PPARalpha siRNA-treated expression profiles uncover the causal sufficiency network for compound-induced liver hypertrophy

Abstract

Uncovering pathways underlying drug-induced toxicity is a fundamental objective in the field of toxicogenomics. Developing mechanism-based toxicity biomarkers requires the identification of such novel pathways and the order of their sufficiency in causing a phenotypic response. Genome-wide RNA interference (RNAi) phenotypic screening has emerged as an effective tool in unveiling the genes essential for specific cellular functions and biological activities. However, eliciting the relative contribution of and sufficiency relationships among the genes identified remains challenging. In the rodent, the most widely used animal model in preclinical studies, it is unrealistic to exhaustively examine all potential interactions by RNAi screening. Application of existing computational approaches to infer regulatory networks with biological outcomes in the rodent is limited by the requirements for a large number of targeted permutations. Therefore, we developed a two-step relay method that requires only one targeted perturbation for genome-wide de novo pathway discovery. Using expression profiles in response to small interfering RNAs (siRNAs) against the gene for peroxisome proliferator-activated receptor alpha (Ppara), our method unveiled the potential causal sufficiency order network for liver hypertrophy in the rodent. The validity of the inferred 16 causal transcripts or 15 known genes for PPARalpha-induced liver hypertrophy is supported by their ability to predict non-PPARalpha-induced liver hypertrophy with 84% sensitivity and 76% specificity. Simulation shows that the probability of achieving such predictive accuracy without the inferred causal relationship is exceedingly small (p < 0.005). Five of the most sufficient causal genes have been previously disrupted in mouse models; the resulting phenotypic changes in the liver support the inferred causal roles in liver hypertrophy. Our results demonstrate the feasibility of defining pathways mediating drug-induced toxicity from siRNA-treated expression profiles. When combined with phenotypic evaluation, our approach should help to unleash the full potential of siRNAs in systematically unveiling the molecular mechanism of biological events.

Figure 1. Effective Modification of Ppara mRNA and Its Known Target Genes by siRNA In Vivo

Ppara was monitored with three probes, Ppara 1, 2, and 3, targeting different fragments of the whole Ppara transcript. The mRNA levels of Ppara, as detected by all three Ppara probes (A), and by its known target genes Cpt1a (B) and Acadl (C) decreased coordinately in the livers of mice treated with Ppara siRNA1 and Ppara siRNA2. The mRNA levels of Ppara and its known target genes, Cpt1a or Acadl, were also measured in Ppara ^−/− mice as a positive control for the silencing effect of Ppara siRNA. Similar concordant regulation between Ppara and its known targets is observed in Ppara ^−/− mice when compared with wild-type mice at day 1 (D1) or day 7 (D7) after vehicle treatment. The mRNA level for each mouse is indicated by a vertical bar.

Figure 2. Schematic of the LCA That Determines Genes Mediating Certain Cellular Responses from Expression Profiles after Perturbation by siRNA

The three major scenarios for the effect of siRNA on gene expression and biological response are illustrated in (A): the direct target or the on-target gene for siRNA and/or its downstream regulated genes resulting in a cellular response; (C): the off-target genes and their downstream regulated genes causing the cellular response; and (E): the cellular effect of an siRNA toward on-target and/or off-target genes and their downstream regulated genes leading to the cellular response. The corresponding causal models among the instrumental variable W, the top gene X, and the bottom gene or response Y for these scenarios are shown in (B,D,F). The hidden variable, H, is not measurable by expression profiles. The flow chart for two-step relay is illustrated in (G).

Figure 3. Identification of the Top and Bottom Gene Pairs and the Causal Genes for PPARα–AILH by the Two-Step Relay Approach

Conditional independency between the instrumental variable (Ppara) and the bottom gene given the top gene in selected gene pairs (A–C) is illustrated. When conditioned on the identified top genes, indicated in red (A), the correlation between siRNA effect, manifested by Ppara mRNA level and the bottom genes, vanishes. However, when conditioned on the bottom genes, indicated in red (B), the correlation between Ppara and the selected top genes remains significant. Similarly, the dependency between the top genes and liver hypertrophy is abolished when conditioned on the bottom genes (p < 0.01), indicated by red dots in (D). The dependency between the top genes and bottom genes remains when conditioned on liver hypertrophy (E), as does the dependency between liver hypertrophy and the bottom genes (p < 0.01) when conditioned on the top genes, indicated by red dots (F). To assess the false-positive rate among the derived relationships after multiple tests, Monte Carlo simulation was done. The estimated false-positive rate for the 15 genes and one EST inferred to mediate PPARα–AILH after testing all of the top and bottom gene pairs is 0.002.

Figure 4. Predictability of Liver Hypertrophy Induced by Non-PPARα Compounds from Inferred PPARα–AILH Causal Genes

Nine selected transcripts for the logistic regression model from the 16 inferred causal genes were coordinately regulated in the training set (A). A total of 211 expression profiles for the nine measured transcripts were aligned with liver hypertrophy, defined by the liver/body weight ratio (B) in corresponding rats (color bar: −0.3:0.3 at log₁₀ scale). The expected normal liver and liver hypertrophy (LH), based on measured liver/body weight ratio in the training set, are indicated by blue dots and diamonds, respectively, while the predicted normal liver and LH derived from the established model are indicated by red dots and diamonds, respectively (B). Based on the logistic regression model built from the training set, the pathological condition of the liver in the independent testing set (normal or hypertrophy) was predicted (D). The expected normal liver and LH based on measured liver/body weight ratio in the testing set are illustrated by blue dots and diamonds, respectively. The predicted normal liver and LH derived from the established model from the training set are indicated by red dots and diamonds, respectively. Distinctive patterns of the selected biomarkers are evident between the normal and hypertrophic livers [(A,C); color bar: −0.3:0.3 at log₁₀ scale]. The probability of obtaining such a set of predictive biomarkers from the genes correlated with the liver/body weight ratio in the PPAR minicompendium was significantly small in both the training dataset [(E), p < 0.001] and the testing dataset [(F), p < 0.005]. The AUCs for the ROC of the built model based on causal transcripts are indicated by the blue and red bars for the training and testing dataset, respectively, among the distribution of AUCs for 10,000 trials of 16 genes randomly selected from 757 genes correlated with the endpoint in the minicompendium (E,F).

Figure 5. The Sufficiency Order Network of the 15 Causal Genes Mediating PPARα–AILH, as Determined by CI Tests against Liver Hypertrophy