A Linear Prediction Rule Based on Ensemble Classifiers for Non-Genotoxic Carcinogenicity

ABC gene-ranking for prediction of drug-induced cholestasis in rats

As legacy toxicogenomics databases have become available, improved data mining approaches are now key to extracting and visualizing subtle relationships between toxicants and gene expression. In the present study, a novel “aggregating bundles of clusters” (ABC) procedure was applied to separate cholestatic from non-cholestatic drugs and model toxicants in the Johnson & Johnson (Janssen) rat liver toxicogenomics database [3]. Drug-induced cholestasis is an important issue, particularly when a new compound enters the market with this liability, with standard preclinical models often mispredicting this toxicity. Three well-characterized cholestasis-responsive genes (Cyp7a1, Mrp3 and Bsep) were chosen from a previous in-house Janssen gene expression signature; these three genes show differing, non-redundant responses across the 90+ paradigm compounds in our database. Using the ABC procedure, extraneous contributions were minimized in comparisons of compound gene responses. All genes were assigned weights proportional to their correlations with Cyp7a1, Mrp3 and Bsep, and a resampling technique was used to derive a stable measure of compound similarity. The compounds that were known to be associated with rat cholestasis generally had small values of this measure relative to each other but also had large values of this measure relative to non-cholestatic compounds. Visualization of the data with the ABC-derived signature showed a very tight, essentially identically behaving cluster of robust human cholestatic drugs and experimental cholestatic toxicants (ethinyl estradiol, LPS, ANIT and methylene dianiline, disulfiram, naltrexone, methapyrilene, phenacetin, alpha-methyl dopa, flutamide, the NSAIDs–—indomethacin, flurbiprofen, diclofenac, flufenamic acid, sulindac, and nimesulide, butylated hydroxytoluene, piperonyl butoxide, and bromobenzene), some slightly less active compounds (3′-acetamidofluorene, amsacrine, hydralazine, tannic acid), some drugs that behaved very differently, and were distinct from both non-cholestatic and cholestatic drugs (ketoconazole, dipyridamole, cyproheptadine and aniline), and many postulated human cholestatic drugs that in rat showed no evidence of cholestasis (chlorpromazine, erythromycin, niacin, captopril, dapsone, rifampicin, glibenclamide, simvastatin, furosemide, tamoxifen, and sulfamethoxazole). Most of these latter drugs were noted previously by other groups as showing cholestasis only in humans. The results of this work suggest that the ABC procedure and similar statistical approaches can be instrumental in combining data to compare toxicants across toxicogenomics databases, extract similarities among responses and reduce unexplained data varation.

Release of (and lessons learned from mining) a pioneering large toxicogenomics database

AIM

We release the Janssen Toxicogenomics database. This rat liver gene-expression database was generated using Codelink microarrays, and has been used over the past years within Janssen to derive signatures for multiple end points and to classify proprietary compounds.

MATERIALS & METHODS

The release consists of gene-expression responses to 124 compounds, selected to give a broad coverage of liver-active compounds. A selection of the compounds were also analyzed on Affymetrix microarrays.

RESULTS

The release includes results of an in-house reannotation pipeline to Entrez gene annotations, to classify probes into different confidence classes. High confidence unambiguously annotated probes were used to create gene-level data which served as starting point for cross-platform comparisons. Connectivity map-based similarity methods show excellent agreement between Codelink and Affymetrix runs of the same samples. We also compared our dataset with the Japanese Toxicogenomics Project and observed reasonable agreement, especially for compounds with stronger gene signatures. We describe an R-package containing the gene-level data and show how it can be used for expression-based similarity searches.

CONCLUSION

Comparing the same biological samples run on the Affymetrix and the Codelink platform, good correspondence is observed using connectivity mapping approaches. As expected, this correspondence is smaller when the data are compared with an independent dataset such as TG-GATE. We hope that this collection of gene-expression profiles will be incorporated in toxicogenomics pipelines of users.