animals were acclimated for approximately 1 week and randomized into the treatment and control groups. In general, 5 rats were included in each treatment group by oral gavage. During the study, WI(HAN) rats were fed ad lib, and SD animals were maintained on a caloric-optimized diet (Hubert et al., 2000). Doses were calculated based on animal body weight, and the last dose was given 24 h prior to necropsy in studies with daily dosing. All animals were fasted overnight prior to necropsy. Serum biochemical parameters were measured for each animal at necropsy. Approximately 2 ml of blood were collected from the vena cava, and routine clinical chemistry endpoints measured (eg, alanine aminotransferase [ALT], aspartate aminotransferase [AST], blood urea nitrogen [BUN], serum creatinine [sCr]). Additional nonroutine endpoints were also assessed on specific studies (eg, creatine kinase [CK] on skeletal muscle studies, cardiac Troponin I [cTnI] on heart and skeletal muscle studies, or amylase and lipase on pancreas studies).

Tissue samples were collected from each animal at necropsy. Tissue sections were removed and processed for histopathology assessment, and an additional adjacent tissue section removed and placed immediately on dry ice for genomic analyses, allowing for direct comparison of the gene expression response to histopathology outcome. For liver, left lateral lobe was selected, and for the kidney, a section was collected including both the medulla and the cortex. For the skeletal muscle, the quadriceps muscle group was selected as a representative muscle, consisting of 3 muscles (vastus lateralis, vastus medialis, and rectus femoris) comprised predominantly Type II (glycolytic) muscle fibers, and 1 muscle (vastus intermedius) comprised predominantly Type I (oxidative) muscle fibers. Tissue samples were stored at −70°C until RNA extraction.

For histopathology assessment, tissues were immersion fixed in 10% neutral buffered formalin for approximately 24 h, processed and embedded in paraffin blocks, cut into 4–6 micron sections, and stained with hematoxylin and eosin (H&E) as per internal standard operating procedures. Stained tissue sections were examined microscopically by pathologists, and a peer review was performed according to routine laboratory work practice. Histomorphologic changes were qualitatively graded using a severity score scale of 0–5: 0 (no observable pathology), 1 (minimal or very slight), 2 (mild or slight), 3 (moderate), 4 (marked), or 5 (severe). Additional tissues were collected on a subset of studies, including testes, GI (stomach, duodenum, jejunum, ileum, and colon), bladder, and/or pancreas, and processed in a similar manner with one exception: pancreas tissue was harvested first and placed immediately in liquid nitrogen in order to maintain RNA integrity in the pancreas tissue samples. Supplementary Table 1 provides additional summaries of the histopathology outcomes for all studies presented in Table 1.

RNA preparation from frozen tissues

The RNA isolation for all tissues except pancreas was performed using methodologies that eliminates the need for nucleic acid precipitation. Briefly, Trizol reagent (1 ml per 100 mg of tissue) was added to the tissue, and samples homogenized immediately. An aliquot of the homogenate was then transferred to a separate vial and chloroform extracted to remove proteins. The remaining tissue homogenate was stored at −70°C for future use. The supernatant was then transferred to a separate vial, and RNA isolated using Qiagen RNeasy columns or 96-well plates as described by the manufacturer. For pancreas, RNA was isolated using an isopropanol precipitation step rather than the column purification to ensure high quality non-degraded RNA. This was performed after the Trizol and chloroform extraction steps. Following RNA isolation, samples were treated with DNase to remove contaminating DNA, quantified, and appropriate amounts were aliquoted for molecular profiling and quantitative PCR (qPCR) analyses.

Sample processing for microarray analysis

Total RNA was profiled for gene expression changes using the Agilent/Rosetta ink-jet microarray platform as described (Hughes et al., 2001). Briefly, total RNA was reverse transcribed using an oligo-dT primer containing a T7 RNA polymerase promoter site, followed by complementary strand synthesis using random hexamers. In vitro transcription was then performed using RNA polymerase. The RNA was then labeled with 1 of 2 fluorescent dyes, Cy3 and Cy5. Each sample was labeled twice, with either Cy3 or Cy5, to hybridize with appropriate control samples labeled with Cy5 and Cy3, respectively. The reversal of the fluorophores compensates for potential biases due to the differing properties of the dyes or to the normalization process. The fluor-reversed data pairs were combined statistically, and the normalized data was subjected to appropriate quality control metrics (eg, overall array signal, uniformity, internal spiked control signals used to assess array acceptance criteria) (Hughes et al., 2001). The microarray data was then uploaded into a database for gene expression analysis using several data analysis platforms including R, MathWorks Matlab and Rosetta Resolver. The microarray data have been deposited in NCBI’s Gene Expression Omnibus at accession number GSE166229. Pathway enrichment analyses were performed by comparing input sets to GeneGo (www.genego.com), Ingenuity (www.ingenuity.com), and KEGG (www.genome.jp/kegg) pathway sets. Bonferroni-corrected hypergeometric p values (expectation [e] values) of less than .1 were considered significant overlap between sets.

qPCR analyses

Complimentary DNA (cDNA) synthesis was carried out using Reverse Transcription Reagents from Thermo Fisher Scientific (Waltham, Massachusetts). Total RNA was reverse transcribed using random hexamers at a concentration of 5 ng per µl. qPCR was performed with 100 ng cDNA in a 100-µl volume on the Life Technologies sequence detection system. Primer and probe sets were purchased as Assay on Demand Gene Expression products from Life Technologies in a low-density array format. Two different array formats were utilized: single measurements of 96 targets, 4 samples per array = 384 wells total, or duplicate measurements of 24 targets per sample, 8 samples per array = 384 wells total.

Calculations for qPCR analyses

RNA targets were normalized to 18S ribosomal RNA content for liver, kidney, heart, and skeletal muscle samples. The relative amount of each RNA endpoint was calculated from duplicate reactions using the delta C_t method, where C_t is the cycle at which the fluorescence crosses a user-defined threshold. For each sample, the C_t was determined for each target gene and for 18S. The difference between the C_t of the target gene and 18S was defined as the “delta C_t” for that sample. The fold change values for treated samples relative to an untreated control were calculated using the formula 2^^{–(delta Ct treated minus deltaCt untreated)}. All individual animal data were therefore expressed as the fold change relative to the mean value of the concurrent controls.

RESULTS

Rat Study Selection and RNA Transcriptome Profiling for Initial Gene Signature Discovery

Fourteen initial short-term studies in SD rats (4–14 days) were conducted with 3–5 compounds selected for each tissue: liver (lithocholic acid, genipin, thioacetamide), kidney (allopurinol, puromycin, d-serine, tobramycin, hexachloro 1,3 butadiene), skeletal muscle (MSD A [PPARα], MSD B [Antimalarial], cerivastatin), and heart (allylamine, angiotensin II, isoproterenol). Test compounds were typically administered at 2–3 dose levels, with one dose level anticipated to result in injury to the target tissue, and at least one dose level that would not. Tissue injury within each of the 4 tissues was assessed using microscopic histopathologic analyses and defined as degeneration/necrosis for each compound. Other microscopic histologic alterations such as vacuolation or hypertrophy were noted on some studies but not considered as tissue injury in these analyses (see Supplementary Table 1 for additional descriptions of histopathologic outcomes). Tissue samples from dose levels confirmed by histopathology to have degeneration/necrosis in the majority of the animals were selected for RNA transcriptional profiling along with tissue samples from the study concurrent vehicle control animals, and gene expression changes in treated animals determined relative to the control tissue means.

Initially, tissues were analyzed separately, and genes with expression changes associated with degeneration/necrosis across treatment groups within each tissue were identified. 1087, 164, 50, and 49 genes were identified as differentially expressed >2-fold with t test p-value <.0001 (<1 false discovery expected per tissue) in liver, kidney, heart, and skeletal muscle, respectively. Although the numbers of differentially expressed genes in kidney, heart, and skeletal muscle were much smaller than that in liver, the genes identified in these tissues overlapped extensively. For example, of the 230 genes identified in kidney, muscle, or heart, 134 (58%) were differentially expressed by the above criteria in more than one of the 4 tissues. In order to identify additional degeneration/necrosis associated changes that were indicative of the early onset of degeneration/necrosis irrespective of tissue, analysis was performed using all 4 tissues combined. A total of 412 genes were differentially expressed >50% with p < .0001 across all 4 tissues. Combining these 412 genes with the 135 identified in multiple tissues above, a union of 426 common genes were identified and termed as “universal” gene expression biomarkers of drug-induced tissue injury (Figure 1). Greater than 80% of these genes were upregulated across tissues. The upregulated genes were enriched in numerous pathways including in innate and adaptive immune response, cell adhesion and migration, and cell cycle pathways (Supplementary Table 2). Of the 80 genes that were downregulated, the changes occurred predominantly in liver and kidney, and were enriched in pathways associated with lipid and bile acid metabolism/transport (Supplementary Table 3).

Optimization of Gene Sets as Signatures of Drug-Induced Tissue Injury for Each Tissue

The set of 426 genes was down-selected based on magnitude of response, consistency across tissues and consideration of nonredundant biological functional categories to develop a 96-well format qPCR panel for confirmation studies. Specifically, this array contained qPCR probes for 94 of the universal biomarker genes and 2 housekeeper probes for normalization (18S and GAPDH, see Supplementary Table 3). Only upregulated genes were selected for the 96-gene array, due to the more significant increases observed relative to the downregulated genes and more consistent correlation with tissues injury. The same tissue samples from the initial 14 studies used for the microarray profiling (see Figure 1) were assessed on this qPCR platform, and an additional 8 compounds were added to supplement the number of samples analyzed (additional 2 compounds per target organ toxicity). From these data, tissues were analyzed separately using only those with positive histopathology scores, and a final set of 12 prioritized genes was selected for each tissue independently based on statistical significance (t test p-values) and magnitude of response in each tissue, as well as ensuring diverse coverage over biological functional categories. The final gene sets for each tissue are presented in Table 2 and comprised a union of 22 genes of which 16 were selected as top performers across 2 or more tissues. Specifically, 3 genes (Spp1, Gpnmb, and Timp1) were among the 12 top performers in all 4 tissues, 4 genes (Egr2, Olr1, Serpine1, and Tnfrsf12a) were among the top 12 in 3 of the 4 tissues, and 9 other genes were among the top 12 performing tissue toxicity transcripts for 2 of the 4 tissues. We refer to these 16 common genes as “universal” tissue injury genes. All 16 were upregulated with injury, and their functions include multiple roles in the injury response encompassing the 3 major biological functions observed among the initial set of upregulated genes: cell growth/proliferation/differentiation/de-differentiation (Bcl2a1, Cdk1, Egr2, Myc, Runx1, and Tnfrsf12a), cell adhesion/structural/remodeling/chaperone (Anxa2, Gpnmb, Pvr, S100a4, Serpine1, and Timp1), and inflammatory response/inflammatory mediators (Cxcl1, Fcnb, Olr1, and Spp1).

Among the final set of 22 selected gene transcripts (Table 2), 3 transcripts (Spp1, Gpnmb, and Timp1) were among the top performers to warrant inclusion across all 4 tissues. Four gene transcripts (Egr2, Olr1, Serpine1, and Tnfrsf12a) were among the top performing tissue toxicity biomarkers in 3 of the 4 tissues. Nine other transcripts remained among the top performing tissue toxicity transcripts for 2 of the 4 tissues, but all 16 of these core universal genes derived from the list of 100 as described above that were discovered to be highly responsive to tissue toxicants in all 4 tissues.

In addition to these 16 core universal genes in the final gene sets, 6 tissue-selective transcripts were also included in the individual tissue gene sets, including Aurkb, Kim-1 (Havcr1), Clu, and Tff3 for kidney, Reg3b for heart, and Lgmn for skeletal muscle (Table 2). These 6 genes were discovered to be among the most highly regulated for each of the 3 tissues, outperforming members of the 16 core universal tissue injury genes. It should be noted that Tff3 was the only downregulated gene selected as a tissue-specific transcript. This gene was first identified in previous internal analyses as a downregulated marker of kidney degeneration/necrosis, and when compared with the performance of the universal genes, demonstrated superior fold change response and correlation to tissue injury, and was thus included as one of the 12 in the kidney algorithm. Also note that “tissue specific” is used in Table 2 in the context of the 4 tissues profiled for a common injury response to a diverse spectrum of toxicants, but not to imply these gene transcripts are specific to a given tissue. Kim-1 is also highly expressed in T-cells, Reg3b is highly expressed in pancreas, Lgmn is expressed in thyroid and lung, as examples.

Mathematical Algorithm Development for Each Tissue

A logistic model was then utilized to convert the gene expression signatures (fold change values) to a single toxicity score. The algorithms were developed on an individual tissue basis, and coefficients determined empirically. First a summation of the qPCR fold change for the genes for each tissue was calculated using the following equation: X = a[(∑ log10 (fold change gene_1–12)) – b. The “factor” value (a) changes the slope of the line, whereas the “bias” value (b) shifts the zero point. Second, the final toxicity score (π) summation was transformed to scale between 0 and 1 using the equation: π = exp^X/(1 − exp^X). All 12 genes within the algorithms were weighted equally, and their log₁₀ values utilized to minimize the impact of a large fold change increase observed for one or two genes. For the kidney algorithm, the absolute value of the downregulated Tff3 transcript was used in the summation of X.

For each of the 4 tissue algorithms developed, the factor (a) and bias (b) values were determined empirically to maximize sensitivity and specificity performance, using an iterative approach and a toxicity score of ≥ 0.70 considered as positive. The optimal factor (a) selected was a = 0.5 for all 4 tissues, and the bias (b) was b = 6 for liver, kidney, and heart, and b = 7 for skeletal muscle. The training set used for the algorithm development consisted of qPCR data for the initial 14 studies and 8 additional studies evaluated on the qPCR arrays. Normal variability determined in control animals indicated in a cutoff of π  ≥ 0.2 as indicating meaningful change from baseline. Elevations above 0.2 but below the 0.7 positive threshold were seen occasionally in animals treated with tissue toxicants but not exhibiting remarkable histopathologic degeneration/necrosis, suggesting the sensitivity of these gene expression signatures in potentially detecting tissue injury in animals prior to them reaching a level of injury that would be observed microscopically.

Algorithm Performance in Diagnosing Tissue Degeneration/Necrosis

For each tissue, the performance of the gene signatures and algorithm was initially evaluated in studies conducted with Sprague Dawley rats. The performance evaluation set of studies included the reassessment of the 22 studies used in the training set, and an additional 8 studies with paradigm toxicants (2 per tissue), and approximately 100 internal drug candidate dose-limiting toxicity 7-day studies (with a full assessment of tissue pathology and known outcomes; designated MSD J through ZZZ in Table 1) totaling over 130 different compounds (Table 1). Due to the nature of the algorithm, a summation of log10 fold change values from 12 genes all weighted equally, the algorithm did not appear to have bias from overfitting. Further, assessment of the algorithm on the test set shows comparable sensitivity and specificity performance, supporting the lack of bias from overfitting the algorithm. As such, all animals were included in the overall assessment of performance. Approximately 1000 individual animal samples were assessed, which included approximately 20% with positive histopathology (degeneration/necrosis, scores ranging from grade 1 to 4 at approximately equal percentages) in liver, kidney, heart, and skeletal muscle to assess sensitivity (eg, liver toxicants to assess sensitivity of the liver toxicity gene signature), and approximately 80% without histopathology (degeneration/necrosis) in a given tissue used to assess specificity (eg, liver samples assessed from studies conducted with known skeletal muscle toxicants to assess specificity of the liver toxicity gene signature). For these assessments, individual tissue histopathological scores were not considered, but rather binned as either “positive” (any score ≥ 1) or “negative” (score = 0) histopathology.

Table 3 summarizes the animal and compound numbers used in the performance assessment for SD rats (as noted in the Materials and Methods section, original development and qualification of the gene expression signature was performed in SD rats). Following qPCR, the toxicity score was calculated on an individual animal basis, where tox scores that exceeded 0.70 were defined as “positive” and individual animals with a tox score not reaching 0.7 as “negative.” Although considered negative only for the purpose of these performance calculations, as noted above, those scores between 0.2 and 0.7 most likely are indicative of some underlying events and a more ambiguous than negative conclusion should be considered. Using histopathology as truth, sensitivity and specificity were calculated on an individual animal and compound level. At an individual animal level, sensitivity is defined as the number of samples with a positive gene expression toxicity score (greater than or equal to 0.70) that are also positive by microscopic histopathology (degeneration/necrosis) divided by the total number of samples with positive toxicity by histopathology. Specificity is defined as the number of samples with a negative gene expression toxicity score (less than 0.70) that did not have histopathology findings divided by the total number of samples without histopathology. At the compound and dose level, positive compounds/dose levels were defined as those inducing tissue degeneration/necrosis within the given tissues, and if any animals treated with that compound at that test dose scored above 0.7. The summary statistics from this performance evaluation are presented in Table 4. For liver, kidney, and skeletal muscle, the overall algorithm had approximately 80% sensitivity and approximately 95% specificity on an individual animal basis and had approximately 90%–100% sensitivity and approximately 95% specificity at the compound/dose level. Heart was an exception, where the individual animal sensitivity was only 54% although specificity was 99%. This was hypothesized to be due to the fact that the tissue samples collected for histopathology and for RNA isolation were from different regions, and some of the compounds assessed induced focal injury within a certain location within the heart. Therefore, to supplement the gene expression performance, assessing cTnI in plasma was added and this greatly enhanced the performance of detecting drug-induced heart injury, with a combined performance (toxicity gene expression signatures plus plasma cTnI) of approximately 75% sensitivity at the individual animal level (using cTnI alone resulted in only approximately 55% sensitivity). It should be noted that the overall performance of the gene expression signatures in detecting tissue degeneration/necrosis in liver, skeletal muscle, or kidney was considerably higher than that of other serum chemistry parameters (data not shown), mainly due to the lack of specificity for ALT and AST (liver and skeletal muscle), or lack of sensitivity for CK, BUN, and sCr (skeletal muscle and kidney). Therefore, supplementing liver, skeletal muscle, or kidney with serum chemistry endpoints did not add additional value.

An example of gene expression changes in liver, and the resulting toxicity scores and histopathology outcome are presented in Figure 2, demonstrating the performance of the algorithm genes and toxicity scores in diagnosing degeneration/necrosis. Of note in Figure 2 are those animals with significant gene expression changes and a positive toxicity score that associate with positive histopathology findings. Conversely, those treated animals that did not have a positive histopathology score or only had liver hypertrophy did not produce appreciable gene expression changes and therefore had low toxicity scores.

Following an internal organizational decision to switch to the Wistar rat strain for all routine rat toxicity study work, 13 bridging studies were conducted in Wistar to confirm that strain difference or differing diets (eg, caloric-optimized diet for SD and ad lib for Wistar) did not have an impact on the gene expression biomarkers originally identified and qualified in SD rats. Specifically, these studies examined if the individual genes used in the algorithms and the final algorithm scores performed consistently across these different rat strains. These additional studies were conducted using a core set of compounds shown previously to induce toxicity in 1 of the 4 tissues in SD rats (Table 1; compounds in italics and designated as Wistar Bridge). Comparable numbers of individual animal samples were included in this assessment, qPCR assays were performed, toxicity scores calculated using the previously developed algorithms in SD, and performance compared with histopathology. A summary of animal numbers and algorithm performance is presented in Table 5. The performance was similar and perhaps slightly better in Wistar than observed in SD, with approximately 80%90% sensitivity and 100% specificity for liver, kidney, skeletal and muscle, and approximately 60% sensitivity and 100% specificity for heart at the individual animal level. This observed similar or slightly better sensitivity and specificity performance in Wistar, an independent sample “test” set, further supports the use of the algorithm without concerns of bias or overfitting. As with SD rats, supplementing with plasma cTnI increased the sensitivity for heart to approximately 80%. Performance at a compound/dose level was 100% for sensitivity and specificity for Wistar (not presented in Table 5), but it should be noted that only a limited number of compounds per target tissues were assessed (3–5 compounds per tissue, 13 studies total).

Gene Signature Expansion to Other Tissues

Based on the universal nature of the tissue degeneration/necrosis gene signatures, studies targeting other tissue toxicities were conducted and algorithms developed. These tissues included pancreas, GI (including sections of stomach, duodenum, jejunum, ileum and colon), bladder, and testes, where drug-induced toxicities are observed less frequently in nonclinical drug development, but are nevertheless observed on occasion or may be of concern for a specific drug development program (known target expression in these tissue or based on additional sources of knowledge identifying liabilities in these tissues). For gene expression analyses of the GI, 5 individual segments were originally collected and each assessed separately (stomach, duodenum, jejunum, ileum, and colon). The duodenum became the focus of our assessment because the majority of degeneration/necrosis scores and most robust gene expression changes where observed in this region. Fifteen studies in total were used in this expansion to other tissues, 5 for GI (duodenum), 4 for pancreas, and 3 each for both bladder and testes (see Table 1). The number of studies for these tissues are far fewer than the number of the qualification studies conducted for liver, kidney, heart, and skeletal muscle, and more comparable to the 3–5 studies per tissue described here in our initial signature discovery efforts for liver, kidney, heart, and skeletal muscle. Relying upon just the core universal 16-gene transcripts, and leveraging a similar 12-gene optimized algorithm for each additional tissue, the following 12 genes were selected for the algorithms in these additional tissues: Pancreas = Bcl2a1, Cdk1, Cxcl1, Egr2, Fcnb, Gpmnb, Runx1, S100a4, Serpine1, Spp1, Timp1, Tnfrsf12a; GI (duodenum) = Anxa2, Bcl2a1, Cxcl1, Egr2, Fcnb, Gpmnb, Pvr, Runx1, S100a4, Serpine1, Timp1, Tnfrsf12a; Bladder = Anxa2, Bcl2a1, Cdk1, Cxcl1, Egr2, Fcnb, Gpmnb, S100a4, Serpine1, Spp1, Timp1, Tnfrsf12a; Testes = Anxa2, Bcl2a1, Cxcl1, Egr2, Gpmnb, Myc, Pvr, Runx1, S100a4, Serpine1, Timp1, Tnfrsf12a. Here we observed approximately 85%–90% sensitivity and ≥95% specificity for bladder and pancreas, and approximately 65%–70% sensitivity and ≥95% specificity for GI and testis (Table 6) on an individual animal level. As was reasoned above for lower sensitivity observed in heart, the two tissues with the lower sensitivity values (duodenum and testes) may be due to more focal drug-induced toxicity (eg, limited to only one region) and RNA isolation/assessment from different regions than those processed for microscopic evaluation. Additionally, gene expression signals from lower grade focal histopathological lesions may be diluted when isolating RNA from larger sections of these tissues. When drug injury is suspected in these tissues, investigations could include multiple smaller sample collections, processing, and analyses to cover a larger section of the tissue and increase the signal with these multiple smaller samples.

DISCUSSION

Here we describe the discovery, development, evaluation, qualification, and implementation of gene expression signatures to diagnose acute drug-induced tissue toxicity (degeneration/necrosis) in rats using tissues anchored with microscopic histopathologic diagnoses of tissue degeneration/necrosis. Gene expression signatures can provide an assessment of tissue toxicity without the need for tissue processing and microscopic analyses of histopathology. These signatures were initially developed in liver, kidney, skeletal muscle, and heart, the tissues that comprise approximately 85% of the tissue toxicities observed in preclinical rat toxicity studies. Noting that a common or “Universal” gene signature for degeneration/necrosis was tissue independent provided the foundation for a single qPCR platform to monitor different tissues, as well as allowing for expansion of gene sets to other tissues (eg, pancreas, GI, bladder, and testes). Figure 3 provides a detailed schematic summary of the development process for these gene expression toxicity biomarker panels, for both SD and Wistar and for all tissues examined.

The resultant composite biomarkers consist of genes that were intentionally selected based both on performance and representation from a diverse set of biological responses to degeneration/necrosis. Specifically, the “Universal” genes broadly represent immune response, tissue regeneration and remodeling in response to degeneration/necrosis. It is worth noting that cell death processes including necrosis and apoptosis were not the predominantly enriched pathways among the set of 426 most robust necrosis/degeneration associated biomarker genes (Figure 1). For example, these processes were not among the top 100 enriched GeneGo/Ingenuity/KEGG pathways (see Supplementary Table 2), or in a similar assessment of GO Ontology terms (geneontology.org). This is not unexpected in bulk tissue expression data and is consistent with our findings from profiling the transcriptional effects across hundreds of compounds in rat liver (Monroe et al., 2020; Podtelezhnikov et al., 2020). First, relative to the bulk tissue, there are a small number of cells actively undergoing cell death at any one point in time, and these signals are thus difficult to detect when averaged across all cells in the tissue. Second, the infiltration of immune cells in response to tissue damage results in dramatic fold increases in immune cell-specific transcripts that dominate the genomic profiles. Indeed, we found that the majority of the most robustly regulated transcripts universally associated with degeneration/necrosis where involved in immune response and/or regenerative processes. Nevertheless, two of the final universal genes (Myc and Tnfrsf12a) and the kidney biomarker gene Clu are involved in promoting cell death processes and could reflect signals from necrotic cells. Future studies using single-cell sequencing technologies will enable assessment of the specific cell types responsible for driving the universal gene expression response biomarkers, and the resolution of more subtle transcriptional responses within resident cell types of degenerating/necrotic tissue.

Upon the successful development of gene expression signatures to monitor drug-induced tissue toxicity with high sensitivity and specificity, a rat study paradigm was implemented as an early safety lead-optimization (SLO) model to de-risk new structural motifs, and later in lead optimization to help ensure selection of more optimal pre-clinical drug candidates. These rat exploratory studies consisted of a 4-day dosing regimen with 4 male rats per group and 2 dose levels. Generally, a low dose was selected to provide exposures approximately 5-fold over targeted clinical exposures, and a high dose of 600–750 mg/kg/day that could help define dose limiting toxicities and inform a maximum tolerated dose for subsequent studies. This paradigm had a low compound requirement (approximately 3.5 g) and can assess one or more compounds per study. This study design was evaluated and compared with routine 7-day dosing exploratory dose-limiting toxicity studies using several model toxicants, ensuring that induced toxicities manifested by day 5 and animal numbers were sufficient to detect tissue injuries, while allowing for a wide range of exposures to be assessed in a 5-day study that can be completed within 1 week. The diagnostic gene expression signatures are assessed as a surrogate to histopathology, providing a rapid determination of tissue toxicity. This allows development teams to quickly determine compound liabilities without the necessity of performing time-consuming histopathology and before committing both time and resources to scale up compound syntheses before final candidate selection and prior to entry to GLP phases of preclinical development in both rats and non-rodent species. These tissue toxicity transcriptional endpoints can also be assessed in other tissues (eg, pancreas, GI, bladder, testes), and other additional endpoints included on the studies as needed such as hematology, circulating inflammatory biomarkers, and bone marrow micronucleus assessment to further capitalize on this early SLO model to provide discovery teams added information to identify compound liabilities. These endpoints may be included when a target has a potential toxicity in those tissues or other lead compounds from the program have demonstrated such liabilities in preclinical development toxicity studies. We have also noted case examples where tissue injury algorithm calculations lie between a score of 0.2–0.7 but animals in that dose group do not cross the 0.7 threshold value for positive. In these situations, the individual genes within the algorithm are modulated similarly to what is observed in animals with scores > 0.7, however, collectively do not meet the fold-change values required to exceed the positive threshold. Generally, but not always, we have observed that rats treated with higher exposures and/or longer dosing periods are likely to test positive (data not shown), suggesting that toxicity scores within the range are still informative and can add value in understanding the potential liabilities for drug candidates.

This SLO study paradigm initially developed to identify molecules early with an unequivocal tissue injury liability also offers opportunity for gene expression panel expansion. We have described elsewhere (Monroe et al., 2020; Podtelezhnikov et al., 2020; Qin et al., 2019) additional gene expression signatures that have been developed, which may inform on the high potential for activating key molecular initiating mechanisms of more insidious toxicities that do not present within the first week of dosing. These include the potential for drug-induced liver injury and rodent carcinogenicity mechanisms. The wide utility of such an early de-risking in vivo model provides not only lead candidate selection, but also provides insight into dose selection for subsequent toxicity studies. Since the implementation of the SLO de-risking model, this paradigm has contributed to the significant reduction in compound attrition due to animal toxicity seen within our organization (Sistare et al., 2018).

SUPPLEMENTARY DATA

Supplementary data are available at Toxicological Sciences online.

ACKNOWLEDGMENTS

The authors thank colleagues in MRL Safety Assessment and Laboratory Animal Resources for their support in conducting studies and providing data interpretation, and to Dr Jose Lebron and Katerina Vlasakova for critical review of the manuscript.

DECLARATION OF CONFLICTING INTERESTS

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

REFERENCES