DILI uncovered: from animal models to a weight-of-evidence future

If you work in drug discovery or development, drug-induced liver injury – DILI – might be a term you’d rather forget. It remains a leading cause of attrition and post-market withdrawal ​(1, 2)​, and despite decades of research, it keeps catching teams off guard. The hardest cases are the ones that surface late, when the cost of failure is at its highest.

The challenge is both scientific and strategic. When do you look? What models do you trust? And how do you make confident go/no-go calls when the biology is complex and the predictive tools are still being fully developed?

I want to work through what we know about DILI biology, what current preclinical models can and can’t tell us, and where the new approach methodologies (NAMs) – the in vitro, microphysiological, and in silico tools – are starting to change the picture.  The short version is that no single assay reveals every liability. The future is weight of evidence, where drug properties, in vitro data, in silico predictions, and targeted in vivo work, are evaluated together, with each test characterized honestly for what it can and can’t detect.

Why the liver gets hit so often

The liver is one of the most versatile organs in the body. It handles carbohydrate, fat, protein, and hormone metabolism, stores vitamins and iron, and metabolizes bilirubin into bile. It’s also where most of the metabolism of foreign substances – drugs – happens, before they’re excreted.

Hepatocytes make up about 80% of liver cells. The rest are non-parenchymal cells, such as endothelial cells, Kupffer cells (the resident macrophages, central to hepatic immune response), and stellate cells. The microarchitecture matters because of the dual blood supply. Around 75% of hepatic inflow comes via the portal vein, carrying nutrient-rich blood straight from the gastrointestinal tract before it reaches systemic circulation. That means high local drug concentrations. Add in metabolic activation – where the parent drug is converted into toxic metabolites inside the hepatocytes themselves – and you get a tissue structurally predisposed to drug-induced injury.

DILI causes roughly 50% of acute liver failure cases ​(3, 4)​. Viral hepatitis, the second most common cause, sits around 10–20% ​(5, 6)​. Diagnosing DILI is its own problem because patients are often on multiple medications, which hampers attribution. The standard biomarkers – alanine aminotransferase  (ALT), aspartate aminotransferase (AST), alkaline phosphatase (ALP), total bilirubin – aren’t fully DILI-specific and don’t classify severity. Diagnosis usually relies on excluding other causes, with liver biopsy reserved for unclear cases. Newer candidates (certain miRNAs and cytokines) look promising but aren’t in routine clinical use yet.

Three faces of DILI

DILI doesn’t behave like one disease. It splits into three categories, and the distinction shapes how you can – or can’t – test for it.

Intrinsic (direct) DILI is the most common type. Dose-dependent, predictable, reproducible in animal models, and fast in onset. The drug is intrinsically hepatotoxic at sufficiently high doses. This is the version current preclinical testing handles well.

Idiosyncratic DILI is rare, not dose-dependent at therapeutic doses, not reliably predicted by animal models, and develops over highly variable latencies – sometimes years. The mechanism is often immunological, triggered by repeated administration. This is the version current preclinical testing handles poorly.

Indirect DILI is the newest classification. Incidence sits between the other two. It’s partially predictable because the mechanism tends to relate to the drug’s pharmacology – an exaggerated on-target effect, or indirect immune activation, sometimes unmasking underlying liver disease. It doesn’t show classical dose response and isn’t generally reproducible in animal models.

These classifications matter because current testing approaches simply cannot assess every DILI signal. Knowing which category you’re worried about determines whether your preclinical strategy is fit for purpose.

What’s going wrong in the cell?

The cellular mechanisms behind DILI have been studied extensively ​(4, 7, 8)​. The major ones are:

• Mitochondrial dysfunction and oxidative stress – present in most DILI cases
• Bile efflux inhibition leading to cholestasis – particularly via bile salt export pump (BSEP) and multidrug resistance-associated protein (MRP) transporters
• Lysosomal disturbance and phospholipidosis – often triggered by cationic amphiphilic drugs (amiodarone, chloroquine, imipramine)
• Steatosis – excess lipid accumulation, often downstream of mitochondrial dysfunction
• Endoplasmic reticulum stress – linked to drug-induced cholestasis and liver lesions
• Immune-mediated responses – central to idiosyncratic DILI in particular
• Reactive metabolite formation – adduct formation and downstream injury

These mechanisms don’t operate in isolation. Several can be involved in a single hepatotoxic response, and one often feeds another – oxidative stress driving mitochondrial dysfunction, reactive metabolites depleting glutathione, and so on.

Drug properties are themselves risk factors. It’s been known for a long time that a dose defines whether a substance is a poison. Drugs given at high doses, or producing high Cmax values, carry higher DILI risk. So do high lipophilicity (LogP) and a low fraction of sp3-hybridized carbons (FSP3) – flatter, more planar molecules are riskier. Bioactivation potential is another property-level red flag, along with ionization state, polar surface area, and hepatic partitioning.

The clinical phenotype is the sum of drug properties and host factors – genetics, age, sex, lifestyle, comorbidities, co-medications – interacting through these cellular mechanisms. Worth keeping in mind whenever you’re trying to predict a clinical outcome from preclinical data.

Where current models fall short and what NAMs change

Current preclinical DILI assessment leans heavily on in vivo GLP toxicology: sub-chronic and chronic studies in rodents and non-rodents, taking months to years to complete. There are no regulatory GLP in vitro studies specifically for DILI; the GLP in vitro studies that do exist cover mostly genotoxicity and skin toxicity.

This is a problem because idiosyncratic and indirect DILI are exactly the categories animals don’t reproduce well, and yet the liver is one of the most prominent organs for safety failure at both preclinical and clinical phases – hepatotoxicity drove 27% of drug withdrawals between 1990 and 2010 ​(9)​. While the US Food and Drug Administration’s FDA’s 2009 industry guidance focuses on clinical detection (ALT/AST plus total bilirubin), not earlier prediction ​(10)​.

New approach methodologies (NAMs) are the proposed bridge. The term covers 2D and complex in vitro systems, organ-on-chip, in chemico, and in silico modelling. The US FDA’s 2025 roadmap encouraged their use ​(11)​, with initial focus on modalities like monoclonal antibodies, where in vivo models translate especially poorly. In 2026, FDA published a draft guidance on the validation of NAMs for specific contexts of use ​(12)​.

In practice, a tiered in vitro approach makes sense:

Screening
Models: 2D monolayers – HepG2, HepaRG, primary human hepatocytes (PHHs).

Outputs: High-throughput viability, cytotoxicity, mitochondrial impairment, oxidative stress, morphology, impedance. First hints of liabilities that can still be designed out in med chem.

Mechanistic/repeat dose
Models: PHH sandwich cultures; spheroids and microtissues:

Outputs: Maintained metabolic competence; long-term and repeated exposure; mechanism-specific endpoints (BSEP/MRP inhibition, glutathione depletion, ATP depletion, caspase activation).

High physiological relevance
Models:Organoids; liver-on-chip.

Outputs: Liver-like architecture and function, including nutrient flow. Suitable for confirmation, not initial screening – lower throughput, longer prep, higher cost.

FDA’s ISTAND program is qualifying complex in vitro systems for regulatory use ​(13)​. A liver-on-chip technology ​(14)​ has been accepted for DILI qualification, though to my knowledge no microphysiological system (MPS) model has yet been fully qualified for DILI.

In silico tools complete the picture. Quantitative structure-activity relationship (QSAR) is still the most widely used, particularly for genotoxicity and impurity assessment. Physiologically based pharmacokinetic (PBPK) and Toxicokinetic (PBTK) support ADME prediction. Machine learning and deep learning models are now being trained to predict hepatotoxicity and cardiotoxicity with reasonable accuracy. Most of these tools sit in early discovery, supporting internal decision-making rather than regulatory submission – but the trajectory here mirrors the in vitro story.

The future is weight of evidence

No single assay reveals every DILI liability. The future of DILI assessment is the weight-of-evidence approach: drug properties, in vitro data (across screening and complex models), in silico predictions, and refined in vivo studies, all evaluated together – with the performance of each model (accuracy, sensitivity, specificity) explicitly characterized.

For drug developers, the practical implications:

• Use in vitro methods as early as possible. High-throughput screening first, then complex models as the candidate list narrows
• Use drug-property predictors routinely. LogP, FSP3, dose projection, bioactivation alerts – signals worth catching early
• Animal studies continue, but are refined. Reduced durations, virtual control groups, fewer repeats where NAM data can substitute
• Be aware of what each assay can and can’t detect. Especially for idiosyncratic DILI, where no current model is comprehensive

Rather than trying to fully replace in vivo testing, the goal is to build a more human-relevant, mechanistically informative picture of risk – earlier, cheaper, and with fewer surprises in the clinic.

Questions and Answers

Which in vitro models would you recommend for preclinical DILI assessment?

A tier-based approach. Start in the very early phase with simple 2D hepatocyte cultures – that’s where you flag compounds with the obvious liabilities. As you move further down the pipeline, bring in the more complex in vitro methods (spheroids, organoids, organ-on-chip), and hopefully, in future, into the regulatory phase as well.

What is the role of transporters in DILI, and how should we address them?

This is within the bile efflux inhibition mechanism. The transporters that move bile from hepatocytes into the bile duct are the key mediators here. BSEP is the best-known one, and BSEP inhibition is a major contributor to drug-induced cholestasis. Some MRP transporters are also relevant. All of them can be screened for inhibition as part of a DILI risk program.

Does reactive metabolite formation always mean hepatotoxicity risk?

Well, not always – it depends. Host repair mechanisms can sometimes circumvent the injury, and if no adducts form, you might not see toxicity at all. That said, reactive metabolite formation is a flag for hepatotoxicity – I wouldn’t take it lightly if I saw that result. Treat it as a signal that needs follow-up, not a verdict on its own.

Using a tiered approach starting with 2D models – how relevant are HepG2 and other cell lines? Are these models equally good?

There are real differences between cell lines. HepG2 versus HepaRG, for example: HepaRG has more metabolic capacity, so you’ll see metabolite-mediated effects that HepG2 would miss. The more physiological the metabolism, the better the model – so cell lines are fine for initial screening, but as you progress, HepaRG or primary human hepatocytes are the better choice.

How can mitochondrial toxicity and oxidative stress contribute to preclinical DILI signals?

Mitochondrial toxicity and oxidative stress are among the highest-incidence mechanisms behind DILI – they show up in most hepatotoxic cases. That alone makes them worth assessing early. There are several well-established assays: imaging-based readouts of mitochondrial membrane potential (using probes like TMRM or MitoTracker), the glucose/galactose (Glu/Gal) assay (which forces cells to rely on oxidative phosphorylation, exposing mitochondrial toxicants that would otherwise be masked by glycolysis under the Warburg effect), oxygen consumption rate measurements, and GSH/GSSG ratios for oxidative stress. These are important mechanisms to characterize.

If you want to go deeper on the in vitro toxicology assays referenced here – from the mechanism schematics to the specific cell models and endpoints – Admescope’s in vitro toxicology ebook covers the same territory in detail.

References