Toxicology has traditionally relied on apical observations in animal models with limited use of high-throughput transcriptomics and AI. Yet these single endpoints with limited predictive power often miss early mechanistic indicators of toxicity, contributing to costly late-stage compound failures.
High-throughput transcriptomics provides a scalable molecular layer that complements conventional toxicology and gives AI the high-dimensional, experimentally grounded data it needs to make more mechanistic predictions. Growing evidence indicates that large-scale gene expression profiling helps prioritize chemicals earlier in development, identifies early transcriptional points of departure (tPOD) for better safety assessment, and predicts systems-level toxicity when paired with AI models.
In this article, we examine the consequences of missed compound toxicity and how high-throughput transcriptomic technologies paired with AI can improve clarity in safety assessment decisions.
Key Takeaways for High-Throughput Transcriptomics in Toxicology
- High-throughput transcriptomics adds a scalable molecular layer that complements conventional toxicology by measuring broad gene-expression responses to chemical perturbations.
- AI can use transcriptomic profiles to identify patterns associated with toxicity, mechanisms of action, biomarkers, and chemical potency.
- Dose-resolved high-throughput transcriptomics can support tPOD derivation, providing quantitative measures of molecular perturbation.
- Combining high-throughput transcriptomics with Cell Painting captures complementary biological information, potentially improving compound MoA and toxicity prediction.
- Screening-scale transcriptomics generates proprietary AI-ready datasets, allowing experimental data generation and model development to operate as an iterative learning cycle.
What is High-Throughput Transcriptomics?
High-throughput transcriptomics measures changes in gene expression transcriptome-wide and across large numbers of chemical perturbations, doses, and biological models, often in high-throughput screening formats. Unlike conventional toxicity assays that typically measure one or a small number of endpoints, high-throughput transcriptomics captures broad molecular responses that can be used to identify pathways, mechanisms, biomarkers, and dose-dependent responses to perturbations, while enabling gene-level resolution.
Unlike conventional low-throughput RNA-seq workflows, high-throughput transcriptomic assays are miniaturized and optimized for large-scale screening. Depending on the technology, experiments can be performed in 96-, 384-, or 1536-well formats, enabling researchers to profile hundreds to thousands of chemical–dose–model combinations in a single study.
This scalability makes high-throughput transcriptomics particularly valuable for AI-enabled toxicology. Each perturbation generates a standardized, high-dimensional biological response profile that can be used for compound prioritization, mechanism-of-action analysis, potency assessment, biomarker discovery, and training predictive models.
See the technical performance and reproducibility of screening-scale MERCURIUS™ 1536 DRUG-seq across multiple independent 1536-well plate experiments and cell types in our recent tech note.
What is the Cost of Missed Toxicity Without High-Throughput Transcriptomics?
Drug safety is a major bottleneck of drug development. Hepatotoxicity accounts for over a quarter of drug failures in clinical trials and one third of market withdrawals caused by adverse drug reactions (1,2).
With drug development costing upwards of US$2.6 billion and taking approximately 12 to 15 years on average, late-stage or on-market failures due to unforeseen toxicity can be immensely expensive, disruptive to pipelines, and represent an overwhelming opportunity cost (3).
These outcomes are among the main drivers of a push to adopt Next Generation Risk Assessment (NGRA) frameworks and new approach methodologies (NAMs), such as high-throughput transcriptomics paired with more human-relevant cellular models and AI (4).
How The Biological Depth of High-Throughput Transcriptomics Paired with AI Supports Early Toxicity Detection
One high-profile late-stage failure was fasiglifam (TAK-875). The GPR40 agonist was developed to treat type 2 diabetes mellitus and demonstrated significant efficacy in improving glycemic control, with a low risk of hypoglycemia.
Despite a clean safety profile in early-stage animal models, development was halted due to a small number of serious drug-induced liver injury (DILI) cases in phase III clinical trials (5,6). Subsequent in vitro, in vivo, and in silico studies aimed to determine why fasiglifam induced severe DILI and why preclinical evidence had not adequately anticipated the clinical liver injury, so they investigated whether transcriptomic changes could provide additional mechanistic insight.
Subsequent transcriptomic studies provided evidence of molecular changes associated with fasiglifam-induced liver injury that were not apparent from conventional histopathological and circulating-biomarker assessments. For instance, gene expression profiling of livers from different mouse strains post-fasiglifam treatment identified hundreds of treatment-responsive genes associated with immune response, bile acid homeostasis, oxidative stress, and mitochondrial dysfunction, despite no liver histological changes and few differences in plasma biomarkers of hepatotoxicity (5,7).
Toxicology researchers now have access to large-scale resources such as DILImap, which provides transcriptome-wide gene expression profiles for 300 compounds across multiple concentrations in primary human hepatocytes (PHHs) (8). When combined with the AI model, ToxPredictor, it predicted liver injury with 88% sensitivity and 100% specificity.
ToxPredictor identified hepatotoxic gene expression signatures from fasiglifam and other late-stage failures, such as evobrutinib and BMS-986142, classifying them as ‘high-risk’ for DILI (8). Therefore, unlike single-endpoint toxicity assays that generally provide a binary result, high-throughput transcriptomics provides high-dimensional data that can allow AI models to identify stress signatures and DILI-associated mechanisms that may precede overt phenotypic effects in some cases.
Together, these findings illustrate how transcriptomic profiles can reveal molecular signatures associated with toxicity that may not be apparent from conventional endpoints. When combined with AI, these high-dimensional profiles can support earlier identification and prioritization of potential safety liabilities.
Want to build AI-ready toxicology datasets tailored to your own compounds? Learn how high-throughput transcriptomics can support proprietary dataset generation for AI model development.
Why Do Traditional Toxicology Assays Sometimes Miss Early Toxicity Indicators?
Traditional pre-clinical toxicology assays can miss early toxicity indicators because they generally rely on animal-based apical endpoints to determine a compound’s no-observed-adverse-effect level (NOAEL). These apical endpoints and associated assays often include fluctuations in organ weight, elevated liver enzymes in the blood, or histopathological evidence of necrosis or inflammation, but they lack high-resolution, systems-level information on early changes in gene expression that often precede overt phenotypic changes (Table 1) (9).
| Assay Type | Where It’s Used | What It Measures | Why It’s Used for Decision-Making | Limitations for Early Toxicity Detection |
|---|---|---|---|---|
| Histopathology | In vivo (rodent, non-rodent tox studies) | Tissue morphology (necrosis, inflammation, fibrosis, steatosis) | Gold standard for identifying target organ toxicity and defining NOAELs | Detects end-stage tissue damage only; requires animal sacrifice; misses early molecular initiating events; subjective interpretation limits reproducibility |
| Serum liver enzymes (ALT, AST, ALP, bilirubin) | In vivo and clinical chemistry panels | Circulating biomarkers of hepatocellular injury | Widely accepted clinical and preclinical biomarkers of liver toxicity | Lagging indicators; only elevated after significant damage; poor specificity; no mechanistic insight |
| In vivo repeated-dose toxicity studies (28-/90-day) | Regulatory preclinical safety packages (OECD TG 407/408) | Systemic toxicity via apical endpoints (clinical chemistry, haematology, histopathology, organ weights) | Central to regulatory approval and dose selection (NOAEL, MTD) | Time-consuming, expensive, and animal-intensive; species differences limit human translation; apical endpoints miss early pathway changes |
| hERG channel inhibition assay | In vitro safety pharmacology | Inhibition of cardiac hERG potassium channel (QT prolongation risk) | Regulatory requirement for assessing proarrhythmic risk | Single-target assay; fails to detect broader cardiotoxicity mechanisms |
| CYP450 inhibition/induction assays | In vitro (microsomes, hepatocytes) | Modulation of cytochrome P450 enzyme activity (drug-drug interaction risk) | Essential for predicting metabolic interactions and clearance | Focused solely on metabolism; no insight into downstream consequences; does not capture system-level toxicity or adaptive stress responses |
While these assays remain central to established toxicology packages, many conventional endpoints are relatively late indicators of biological injury and provide limited molecular resolution. They can show that tissue damage has occurred, but generally provide less information about the early molecular perturbations and pathways that precede that outcome. High-throughput transcriptomics complements these endpoints by providing a broader, earlier view of cellular responses (9).
A systematic review from 2000 found that animal models failed to identify about half of pharmaceuticals associated with clinical DILI, highlighting the challenge of predicting human-specific responses (10). This has likely contributed to the critical need for more relevant, human-systems-focused safety assessment approaches, such as pairing high-throughput transcriptomics with human-focused NAMs like 3D human organoid models (10).
How Does High-Throughput Transcriptomics Help AI Toxicology?
High-throughput transcriptomics supports AI-enabled toxicology by turning chemical perturbations into standardized, quantitative biological-response profiles at a much greater scale than many conventional molecular assays. Scalable transcriptomics provides AI-enabled toxicology workflows with a much richer intermediate layer between chemical structure and toxicological outcome than a simple toxic/non-toxic label, transforming toxicology from a collection of endpoint measurements into a biological data-generation system that AI can learn from.
Assays like MERCURIUS™ DRUG-seq also address the limitations of apical assays or animal studies by shifting the focus from physical damage to systems-level mechanistic insights at scale. By detecting changes in gene expression associated with toxicity mechanisms, such as activation of oxidative stress, DNA damage, or ER stress pathways, high-throughput transcriptomics enables toxicologists to run large-scale dose-response screens that identify potential early safety concerns long before apical symptoms manifest (4).
Benefits of high-throughput transcriptomics for AI-enabled toxicology:
- Rapid, cost-effective generation of high-dimensional data capturing both overt and nuanced biological responses
- Early detection of molecular and pathway perturbations before changes in conventional toxicological endpoints emerge
- Unbiased, simultaneous assessment of multiple biological pathways and cellular processes
- Identification of early molecular biomarkers and transcriptional signatures predictive of adverse outcomes
- Improved characterization of molecular initiating events (MIEs), mechanisms of action (MoAs), and adverse outcome pathways (AOPs)
- Quantitative comparison of chemical potency and derivation of transcriptomic points of departure (tPODs)
- Automatable, standardized, and reproducible generation of AI-ready data, enabling continuous model improvement and rapid lab-in-the-loop iteration
Applications of high-throughput transcriptomic approaches in toxicology:
- Compound prioritization and hazard triage
- Chemical categorization and biological read-across
- Mechanism-of-action and pathway analysis
- Chemical potency and dose–response comparison
- tPOD derivation
- Hypothesis generation, testing and biomarker discovery
- Generation of proprietary AI-ready toxicology data and predictive models
Learn how gene expression biomarkers can improve early toxicity detection and drug development in our in-depth explainer article.
Which High-Throughput Transcriptomic Technologies Are Suitable for Toxicology?
For toxicology, the most suitable high-throughput transcriptomic technologies combine scale, dose–response information, reproducibility, and compatibility with automated 384-/1536-well screening. 3’ mRNA-seq technologies such as DRUG-seq and Alithea’s optimized commercial version, MERCURIUS™ DRUG-seq, are specifically designed for integration into large-scale toxicology screening pipelines and proprietary AI-ready toxicology dataset generation workflows (11).
Targeted expression-profiling approaches such as the S1500+ gene set and the LINCS L1000 platform also enable highly scalable profiling, but measure a substantially smaller set of genes than whole-transcriptome approaches. This creates different trade-offs among coverage, throughput, cost, and the ability to discover previously uncharacterized biological responses (Table 2) (12,13)
In contrast, MERCURIUS™ DRUG-seq offers broad profiling of the whole transcriptome across up to 1,536 samples in a single tube. Options are also available to generate 3’ mRNA or full-length total RNA readouts for 2D cells or 3D spheroid models. What once started with low-throughput microarrays and sample-by-sample RNA-seq as a proof of concept for toxicology has now evolved into high-throughput whole-transcriptome profiling.
| Assay Type | Methodology | Coverage | Throughput | Key Advantages | Key Disadvantages |
|---|---|---|---|---|---|
| MERCURIUS™ DRUG-seq | Bulk 3′ mRNA-seq with early barcoding and pooling. | ~20,000+ genes (whole transcriptome, depending on sequencing depth) | Ultra-high | Screening-scale throughput. No RNA extraction required. Eliminates individual library prep costs. | 3′ expression readouts. MERCURIUS™ Total DRUG-seq required for full-length transcript detection. |
| TempO-Seq | Targeted sequencing of specific probe-ligated RNA. | Targeted (e.g., ~3,000 genes) or whole transcriptome | High | No RNA extraction required. Suitable for FFPE samples. | Limited to the specific probe set designed. |
| L1000 | Luminex bead-based measurement of 978 genes. | 978 measured; ~12,000 inferred | High | Large reference database (CMap) for comparison. | Relies on computational inference for ~90% of the transcriptome. |
| Traditional RNA-Seq | Classic Next-Generation Sequencing (NGS) technology | ~20,000+ genes (whole transcriptome) | Low | Detects novel transcripts, isoforms, and non-coding RNA. | High cost per sample. Requires prior RNA extraction. Not compatible with screening infrastructure. |
Compare high-throughput transcriptomics platforms for toxicology. See how MERCURIUS™ DRUG-seq compares with TempO-Seq and L1000 for scale, transcriptomic coverage, and screening applications.
How Can Pairing High-Throughput Transcriptomics with Cell Painting Improve Toxicology Predictions?
Pairing high-throughput transcriptomics with Cell Painting can substantially improve AI-enabled toxicology because the two assays capture complementary layers of the cellular response. Cell Painting is an imaging-based assay that analyzes cellular responses to perturbations by staining organelles with six fluorescent dyes, then inferring outcomes for over 1,000 features per cell using machine learning (14). By capturing morphological changes, such as size, shape, and intensity, it creates a “phenotypic profile” of cells, allowing researchers to determine the MoA of compounds or early safety concerns. Many safety assessment teams have already largely embraced other high-throughput, data-rich assays, such as morphological profiling with Cell Painting (14).
However, morphological profiling and transcriptomics capture different dimensions of the cellular response. Our recent study suggests that integrated Cell Painting and MERCURIUS™ Total DRUG-seq profiles consistently outperformed either modality alone for MoA classification and DILI prediction. The two approaches weren’t redundant, as each captured relevant signal the other missed, depending on cell type. Integration recovered both, suggesting new avenues for more predictive safety science across diverse compound classes.
See how combining morphological profiling with high-throughput transcriptomics can improve MoA and DILI prediction. Explore our AGBT 2026 study where we performed Cell Painting and MERCURIUS™ Total DRUG-seq across 140 compounds and three human DILI-relevant cell models.
Transcriptomic Points of Departure Provide Quantitative Precision
A growing number of studies now indicate that tPODs offer greater sensitivity and speed for detecting early molecular toxic effects of compounds and informing go/no-go decisions than traditional PODs (15).
An early low-throughput proof of principle in 2013 used microarrays to demonstrate that tPODs derived from short-term, 5-day exposures are highly concordant with apical points of departure (like organ weights and histopathology) from 13-week studies. It established that molecular changes are reliable early predictors of chronic outcomes (16).
Since then, toxicologists have shown that tPODs can provide earlier, quantitative measures of biological perturbation and may complement or, in defined contexts, inform conventional apical points of departure. For instance, transcriptomics boosted the sensitivity of toxicology testing while shortening assay time, moving from long-term 90-day subchronic rodent tests to shorter-term, week-long studies (15). tPODs are generally derived by modeling the dose–response behavior for each gene independently and then aggregating the gene-level data, which works well in most instances, but was previously impossible to achieve due to the prohibitive cost and poor throughput of RNA-seq or microarray. High-throughput transcriptomics now make tPOD derivation substantially more practical at screening scale (15).
Learn how transcriptomics in short-term in vitro toxicology studies now delivers the mechanistic depth and reproducibility needed to make them viable alternatives to long-term animal models.
High-Throughput Transcriptomics: from Safety Signal to the Future Safety Standard
In summary, high-throughput transcriptomics makes it economically feasible to generate massive, standardized, dose-resolved molecular-response datasets. When paired with phenotypic profiling and AI, these datasets can create a continuously improving experimental foundation for predictive toxicology. Recent studies demonstrate that AI models trained on large, well-characterized transcriptomic datasets can achieve strong predictive performance for specific toxicological endpoints, including DILI.
The opportunity extends beyond earlier toxicity detection. Transcriptomic profiles can support compound prioritization, mechanistic investigation, potency comparison, tPOD derivation, biomarker discovery, and biological read-across. When generated at screening scale, each experiment can simultaneously produce a toxicological result, mechanistic insight, and new training data for predictive models.
The next step is therefore not simply to add transcriptomics as another endpoint to conventional toxicology workflows, but to integrate high-throughput molecular profiling into an iterative screen–learn–predict–screen cycle. Each experiment generates robust mechanistic information and new chemical–biology relationships that can strengthen subsequent predictive models. As the experimental dataset grows, AI models can become better informed by increasingly diverse, experimentally grounded biological responses.
Ready to explore high-throughput transcriptomics for AI-enabled toxicology? Whether you’re evaluating compound prioritization, DILI prediction, MoA characterization, tPOD derivation, or AI-ready data generation, our team can help you design a screening strategy around your biological model, compound library, and decision-making goals. Talk to our team about your toxicology workflow.
References
- Watkins PB. Drug safety sciences and the bottleneck in drug development. Clinical Pharmacology & Therapeutics. 2011 Jun;89(6):788-90.
- Dirven H, Vist GE, Bandhakavi S, Mehta J, Fitch SE, Pound P, Ram R, Kincaid B, Leenaars CH, Chen M, Wright RA. Performance of preclinical models in predicting drug-induced liver injury in humans: a systematic review. Scientific reports. 2021 Mar 18;11(1):6403.
- DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry: new estimates of R&D costs. Journal of Health Economics. 2016 May 1;47:20-33.
- Mitchell CA, Wehmas L, Rouquié D, Caiment F, Currie RA, Crute CE. Navigating complexity in modern toxicology: the role of omics in short-term in vivo studies. Toxicological Sciences. 2025 Nov;208(1):32-7.
- Urano Y, Oda S, Tsuneyama K, Yokoi T. Comparative hepatic transcriptome analyses revealed possible pathogenic mechanisms of fasiglifam (TAK-875)-induced acute liver injury in mice. Chemico-biological interactions. 2018 Dec 25;296:185-97.
- Marcinak JF, Munsaka MS, Watkins PB, Ohira T, Smith N. Liver safety of fasiglifam (TAK-875) in patients with type 2 diabetes: review of the global clinical trial experience. Drug safety. 2018 Jun;41(6):625-40.
- Mosedale M, Cai Y, Eaddy JS, Kirby PJ, Wolenski FS, Dragan Y, Valdar W. Human-relevant mechanisms and risk factors for TAK-875-Induced liver injury identified via a gene pathway-based approach in Collaborative Cross mice. Toxicology. 2021 Sep 1;461:152902.
- Bergen V, Kodella K, Srikrishnan S, Barrandon O, Anderson S, Rogers-Grazado M, Fowler C, Beyene H, Robichaud N, Fulton T, Lapchyk N. A large-scale human toxicogenomics resource for drug-induced liver injury prediction. Nature Communications. 2025 Nov 13;16(1):9860
- Worth AP, Berggren E, Prieto P. Chemicals 2.0 and why we need to bypass the gold standard in regulatory toxicology. Alternatives to Laboratory Animals. 2025 Jan;53(1):21-5.
- Olson H, Betton G, Robinson D, Thomas K, Monro A, Kolaja G, Lilly P, Sanders J, Sipes G, Bracken W, Dorato M. Concordance of the toxicity of pharmaceuticals in humans and in animals. Regulatory toxicology and pharmacology. 2000 Aug 1;32(1):56-67.
- Ye C, Ho DJ, Neri M, Yang C, Kulkarni T, Randhawa R, Henault M, Mostacci N, Farmer P, Renner S, Ihry R. DRUG-seq for miniaturized high-throughput transcriptome profiling in drug discovery. Nature communications. 2018 Oct 17;9(1):4307.
- Mav D, Shah RR, Howard BE, Auerbach SS, Bushel PR, Collins JB, Gerhold DL, Judson RS, Karmaus AL, Maull EA, Mendrick DL. A hybrid gene selection approach to create the S1500+ targeted gene sets for use in high-throughput transcriptomics. PloS one. 2018 Feb 20;13(2):e0191105.
- Subramanian A, Narayan R, Corsello SM, Peck DD, Natoli TE, Lu X, Gould J, Davis JF, Tubelli AA, Asiedu JK, Lahr DL. A next generation connectivity map: L1000 platform and the first 1,000,000 profiles. Cell. 2017 Nov 30;171(6):1437-52.
- Bray MA, Singh S, Han H, Davis CT, Borgeson B, Hartland C, Kost-Alimova M, Gustafsdottir SM, Gibson CC, Carpenter AE. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nature protocols. 2016 Sep;11(9):1757-74.
- O’Brien J, Mitchell C, Auerbach S, Doonan L, Ewald J, Everett L, Faranda A, Johnson K, Reardon A, Rooney J, Shao K. Bioinformatic workflows for deriving transcriptomic points of departure: current status, data gaps, and research priorities. Toxicological Sciences. 2025 Feb;203(2):147-59.
- Thomas RS, Wesselkamper SC, Wang NC, Zhao QJ, Petersen DD, Lambert JC, Cote I, Yang L, Healy E, Black MB, Clewell III HJ. Temporal concordance between apical and transcriptional points of departure for chemical risk assessment. Toxicological Sciences. 2013 Jul 1;134(1):180-94.