(M1030-05-29) Computational Toxicity Prediction for Compounds of Unknown Toxicity using Bulk RNA-Seq: A Case of Liver Toxicity

Purpose: Pharmaceuticals, cosmetics, and even food include chemical components that are closely related to our health and the environment. The information on chemical toxicity is still considered undiscovered territory. To save the time and expense of trying to assess toxicity in animal tests, many computational approaches have been employed to predict the toxicity profile of chemicals. Advances in next-generation sequencing technology have resulted in a flood of biological data, leading to the establishment of open repositories for storing and sharing high-throughput sequencing data. In humans, gene expression profiling is frequently utilized to elucidate the toxicity processes of environmental toxins, medicates, and biomarkers. Toxicogenomics, which evaluates toxicity using expression data from hundreds or even thousands of genes, can be a valuable tool for identifying biological processes related to chemical toxicity¹.

Methods: A comprehensive database was integrated using the gene expression information from DrugMatrix² and Connectivity Map³. In addition, the terms defining disease-related toxicity were grouped into 217 medical subject headings⁴, each containing related compounds based on information about the association of that compound with the disease⁵. The method of connectivity score calculation was used for a comparative analysis of the integrated database and the gene set. To assign the direction of the gene expression change in each experimental condition, the weighted enrichment score was calculated with the segregation of the upregulated or downregulated genes. Then, normalized values were obtained for signatures to compensate for differences in expression values caused by treatment time and dose. The tau score was used to assess whether the correlation between a query compound and a specific signature was substantially distinct from the association observed between a specific signature and other compounds. Using 217 medical subject headings, the Fisher’s exact test was performed to predict potential toxicity for compounds with a final Tau score above +90 or below -90. As a case of our method, we performed the toxicity prediction using the gene expression profile of trovafloxacin known to cause liver toxicity. RNAs were extracted from immortalized murine Kupffer cells treated with 2.5 μM TVX or DMSO for 24 h, followed by RNA sequencing⁶. Through the RNA-Seq analytic process, differentially expressed gene sets were selected with a p-value of 0.05 or less and a fold change of more than 2-fold. The gene set of trovafloxacin was compared one by one with the gene expression signatures in the integrated database to calculate the connectivity score, and the potential toxicity was predicted using the highly correlated compounds.

Results: To build an integrated database, 57,304 gene expression data from 2,760 chemicals and 1,931,845 compound-disease relationships from 1,323 compounds were collected. The integrated database included 1,401 Affymetrix RG230 arrays, 3,887 Codelink RU1 arrays, and 52,016 Affymetrix HGU133A arrays. Furthermore, 1,931,845 cases of disease-related information were included in the database in which 7,247 cases of disease and 1,323 cases of chemical substances are linked 1:1. The efficacy of toxicity prediction was assessed using a gene set of trovafloxacin known to induce hepatotoxicity. RNA sequencing analysis revealed that 852 genes were upregulated and 1,182 genes were downregulated in the murine immortalized Kupffer cell line, which genes were submitted as queries. Table 1 listed the highest ten compounds that were positively or negatively correlated with the expression following trovafloxacin treatment. Trovafloxacin showed a similar expression pattern to amsacrine and etoposide. And methyldopa, prednisone, and prednisolone all had high Tau scores. SB-225002, albendazole, dasatinib, finasteride, loratadine, griseofulvin, and cerivastatin were the compounds having the opposite expression pattern to trovafloxacin. Toxicity for 1,123 compounds, including 630 positively correlated compounds and 618 negatively correlated compounds, was predicted using Fisher's exact test. The 78 toxicities had p-values less than 0.05. Results are screened based on p values less than 0.05 and sorted by the number of matched compounds (Table 2). The results showed a high correlation with fibrosis, chemical- and drug-induced liver injury, and inflammation.

Conclusion: Toxicity was predicted in this work utilizing gene expression patterns, which differed from earlier structure-activity relationship-based predictive toxicology. Toxicity prediction indicated that the compounds employed in the evaluation were likely to have toxicity to the liver. The expression patterns of trovafloxacin were similar to those of the antineoplastic drugs amsacrine and etoposide. Treatment with these medications at high doses has been shown to cause hepatotoxicity^7,8. SB-225002 has the opposite effect of trovafloxacin, increasing hepatocyte growth and regeneration⁹. Toxicity prediction based on toxicogenomics could be reliably expected to predict the toxicity of unknown chemicals.

References: 1. Chavan-Gautam P, Shah T, Joshi K. Transcriptomics and Epigenomics. Innov Approaches Drug Discov Ethnopharmacol Syst Biol Holist Target. Published online January 1, 2017:235-272.
2. Svoboda DL, Saddler T, Auerbach SS. An Overview of National Toxicology Program’s Toxicogenomic Applications: DrugMatrix and ToxFX. Challenges Adv Comput Chem Phys. 2019;30:141-157.
3. Lamb J, Crawford ED, Peck D, et al. The Connectivity Map: Using Gene-Expression Signatures to Connect Small Molecules, Genes, and Disease. Science (80- ). 2006;313(5795):1929-1935.
4. Lipscomb CE. Medical Subject Headings (MeSH). Bull Med Libr Assoc. 2000;88(3):265.
5. Davis AP, Grondin CJ, Johnson RJ, et al. Comparative Toxicogenomics Database (CTD): update 2021. Nucleic Acids Res. 2021;49(D1):D1138-D1143.
6. Yang H, Park T, Park D, Kang M-G. Trovafloxacin drives inflammation-associated drug-induced adverse hepatic reaction by changing macrophage polarization. Toxicol Vitr. 2022;82:105374.
7. Horstmann MA, Hassenpflug W-A, zur Stadt U, Escherich G, Janka G, Kabisch H. Amsacrine combined with etoposide and high-dose methylprednisolone as salvage therapy in acute lymphoblastic leukemia in children. Haematologica. 2005;90(12):1701-1703.
8. Wolff SN. High-dose carmustine and high-dose etoposide: a treatment regimen resulting in enhanced hepatic toxicity. Cancer Treat Rep. 1986;70(12):1464-1465.
9. Kuboki S, Shin T, Huber N, et al. Hepatocyte signaling through CXC chemokine receptor-2 is detrimental to liver recovery after ischemia/reperfusion in mice. Hepatology. 2008;48(4):1213-1223.

Acknowledgments: This research was supported by the Korea Institute of Toxicology, Republic of Korea (1711159828).

Table 1. List of 10 compounds with the strongest positive or negative correlations with the expression profile of trovafloxacin.



Table 2. List of toxicities predicted by the expression profile of trovafloxacin