Nephrotoxicity Development of a Clinical Decision Support System Based on Tree-Based Machine Learning Methods to Detect Diagnostic Biomarkers from Genomic Data in Methotrexate-Induced Rats

Abstract

Background: The purpose of this study was to carry out the bioinformatic analysis of lncRNA data obtained from the genomic analysis of kidney tissue samples taken from rats with nephrotoxicity induced by methotrexate (MTX) and from rats without pathology and modeling with the tree-based machine learning method. Another aim of the study was to identify potential biomarkers for the diagnosis of nephrotoxicity and to provide a better understanding of the nephrotoxicity formation process by providing the interpretability of the model with explainable artificial intelligence methods as a result of the modeling. Methods: To identify potential indicators of drug-induced nephrotoxicity, 20 female Wistar albino rats were separated into two groups: MTX-treated and the control. Kidney tissue samples were collected from the rats, and genomic, histological, and immunohistochemical analyses were performed. The dataset obtained as a result of genomic analysis was modeled with random forest (RF), a tree-based method. Modeling results were evaluated with sensitivity (Se), specificity (Sp), balanced accuracy (B-Acc), negative predictive value (Npv), accuracy (Acc), positive predictive value (Ppv), and F1-score performance metrics. The local interpretable model-agnostic annotations (LIME) method was used to determine the lncRNAs that could be biomarkers for nephrotoxicity by providing the interpretability of the RF model. Results: The outcomes of the histological and immunohistochemical analyses conducted in the study support the conclusion that MTX use caused kidney injury. According to the results of the bioinformatics analysis, 52 lncRNAs showed different expressions in the groups. As a result of modeling with RF for lncRNAs selected with Boruta variable selection, the B-Acc, Acc, Sp, Se, Npv, Ppv, and F1-score were 88.9%, 90%, 90.9%, 88.9%, 90.9%, 88.9%, and 88.9%, respectively. lncRNAs with id rnaXR_591534.3 rnaXR_005503408.1, rnaXR_005495645.1, rnaXR_001839007.2, rnaXR_005492056.1, and rna_XR_005492522.1. The lncRNAs with the highest variable importance values produced from RF modeling can be used as nephrotoxicity biomarker candidates. Furthermore, according to the LIME results, the high level of lncRNAs with id rnaXR_591534.3 and rnaXR_005503408.1 particularly increased the possibility of nephrotoxicity. Conclusions: With the possible biomarkers resulting from the analyses in this study, it can be ensured that the procedures for the diagnosis of drug-induced nephrotoxicity can be carried out easily, quickly, and effectively.

1. Introduction

Kidneys play an important role in many vital tasks such as maintaining water and electrolyte balance in our body, regulating acid–base balance, regulating blood pressure with the renin they secrete, regulating the body’s production of erythropoietin with erythrocyte, activating vitamin D3, and removing drugs and toxins from the body [1]. Kidneys are the target organs for many drugs and toxic substances, especially due to their high metabolic activity, the removal of harmful substances from the body, and active transport functions [2]. Nephrotoxicity is the condition in which the structure of the kidney is damaged by chemicals and decreases by affecting the kidney function negatively. In the case of drug-induced nephrotoxicity, acute kidney injury occurs with cell death due to apoptosis and necrosis, and may lead to death due to kidney failure. Therefore, drug-induced nephrotoxicity may be the primary cause of morbidity and mortality [3]. Today, the use of chemotherapeutic agents has increased considerably due to the increase in cancer cases. MTX, a chemotherapeutic agent, is widely used in leukemia, osteosarcoma, lymphoma, head and neck tumors, breast cancer, lung cancer, and some other cancer types [4]. MTX and its metabolites are largely eliminated by the kidney and enter cells by active transport. In the mechanism of MTX-induced kidney toxicity, the direct toxic effect of MTX, the inhibition of enzymes related to DNA synthesis, and an increase in the production of free oxygen radicals (ROS) can be seen. ROS are directly or indirectly involved in a wide variety of clinical disorders such as atherosclerosis, viral infection, reperfusion injury, macular degeneration, pulmonary toxicity, cataractogenesis, diabetes, cancer, and toxic cell damage [5,6]. Clinical and toxicological assessment of kidney function routinely relies on the measurement of blood urea nitrogen and serum creatinine, but their relatively low sensitivity often precludes the early detection of kidney injury. Therefore, it is important to identify new sensitive and reliable biomarkers of renal nephrotoxicity. In addition, the development of new biomarkers that can accurately detect drug-induced kidney damage is needed for both drug development studies and for the clarification and treatment of the mechanism of this type of kidney damage [7]. One way to discover potential biomarkers is to use omics data [8]. Recently, genomic data have been used a lot in the determination of diseases, and these data have had a very important effect on the creation of personal profiles by examining the diseases on the basis of genes, and in the regulation of personal treatment and side effects. The detection of drug-induced gene changes is of critical importance for the detection of drug-induced kidney damage [9]. With the increasing use of both experimental and computational methods in RNA-Seq technologies, the number of long non-coding RNAs (lncRNA) has increased greatly in the last few years [10,11]. It has been reported that lncRNAs are associated with kidney diseases such as acute kidney disease, chronic kidney disease, and kidney transplantation [12,13]. However, although there are genomic studies for drug-related nephrotoxicity, studies related to lncRNA are not frequently encountered. For this reason, there is a need for such studies in the literature to eliminate the deficiencies in this area.

Machine learning (ML) methods, which have recently been used to a large extent in the health field, help researchers in the early prediction, diagnosis, prognosis, and individual patient care decision-making of various diseases and other medical disorders [14,15]. In addition, in recent years, ML has contributed to the literature in identifying possible biomarkers for many disease states such as cancer [16]. Some methods were needed in order to make the results obtained from modeling ML methods more interpretable and explainable. Based on these requirements, the concept of explicable artificial intelligence (XAI) has been introduced. The use of classification models to diagnose disease in the field of health largely depends on the ability of the models to be interpreted and explained by the researcher [17,18]. The XAI methods used for this purpose provide a patient-specific explanation for a particular classification, thus allowing for a more understandable explanation of any complex classifier in the clinical [19].

This study aimed to model the lncRNA data of kidney tissues taken from rats without pathology and treated with methotrexate with ML methods and to determine possible biomarkers for the early diagnosis of nephrotoxicity by providing the interpretability of the model with XAI methods as a result of the modeling.

2. Materials and Methods

2.1. Dataset

In order to discover probable biomarkers underlying drug-induced nephrotoxicity and classify nephrotoxicity at the clinical level, 20 female Wistar albino rats (weight: 250 ± 20 g; age: 3 months;) were acquired from the Inonu University Experimental Animal Production and Research Center.

• Control group (MK): This group was injected intraperitoneally with physiological saline as a carrier solvent on the first day of the experiment.
• MTX-treated group (M): This group was given a single dose of 20 mg/kg MTX intraperitoneal on the first day of the experiment.

On the fourth day of the experiment, xylazine (24 mg/kg intraperitoneal) and ketamine (225 mg/kg intraperitoneal) were given to the rats under high-dose anesthesia, and kidney tissue samples of the rats were taken. Genomic, histopathological, and immunohistochemical analyses were performed.

2.2. Random Forest Method

The RF technique, proposed by Breiman in 2001, is a machine learning algorithm with several decision trees that combines the bagging and random subspaces methodologies. In the RF algorithm, it is a supervised machine learning algorithm in which calculations of multiple decision trees are combined to produce a final result. Its ease of use and flexibility have accelerated its adoption as it addresses both classification and regression problems [20]. In the RF algorithm, the dataset is first randomly divided into two sections in the RF algorithm: the training data for learning and the validation data for assessing the learning level. Following this, multiple decision trees are randomly generated from the dataset using the “bootstrap method”. The branching of each tree is determined by randomly picked determinants at node positions. The RF outcome estimate is the mean of all the tree’s outcomes. As a result, each tree effects the RF estimation for certain weights. The RF method outperforms other machine learning algorithms because of its capacity to accept training data from subsets at random and generate trees using random approaches. Furthermore, because training is performed on numerous randomly selected sub-datasets via bootstrap sampling, the RF technique minimizes overfitting [21,22].

2.3. Data Analysis and Modeling Tasks

The Shapiro–Wilk test was used to assess the conformance of quantitative data to normal distribution. Non-normal distribution data were presented using the median (minimum-maximum), whereas the normal distribution data were summarized using the mean ± standard deviation. The Mann–Whitney U test was used to compare non-normally distributed data, while the independent sample t-test was used to examine normally distributed data. All analyses were performed using IBM SPSS Statistics 26.0 for Windows (New York, NY, USA). The TMM (trimmed mean of M values) normalization method was employed for the relevant data. In bioinformatics analysis, the false discovery rate (FDR) is utilized to make evaluations.

Within the scope of the investigation, the Boruta technique was applied as the variable selection method. Python programming language was used for the application of a tree-based model planned for use within the scope of the study and for explainable artificial intelligence modeling afterward. In this study, the model performance was evaluated using Se, Sp, B-Acc, Npv, Acc, Ppv, and F1-score metrics. Furthermore, the images used in the visualizations were developed using the R programming language and Excel software.

2.4. Histopathological and Immunohistochemical Analyses

2.4.1. Histopathological Analyses

Rat kidney tissues were divided into tiny fragments of 3–4 mm for histological analysis. Following that, plastic tissue was put in follow-up cassettes and preserved for 24 h in 10% formaldehyde. The tissues were rinsed in running tap water for 24 h after fixation. These were rendered transparent in xylene, dehydrated in various grades of alcohol, and then embedded in paraffin. A Leica RM2145 microtome was used to cut sections from paraffin blocks that were 5 microns thick. To study the overall histological structure, the slices were stained using the hematoxylin and eosin (H&E) procedure. The H&E combination is one of the most frequently used dye combinations in routine histopathological analysis. Hematoxylin is dark blue-violet and stains nucleic acids in a reaction that is still not fully understood. Eosin is pink and stains proteins non-specifically. In a typical tissue, the nuclei are blue-black, and the cytoplasm and extracellular substance are stained in different shades of pink. We can evaluate nuclear and cytoplasmic changes in the cell with H&E staining. Nuclear pyknosis and eosinophilic cytoplasm are indicative of necrotic cells [23]. Renal damage was evaluated in terms of peritubular infiltration, vacuolization of the tubular epithelial cells, shedding, and necrosis. Ten areas were examined at X20 magnification from each section, and histopathological scoring was determined according to the degree and extent of renal damage. According to the severity of the damage, it was rated as 0 (no change), 1 (mild), 2 (moderate), and 3 (severe) [24]. The preparations were examined with a Leica DFC280 light microscope and the Leica Q (Leica Micros Imaging Solution Ltd., Cambridge, UK) image analysis system, scored, and photographed.

2.4.2. Immunohistochemical Analyses

Immunohistochemical staining with the Cystatin C antibody was used to observe tubule damage in the kidney sections. For immunohistochemistry analyses, sections that were deparaffinized and rehydrated were placed in a 2100 Antigen Retriever incubator and boiled in 0.01 M citrate (pH 6.0) for 15 to 20 min. The sections were exposed to 3% hydrogen peroxide for 12 min in order to inhibit the endogenous peroxidase enzyme activity. After washing the sections with phosphate buffer saline (PBS), a protein block (ultra V block) was applied for 5 min. After that, the sections were exposed to the primary antibody for 60 min at 37 °C. The tissues were treated with biotin-based secondary antibodies for 10 min at 37 °C after being rinsed with PBS. Following this process, the slices were treated with streptavidin peroxidase for 10 min at 37 °C. Following hematoxylin staining, the slices with chromogen applied were covered with water-based concealer. Semi-quantitative scoring was used to determine the staining immunoreactivity prevalence (0: 0–25%, 1: 26–50%, 2: 51–75%, and 3: 76–100%) and severity (0: none, +1: mild, +2: moderate, +3: severe) [25].

2.5. Genomic Analyses

2.5.1. Total RNA Isolation and Quality Control from Harvested Tissues

Total RNA was isolated from kidney tissue samples using kits that allow for high-efficiency isolation, even with low-volume samples. The miRNeasy Serum/Plasma Kit (Qiagen, Cat. No./ID: 217184) was developed to purify cell-free total RNA, namely miRNA and other small RNA, from very tiny amounts of serum and plasma. Qubit (Life Technologies, Carlsbad, CA, USA) was used to fluorometrically quantify the amount of RNAs collected. The RNAs were verified for quality using a bioanalyzer before sequencing. RNA integrity number (RIN) ≥7 samples were sequenced with the control.

2.5.2. Preparing and Sequencing NGS Libraries for lncRNA Sequences

The “TruSeq Stranded Total RNA Library Prep Kit” from the Illumina corporation was used to create the sequencing library for lncRNA sequences under the following circumstances:

Ribosomal RNAs (rRNAs) were isolated from the total RNA, and the remaining RNAs were purified and fragmented. The bioanalyzer was used to verify the elimination of rRNA. First-strand cDNA was created by reverse transcription of the remaining RNA fragments using random hexamer sequences. The RNA template was then removed, and the second strand of cDNA (blunt ds cDNA) was synthesized [26]. To prevent the fragments from attaching to one other during the adaptor ligation procedure, a single ‘A’ nucleotide was inserted into the 3′ ends of the blunt ds cDNAs. To hybridize the ds cDNA fragments to the flow cell surface, indexing adapters were introduced. Finally, DNA fragments were enriched, and sample libraries were standardized and pooled. The samples were read using the paired-end (PE) 2 × 150 Illumina NovaSeq 6000 platform with 50M reads as the baseline [27].

3. Results

3.1. Histopathological Results

In the kidney sections of the control group stained with the hematoxylin and eosin staining method, the outer leaf of the Bowman capsule in the renal corpuscle and the glomerular tuft within it had a normal appearance. The macula densa formed by changing the morphology of the distal tubule cells approaching the vascular pole of the renal corpuscle was observed as normal. The Bowman distance between the outer leaf of the Bowman’s capsule and the inner leaf surrounding the glomerular tuft was of normal width.

The epithelial cells surrounding the lumen of the proximal tubule around the renal corpuscles in the cortex were normal in appearance with round and central nuclei and acidophilic cytoplasm. The inner lumen borders were not very clear due to the microvilli located at the apical part of the cells. Distal tubular epithelial cells were easily distinguished from the proximal tubules by their paler staining and wider lumens (Figure 1a). In the MTX-treated group, in the preparations examined by the hematoxylin and eosin staining method, prominent areas of inflammation were observed in the intertubular regions of the cortex (Figure 1b).

In the sections, epithelial cells of some tubules were observed to spill into the lumen (Figure 2a). In this group, vacuolization was also detected in the cytoplasm of some tubule cells (Figure 2b).

Another remarkable finding was the presence of necrotic cells with eosinophilic cytoplasm and dark nuclei in some tubules (Figure 3).

3.2. Immunohistochemical Results

In the control group, positive tubule cells were not found in the sections where the Cystatin C immunohistochemical staining method was applied (Figure 4a). In the MTX-treated group, the density of positively stained tubules in the sections using the Cystatin C immunohistochemical staining method was observed to increase compared to the control group (Figure 4b).

The descriptive statistics for the rats used in the experiment are shown in

Table 1. Descriptive statistics for the rats utilized in the investigation.

Variables	Mean ± Standard Deviation
Rat weight (baseline) (g)	249.15 ± 22.32
Rat weight (endpoint) (g)	252.1 ± 24.05
Kidney weight (g)	0.968 ± 0.1

.

Table 2. Descriptive statistics for the MTX-treatment and control groups.

Variables	Control	MTX Treatment
Rat weight (baseline) (g)	245.3 ± 24.02	253 ± 21.01
Rat weight (endpoint) (g)	252 ± 24.03	252.2 ± 25.37
Kidney weight (g)	0.97 ± 0.08	0.96 ± 0.12

provides the descriptive statistics for the MTX-treatment and control groups.

3.3. Differential Expression Results

There were 16.386 expressions in the dataset that was used for the investigation. The bioinformatics study found that 52 lncRNAs expressed differently in the groups (FDR < 0.05). A total of 35 of them displayed up-expression (logFC > 1), while 17 displayed down-expression (logFC < −1). A presentation of the dataset can be found in the Supplementary Materials.

The distribution of the samples was found to be consistent in terms of lncRNA expression levels in the comparison of the MTX-treated group (M) and the control group (MK) based on the principal component (PCO) analysis. The controls and application examples showed some unity among themselves. However, although this distinction was not sharp, a distinction emerged due to the lower number of lncRNAs showing the total expression changes. When the MTX-treated group and control group samples were compared individually with each other, it was determined that more lncRNAs were exposed to expression level changes in the M2 and MK2 samples. Figure 5 depicts a graphical representation of the PCO analysis.

Figure 6 shows a heatmap representation of the 50 lncRNA expressions with the highest variation in the expression level comparison.

Overexpressed lncRNAs are indicated in red and suppressed lncRNAs are shown in green for the 50 lncRNAs that exhibited the highest variation in the M versus MK comparison. When compared to the control, the application samples had different expression patterns. However, it was determined that some samples (such as M-10, MK-4, and MK-2) deviated from the application and control groups.

Figure 7 shows the volcano plot used to visualize the differentially expressed genes. Figure 7 shows that the red and blue lncRNAs represent the upregulated and downregulated lncRNAs, respectively. The lncRNAs in black are those that did not differ in expression between the two groups.

3.4. Biostatistics Analysis and Modeling Results

The TMM (trimmed mean of M values) normalization approach was used to extract data from 16.386 lncRNAs in the dataset. In the study, 31 lncRNAs that may be associated with the disease state were selected using the Boruta variable selection method, one of the variable selection methods from lncRNAs that show different regulations (up and down), in order to reveal lncRNAs that may be associated with the disease state.

Table 3. Detailed information about the data analysis results.

Gene Name	Chromosome	ID	Group				LogFC	p
			M		MK
			Mean ± SD	Median (Min–Max)	Mean ± SD	Median (Min–Max)
LOC102555118	NC_051337.1	rna-XR_351582.4	226.4 ± 116.41	248 (42–447)	35.6 ± 16.85	34 (17–66)	1.616	0.001 *
LOC106736471	NC_051345.1	rna-NR_133655.1	102.4 ± 76.73	88 (5–257)	11.9 ± 9.48	10.5 (1–34)	2.198	0.005 *
LOC103691349	NC_051336.1	rna-XR_590665.2	281.2 ± 123.78	294 (49–470)	68.4 ± 78.82	46 (26–290)	1.247	0.001 **
LOC108351528	NC_051342.1	rna-XR_001839007.2	454.2 ± 191.95	486.5 (96–661)	117.1 ± 118.09	80.5 (55–449)	1.118	0.001 **
LOC120098801	NC_051336.1	rna-XR_005497310.1	166 ± 92.76	164.5 (29–370)	38.6 ± 36.34	26.5 (13–139)	1.187	0.001 **
LOC120094778	NC_051344.1	rna-XR_005489439.1	140 ± 90.62	125 (28–296)	28.9 ± 14.03	30 (9–51)	1.248	0.004 *
LOC120099280	NC_051336.1	rna-XR_005498350.1	109.6 ± 68.4	96 (13–206)	21.6 ± 21.84	15.5 (7–82)	1.488	0.002 **
LOC120096007	NC_051347.1	rna-XR_005492056.1	134.4 ± 91.26	123.5 (17–332)	32.6 ± 28.96	25.5 (6–111)	1.087	0.004 **
LOC120098788	NC_051336.1	rna-XR_005497230.1	27.3 ± 13.61	29.5 (3–52)	4.6 ± 4.62	2.5 (0–15)	1.751	<0.001 **
LOC120098190	NC_051353.1	rna-XR_005496257.1	85.5 ± 54.7	70 (9–172)	19.5 ± 18.58	16 (4–70)	1.277	0.004 **
LOC108348888	NC_051354.1	rna-XR_005496888.1	71.2 ± 32.64	75.5 (12–112)	17.1 ± 20.59	11.5 (3–74)	1.250	0.002 **
LOC103691816	NC_051338.1	rna-XR_591534.3	210.4 ± 116.14	230.5 (54–421)	49.2 ± 36.54	40.5 (19–147)	1.171	0.001 **
LOC120098816	NC_051355.1	rna-XR_005497370.1	220.6 ± 173.89	186 (48–552)	31.3 ± 22.35	30 (6–73)	1.992	0.007 *
LOC120096731	NC_051349.1	rna-XR_005493563.1	6.6 ± 6.64	3.5 (0–18)	13.2 ± 14.34	8 (3–51)	−1.862	0.093 **
LOC120098521	NC_051354.1	rna-XR_005496784.1	362.1 ± 181.28	349.5 (74–587)	88.8 ± 100.42	58 (33–369)	1.249	0.001 **
LOC120102202	NC_051339.1	rna-XR_005503371.1	84.1 ± 63.57	73 (13–208)	15.6 ± 11.47	13.5 (3–37)	1.559	0.008 *
LOC102549457	NC_051346.1	rna-XR_358189.4	77.7 ± 42.9	75.5 (8–154)	21.1 ± 26.93	12.5 (4–96)	1.078	0.007 **
LOC120102261	NC_051339.1	rna-XR_005503535.1	215.2 ± 138.84	176.5 (16–442)	47 ± 27.2	38.5 (26–116)	1.205	0.003 **
LOC120100781	NC_051337.1	rna-XR_005500805.1	51.1 ± 23.38	49 (11–82)	14.4 ± 20.5	8 (2–71)	1.114	0.002 **
LOC108348808	NC_051353.1	rna-XR_005496283.1	42.2 ± 26.1	37.5 (5–84)	9.1 ± 5.61	9 (2–19)	1.287	0.003 *
LOC103691306	NC_051336.1	rna-XR_005499594.1	6.2 ± 4.47	5.5 (0–12)	0.6 ± 0.52	1 (0–1)	2.178	0.001 **
LOC102552040	NC_051344.1	rna-XR_001839839.2	3.8 ± 4.49	3 (0–15)	0.1 ± 0.32	0 (0–1)	3.296	0.002 **
LOC120099889		rna-XR_005499330.1	282.2 ± 232.78	197 (41–831)	68.2 ± 86.7	35 (24–308)	1.431	0.002 **
	NC_051336.1
LOC120099800	NC_051336.1	rna-XR_005499033.1	53.7 ± 33.95	45 (5–102)	14.2 ± 19.85	9.5 (1–69)	1.176	0.004 **
LOC120097836	NC_051352.1	rna-XR_005495645.1	32.8 ± 16.73	28.5 (13–62)	7.7 ± 4.32	8.5 (1–14)	1.089	0.001 *
LOC120102212	NC_051339.1	rna-XR_005503408.1	18.5 ± 10.54	14.5 (8–42)	4 ± 2.31	3.5 (2–10)	1.313	<0.001 **
LOC102555751	NC_051355.1	rna-XR_005497840.1	54.9 ± 45.9	41 (1–162)	12.1 ± 12.54	8.5 (3–47)	1.431	0.008 **
LOC120102327	NC_051339.1	rna-XR_005503688.1	50.7 ± 46.08	41.5 (1–165)	9.8 ± 8.04	7.5 (3–30)	1.612	0.005 **
LOC120099962	NC_051336.1	rna-XR_005499541.1	1 ± 0.94	1 (0–3)	2.3 ± 0.82	2 (1–4)	2.047	0.005 **
LOC108352129	NC_051345.1	rna-XR_001840278.2	26 ± 18.34	21 (0–59)	5.8 ± 6.94	3 (2–25)	1.282	0.008 **
LOC102554372	NC_051339.1	rna-XR_353438.4	48.4 ± 27.61	49.5 (3–84)	12.1 ± 6.05	11.5 (4–21)	1.037	0.002 *

*: Independent sample t-test; **: Mann–Whitney U test; LogFC: Log fold change; M: MTX-treated group; MK: control group.

contains the selected expressions and dataset descriptions, the descriptors of the expressions chosen for the target variable under consideration, their statistical significance, the log fold change (LogFC) per gene for the target variable, and the data analysis results of these selected expressions.

Table 3 shows that statistically significant differences in lncRNA expression were identified between the rat group with nephrotoxicity and the control group for all lncRNA expressions except for rna-XR_005493563.1 (LOC120096731) (p < 0.05).

The findings of the performance metrics achieved as a result of the tree-based RF model using the selected lncRNAs are shown in

Table 4. The findings of the performance metrics achieved from the tree-based RF model.

Metric	Value (%) (95% CI)
B-Acc	88.9 (76.7–100)
Acc	90 (76.9–100)
Sp	90.9 (58.7–99.8)
Se	88.9 (51.8–99.7)
Npv	90.9 (58.7–99.8)
Ppv	88.9 (51.8–99.7)
F1-score	88.9 (75.1–100)

.

According to the classification performance of the RF model, the B-Acc was 88.9%, Acc was 90%, Sp was 90.9%, Se was 88.9%, NPV was 90.9%, PPV was 88.9%, and the F1-score was 88.9%.

A graph of the RF model’s performance metrics is shown in Figure 8.

The variable importance values of the expressions for the selected genes to explain the target variable (nephrotoxicity) are shown in

Table 5. The variable importance values of the chosen lncRNAs used to explain the target variable.

Gene Name	Variable Importance Value
rna-XR_591534.3	100
rna-XR_005503408.1	80.127
rna-XR_005495645.1	80.02
rna-XR_001839007.2	47.205
rna-XR_005492056.1	45.374
rna-XR_351582.4	42.972
rna-XR_001840278.2	42.9
rna-XR_005496784.1	41.422
rna-XR_005498350.1	39.116
rna-XR_005503371.1	38.433
rna-NR_133655.1	38.301
rna-XR_005497370.1	35.986
rna-XR_005500805.1	33.445
rna-XR_005496283.1	31.788
rna-XR_353438.4	30.313
rna-XR_005499330.1	29.65
rna-XR_005497310.1	29.435
rna-XR_005503535.1	29.232
rna-XR_358189.4	27.716
rna-XR_005499033.1	24.311
rna-XR_005496888.1	24.018
rna-XR_590665.2	23.715
rna-XR_005497840.1	23.365
rna-XR_005503688.1	19.988
rna-XR_005499541.1	18.123
rna-XR_005496257.1	17.68
rna-XR_005499594.1	17.632
rna-XR_005497230.1	15.566
rna-XR_005493563.1	8.101
rna-XR_001839839.2	5.695
rna-XR_005489439.1	0

.

The rna-XR_591534.3.1, rna-XR_005503408.1, rna-XR_005495645.1, rnaXR_001839007.2, and rna-XR_005492056.1 id lncRNAs have the highest five significance of 100%, 80.127, 80.02%, 47.205, and 45.374%, respectively. Figure 9 shows the variable importance levels of the top five expressions with the highest variable significance for the selected genes to explain the output variable.

LIME, a local explainable artificial intelligence method, was applied to the tree-based random forest model. Figure 10, Figure 11 and Figure 12 show the results for the first three rats as a result of the LIME method. Green bars show features that were positively correlated with the target variable, while red bars show features that were negatively correlated with the target variable.

It was estimated that the rat in Figure 10 did not have nephrotoxicity with a 90% probability. This rat had an rna-XR_005499594.1 value less than 1.0, rna-XR_005503408.1 value of less than 3.75, rna-XR_591534.3 value less than 49.50, rna-XR_001839007.2 value less than 89.00, rna-XR_001839839.2 value less than 0.00, rna-XR_005497230.1 value between 2.75 and 12.00, rna-XR_005496888.1 value less than 11.75, rna-XR_005493563.1 value between 7.5 and 9.5, and an rna-XR_005497370.1 less than 32.5 reduced the likelihood of nephrotoxicity. On the other hand, an rna-XR_005489439.1 value between 41.50 and 152.00 increased the probability of nephrotoxicity.

It was estimated that the rat in Figure 11 did not have nephrotoxicity with a probability of 93%. This rat had an rna-XR_005495645.1 value less than 8.50, rna-XR_351582.4 value less than 40.50, rna-XR_351582.4 value less than 35.50, rna-XR_591534.3 value less than 49.50, rna-XR_005503408.1 values between 3.75 and 9.00, rna-XR_005499594.1 value less than 1.00, rna-XR_005497370.1 value less than 32.50, rna-XR_005497230.1 value less than 2.75 and rna-XR_005499541.1 values between 1.50 and 2.00 reduced the possibility of nephrotoxicity. On the other hand, an rna-XR_001839839.2 value between 0.0 and 3.00 increased the probability of nephrotoxicity.

The rat in Figure 12 was estimated to have a 90% probability of nephrotoxicity. This rat had an rna-XR_005503408.1 value greater than 13.50, rna-XR_005499594.1 value greater than 3.75, rna-XR_005497370.1 value greater than 124.5, rna-XR_005497230.1 value greater than 23.00, rna-XR_351582. 4 value greater than 228.50, rna-XR_001839839.2 value between 0.0 and 3.00, rna-XR_005499541.1 value less than 1.00, rna-XR_005503371.1 value greater than 46.50, and an rna-XR_005503535.1 value greater than 159.50, which increased the possibility of nephrotoxicity. On the other hand, an rna-XR_005495645.1 value between 8.50 and 13.50 decreased the possibility of nephrotoxicity.

4. Discussion

Antineoplastic drugs not only kill pathologically growing cancer cells in the body, but they also destroy rapidly proliferating normal cells. Therefore, many cancer drugs also have side effects on tissues including bone marrow, blood cells, and other rapidly proliferating cells. Although kidney cells do not divide fast, their high blood flow, ability to concentrate poisons in the medullary interstitium, and particular transporters in the tubular epithelium make them highly susceptible to toxic injury [28,29]. Tubulopathies, acute renal failure, and glomerulopathies as prevalent clinical manifestations, and nephrotoxicity, defined as any kidney injury directly or indirectly caused by drugs, occur when kidney-specific detoxification and excretion does not work properly due to damage or destruction of kidney function by exogenous or endogenous toxic substances [3]. Drugs can have harmful effects on many targets in the kidney via diverse biological processes [30]. Because active clearance, reabsorption, intracellular concentration, and the local interstitial accumulation of medicines occur largely at this region in the kidney, the proximal tubule is of special relevance for nephrotoxicity research. The proximal tubule’s continuous exposure to high concentrations of drugs and their toxic metabolites, combined with the high energy demand of epithelial cells in this region, makes it especially vulnerable to noxious stimuli that can lead to tubular cell toxicity and, eventually, to kidney failure [31,32].

Drug toxicity frequently occurs in the kidney, which is the principal control mechanism that maintains the body’s homeostasis and is hence highly vulnerable to xenobiotics [33]. Understanding the harmful mechanisms of nephrotoxicity can help in the creation of medications with fewer side effects and more therapeutic advantages. Mechanisms of drug-induced nephrotoxicity include tubular cell toxicity, inflammation, changes in glomerular hemodynamics, crystal nephropathy, thrombotic microangiopathy, and rhabdomyolysis. New biomarkers that can detect kidney damage early and more precisely must be discovered and developed in order to effectively prevent drug-induced nephrotoxicity. Biomarker candidates for nephrotoxicity assessment have been identified. Although some fail to provide specificity and sensitivity, studies are promising [34,35,36]. The most effective technique for preventing or limiting nephrotoxicity is to have sensitive and specific biomarkers available early in the drug development process, well before clinical trials begin. In preclinical models and clinical settings, these biomarkers should be able to accurately anticipate toxicity, enabling drug developers to successfully counsel patients to modify or abandon potential medicines and switch to alternatives that affect the same target without toxicity [36].

In this study, genomic, histopathological, and immunohistochemical analyses were performed with samples taken from rats with nephrotoxicity induced by an antineoplastic drug, methotrexate, and from rats in the control group, in order to determine biomarkers for drug-induced nephrotoxicity. lncRNA sequence analyses, which are known to be involved in many regulatory mechanisms in the case of transcription and subsequent gene expression as well as fulfill primary functions for quite different biological processes, were performed from tissue samples taken within the scope of the genomic analyses.

According to the histopathological analyses performed in this study, the outer leaf of the Bowman capsule in the renal corpuscle and the glomerular tuft in the cortex was normal in the kidney sections stained with the hematoxylin and eosin staining method in the control group. In the MTX-treated group, significant areas of inflammation were observed in the intertubular areas of the cortex in the preparations examined by the hematoxylin and eosin staining method. In the control group, the macula densa, formed by changing the morphology of the distal tubule cells approaching the vascular pole of the renal corpuscle, was observed to be normal, and the Bowman distance between the outer leaf of the Bowman’s capsule and the inner leaf surrounding the glomerular tuft was also of normal width. In the MTX-treated group, it was observed that epithelial cells of some tubules spilled into the lumen in some stained sections. Another observed condition was the presence of necrotic cells with eosinophilic cytoplasm and dark nuclei in some tubules. With all of these results, it can be said that the general structure of the kidney went beyond what is known and the formation of necrotic cells was observed in the experimental group using MTX. This produces symptoms of kidney damage caused by the drug.

According to the results of the immunohistochemical analysis, positive tubule cells were not found in the control group in the sections where the Cystatin C immunohistochemical staining method was applied, while the density of the positively stained tubules in the MTX-treated group was increased in this group when compared to the control group. It can be said that these differences were based on drug use and liver damage occurred in the drug-administered group compared to the control group. These analyses reveal that histopathologically and immunohistochemically, MTX causes damage to the kidney.

In this study, a genomic dataset containing 16,386 lncRNAs obtained from the kidney tissues of rats with nephrotoxicity and the control group rats was used. According to the findings of the bioinformatic analysis, the rna-XR_005487515.1 id lncRNA showed significantly higher gene expression in the nephrotoxicity group than in the control group. Similarly, rna-XR_361074.4, rna-XR_001839839.2, rna-XR_005486989.1, rna-NR_133655.1, rna-XR_005499594.1, rna-XR_005499333.1, rna-XR_005497370.1, rna-XR_005497230.1, and rna-XR_351582.4 lncRNAs with id had higher gene expression in the group with nephrotoxicity than in the control group. rna-XR_005489095.1, rna-NR_131064.1, rna-XR_005494344.1, rna-XR_005503866.1, rna-XR_005491414.1, rna-XR_360468.4, rna-XR_005501201.1, rna-XR_005493387.1, rna-XR_005499541.1, and rna-XR_005493563.1 lncRNAs with id had very low gene expression in the group with nephrotoxicity compared to the control group.

According to the results of the biostatistical analysis, all genes except rna-XR_005493563.1 (LOC120096731) lncRNA out of the 31 lncRNAs obtained by the Boruta variable selection showed statistically significant differences for the two groups. It shows that it can correctly classify nephrotoxicity according to the performance criteria obtained as a result of the tree-based RF machine learning modeling made by taking the target (nephrotoxicity) variable with 31 lncRNAs selected by the Boruta variable selection method used in the study. In addition, as a result of RF modeling, lncRNAs with the id rna-XR_591534.3, rna-XR_005503408.1, rna-XR_005495645.1, rna-XR_001839007.2, rna-XR_005492056.1, and rnaXR_005492522.1 had the five highest variable significance values. Therefore, these lncRNAs can be used as biomarker candidates for nephrotoxicity. When the LIME results are considered, it was observed that the high level of lncRNAs with id rna-XR_591534.3 and rna-XR_005503408.1 increased the possibility of nephrotoxicity. Currently, there is limited clinical information available regarding the use and effectiveness of lncRNAs rna-XR_591534.3 (LOC103691816) and rna-XR_005503408.1 (LOC120102212) against nephrotoxicity. However, considering that lncRNAs can modulate inflammatory responses, regulate cell death pathways, impact epigenetic modifications, and participate in cellular stress responses, they have the potential to contribute to nephrotoxicity, and further investigation into these specific lncRNAs and their mechanisms can enhance our understanding of the involved molecular pathways and aid in the development of innovative therapeutic approaches for nephrotoxicity. Based on the preliminary results of the current research, targeted or untargeted next-generation sequencing studies on human materials, especially the biomarkers discovered in this study, in the following clinical studies can determine the possible roles of the proposed biomarkers in nephrotoxicity mechanisms and potential drug development efforts.

This study had certain limitations. It was conducted using the data obtained from mouse experiments and serves as a foundation for future research. However, it is essential to conduct human studies to validate the findings and enable the generalization of the results. These human studies would also facilitate their application in potential drug development studies.

5. Conclusions

Considering the findings of the possible biomarkers related to the presence of drug-induced MTX treatment in the current study, the results from the different articles should be evaluated for a comprehensive analysis. In addition, the possible biomarker candidates from the high-performance tree-based modeling and LIME may be administrated for personalized treatment and diagnosis.