TL;DR
This paper introduces a novel "Predictive Toxicogenomics Space" (PTGS) derived from transcriptomics data that accurately predicts liver damage and drug-induced liver injury (DILI). By fusing vast amounts of gene expression and cytotoxicity data, the PTGS offers a powerful new tool for assessing chemical safety and predicting adverse drug reactions with unprecedented accuracy.
What Problem Does This Paper Address?
Predicting the harmful effects of chemicals and drugs is a major challenge in toxicology and drug development. Current safety testing methods, while rigorous, often fail to identify inherent toxicological properties, especially those that manifest only after prolonged exposure or in specific individuals. Drug-induced liver injury (DILI) is a particularly significant problem, as it can be severe and is often not predicted by preclinical tests. Existing approaches using 'omics data to identify toxicity mechanisms face challenges with high dimensionality, scaling across different experimental systems, and accounting for dose-dependent responses. This paper aims to bridge this gap by developing a data-driven approach to create a predictive model for liver toxicity and DILI.
Methods
The core of this work is the development of a Predictive Toxicogenomics Space (PTGS). This was achieved through a "big data compacting and data fusion" approach.
- Probabilistic Component Modelling: The researchers used latent Dirichlet allocation (LDA), a probabilistic modeling technique, to analyze transcriptomics data from the Connectivity Map (CMap) and cytotoxicity data from the NCI-60 cancer cell line screen. This unsupervised method aims to uncover common themes (components) that describe collections of gene expression profiles and link them to compound treatments.
- Data Fusion: The CMap and NCI-60 datasets were fused to create a comprehensive dataset. This allowed for the identification of components that were strongly associated with cytotoxicity.
- PTGS Construction: Through a semi-supervised process, the researchers identified 14 overlapping gene space components that best captured dose-dependent cytotoxicity. These 14 components, comprising 1,331 genes, form the PTGS.
- Cytotoxicity Scoring: A scoring concept was developed based on the activation of PTGS components. This score provides a virtual probability estimate of cytotoxicity.
- Validation: The PTGS was validated against various datasets, including rat liver pathology data from the TG-GATEs database and human hepatocyte experiments, to assess its ability to predict liver cytopathology and human DILI.
Datasets
- Connectivity Map (CMap): A large dataset of gene expression profiles from various cell lines treated with thousands of compounds. This was the primary source for transcriptomic data.
- NCI-60 Cancer Cell Line Screen: Provided cytotoxicity data (GI50, TGI, LC50) for numerous compounds across 59 cell lines. This dataset was crucial for linking gene expression to cytotoxic effects.
- Open TG-GATEs: A toxicogenomics database containing in vitro and in vivo analyses of potentially hepatotoxic compounds, including rat liver pathology data and human/rat hepatocyte gene expression data. This was used for validating predictions of liver pathology and DILI.
- Liver Toxicity Knowledge Base (LTKB): A resource for drug information related to liver damage, used for annotating DILI-related drugs.
Key Insights and Ideas
- Data Fusion for Predictive Power: The study demonstrates that fusing diverse 'omics datasets (transcriptomics and cytotoxicity) using probabilistic modeling can create a powerful predictive space for complex biological outcomes like toxicity.
- A "Gene Space" for Toxicity: The PTGS represents a novel concept of a "gene space" specifically designed to capture and predict toxicological responses, moving beyond individual gene analysis.
- Dose-Dependency Captured: The PTGS effectively captures dose-dependent cytotoxicity, a critical factor often missed in simpler toxicity assays.
- In Vitro to In Vivo Extrapolation: The PTGS shows remarkable ability to extrapolate findings from in vitro cell-based experiments to predict in vivo liver pathology and human DILI, a significant challenge in toxicology.
- Complementary to Existing Tests: The PTGS not only outperforms existing methods like QSAR but also complements them, suggesting a path towards more integrated and accurate safety assessments.
Key Results
- The PTGS, composed of 1,331 genes within 14 components, was constructed using approximately 2.5 x 10<sup>8</sup> data points and 1,300 compounds.
- The PTGS accurately predicts dose-dependent cytotoxicity effects and provides a virtual cytotoxicity probability estimate.
- It successfully predicts chemically-induced pathological states in rat livers from repeated dosing.
- Crucially, the PTGS predicts human drug-induced liver injury (DILI) from hepatocyte experiments with high accuracy.
- When analyzing 68 DILI-annotated drugs, the PTGS tool demonstrated superior performance compared to existing tests, achieving a "hereto-unseen level of DILI prediction accuracy."
- The PTGS-based grouping of compounds outperformed structure-based grouping in predicting cytotoxicity.
Limitations and Future Work
- Cellular Models: While the study uses diverse cell lines, the reliance on tumor cell lines (CMap, NCI-60) for initial PTGS construction might introduce biases. The authors suggest that future studies could explore the PTGS concept using proteomics and metabolomics data.
- RNA-Seq Potential: The current PTGS was developed using microarray data. The authors hypothesize that RNA-seq data, with its higher sensitivity, could potentially detect PTGS activation at even lower doses, which warrants further investigation.
- Broader Organ Toxicity: The study focused on liver toxicity. The authors suggest that applying the PTGS concept to capture further organ toxicities (e.g., heart, kidney) is a promising future direction, as component associations to these conditions were observed.
- Refinement of DILI Prediction: While the DILI prediction is highly accurate, further refinement could be achieved by standardizing drug concentrations relative to therapeutic Cmax more precisely and by incorporating more negative control compounds.
- AOP Integration: The authors propose that PTGS could be integrated into Adverse Outcome Pathway (AOP)-based risk assessment strategies, serving as a "biomarker" for key events.
Summary updated