Prediction Model Construction for Cholestasis by Toxicogenomics Approach
Takashi Matsuda1), Daisuke Kigami2), Koutaro Tamura2), Akira Unami2), Masato Kobori1), Tetsuro Urushidani3)4) and Yasuo Ohno5) 1) Molecular Medicine Research Labs., Drug Discovery Research, Astellas Pharma Inc., 21, Miyukigaoka, Tsukuba-shi, Ibaraki 305-8585 JAPAN 2) Drug Safety Labs., Drug Discovery Research, Astellas Pharma Inc., 1-6 Kashima 2-chome, Yodogawa-ku, Osaka 532-8514 JAPAN 3) Toxicogenomics Informatics Project, National Institute of Biomedical Innovation, Ibaraki, Osaka, 567-0085, JAPAN 4) Department of Pathophysiology, Doshisha Women's College of Liberal Arts, Kyotanabe, Kyoto 610-0395, JAPAN 5) National Institute of Health Sciences, Setagaya-ku, Tokyo 158-8501, JAPAN [email protected]
1. Introduction The Toxicogenomics Project in Japan (TGP)  was started in 2002 and finished in 2007 as a collaborative project by the National Institute of Health Science (NIHS) and 15 pharmaceutical companies. In 2007, the Toxicogenomics Informatics Project (TGP2) was started as a subsequent project of TGP. In these projects, about 150 compounds were comprehensively analyzed and large-scale transcriptome database were constructed. In the present study, we selected cholestasis which is one of serious adverse effects as a target phenotype and constructed classifier for cholestasis based on linear. discriminant analysis (LDA) from microarray profile.
2. Materials and Methods Animal treatment Male Sprague-Dawley rats were daily administered with each compound (3 dose levels) or vehicle control for 3, 7, 14 or 28 days. The animals were sacrificed 24 hours after the last dosing and the liver in each animal was obtained for microarray analysis. Three animals were analyzed for each group (4 time points and 4 dose levels including vehicle control). Microarray analysis Microarray analysis was performed with GeneChip○R Rat Genome 230 2.0 Arrays (Affymetrix, Santa Clara, CA, USA). The raw signal intensity data were normalized by global mean normalization method. Compounds and training dataset From our database, we selected 10 cholestasis positive/negative compounds for training dataset. Alpha-naphthylisothiocyanate (1.5, 5, 15 mg/kg), chlorpromazine (4.5, 15, 45 mg/kg), carbamazepine (30, 100, 300 mg/kg), azathioprine (3, 10, 30 mg/kg) and ethinylestradiol (1, 3, 10 mg/kg) were selected as cholestasis positive compounds. Benzbromarone (20, 60, 200 mg/kg), hexachlorobenzene (30, 100, 300 mg/kg), Wy-14,643 (10, 30, 100 mg/kg), adapin (10,
30, 100 mg/kg) and pemoline (7.5, 25, 75 mg/kg) were selected as cholestasis negative compounds. Statistical analysis We first extracted differentially expressed probe sets of each positive compound by comparing highest dose profile to corresponding control profile for each time point with Welch’s t-test (p<0.2) and mean fold change (fc>2.0). Then we selected probe sets, which were commonly extracted from more than 4 out of 5 positive compounds, for next prediction model construction step. In prediction model construction step, each profile was converted to fold change compared to control mean and classifier was constructed by LDA. To select effective probe sets for classifier from among probe sets selected previous step, we applied forward stepwise selection based on singular values and selected the smallest set of probe sets which provided the best accuracy estimated by leaving one compound out cross-validation (LOCO CV). Furthermore, stepwise feature elimination were performed while estimated accuracy by LOCO CV was not reduced. Finally, we constructed classifier for cholestasis by LDA with selected probe set.
3. Results and Discussion The probe sets selected finally for cholestasis classifier are listed in Table 1 and ROC curve with test compounds profile consist of 7 positive and 6 negative compounds is shown in Figure 1. From the figure it is found that for example we can detect 80 % of cholestasis positive compound if we allow 30 % false positive. Thus the classifier constructed in present study can classify cholestasis positive and negative compound profiles effectively. Table 1 probe set list
1367802_at 1368147_at 1368171_at 1368272_at 1368607_at 1370026_at 1370384_a_at 1371143_at 1373146_at 1374093_at 1374883_at 1375647_at 1376958_at
Sgk Dusp1 Lox Got1 Cyp4a12 Cryab Prlr Serpina7 Ssx2ip EST Mtmr7_predicted EST RGD1562844_predicted
1379027_at 1379513_at 1380351_at 1381748_at 1382443_at 1382451_at 1383047_at 1383248_at 1383486_at 1383946_at 1385005_at 1386922_at 1387022_at
RGD1308329_predicted Tmem30b_predicted Sugt1 Raph1_predicted RGD1562451_predicted Hebp2_predicted Gas6 Fmo5 EST Cldn1 EST Ca2 Aldh1a1
1387139_at 1387197_at 1387270_at 1387583_at 1387985_a_at 1388199_at 1389221_at 1389528_s_at 1390860_at 1395542_at 1397596_at 1398309_at 1398759_at
Hao2 Omd Hhex Cyp26a1 Obp3 Tacstd1 Mmd2_predicted Jun Igf2bp3 EST Trim2 Pigl Tgfb1i4
Acknowledgement This study was supported in part by a grant from the Ministry of Health, Labor and Welfare (H14 and H19-Toxico-001). References 1. Urushidani T, Nagao T, 2005, Handbook of Toxicogenomics – Strategies ans Applicatioins, Wiley – VCH, 623-631.