Drug-Induced Liver Injury (DILI)

Modern DILI models move beyond static PAINS or Brenk filters to predict the formation of reactive metabolites. They incorporate quantum mechanical calculations or deep learning to assess a molecule's propensity for bioactivation by CYP450 enzymes into electrophilic species. These reactive metabolites can covalently bind to hepatic proteins, forming drug-protein adducts that trigger downstream toxicity cascades, a mechanism not captured by simple 2D fingerprint similarity.

Hepatotoxicity is not a single endpoint but a convergence of mechanisms. A robust DILI model integrates predictions for multiple initiating events:

• Mitochondrial Toxicity: Uncoupling of oxidative phosphorylation or inhibition of the electron transport chain.
• BSEP Inhibition: Blockade of the Bile Salt Export Pump, leading to cholestatic hepatocyte damage.
• Lysosomal Trapping: Accumulation of phospholipids, causing phospholipidosis.
• Oxidative Stress: Generation of reactive oxygen species (ROS) depleting glutathione.

State-of-the-art DILI prediction relies on fusing disparate data modalities. A single model might ingest high-content screening (HCS) imaging data from hepatocytes, transcriptomic signatures from toxicogenomics databases like TG-GATEs, and chemical structure information. Multi-task or multi-modal neural networks learn a shared latent representation that correlates structural features with phenotypic cellular responses, providing a holistic view of a compound's hepatotoxic potential.

To overcome species-specific translational gaps, models are trained on data from advanced human-relevant culture systems. These include 3D hepatic spheroids, microphysiological systems (MPS) like liver-on-a-chip, and sandwich-cultured human hepatocytes. These systems maintain metabolic competence and transporter expression for longer durations, enabling the detection of chronic toxicity signals that are invisible to simpler immortalized cell lines.

QST models are mechanistic, mathematical models that simulate the biochemical interaction networks within a hepatocyte. Instead of a black-box prediction, a QST model computationally represents pathways like glutathione depletion, mitochondrial respiration, and apoptotic signaling. By inputting a compound's physicochemical properties and reactivity data, the model simulates the time-resolved cellular response, predicting whether homeostatic defenses will be overwhelmed.

Given the catastrophic cost of late-stage attrition, a DILI prediction without a confidence interval is of limited value. Advanced models employ conformal prediction to provide prediction sets with a guaranteed error rate, or Bayesian neural networks to distinguish between epistemic uncertainty (model ignorance due to sparse data) and aleatoric uncertainty (inherent biological noise). This allows a CTO to make a risk-informed decision, not just a binary classification.

A primary mechanism of DILI where a drug is bioactivated by CYP450 enzymes into a chemically reactive, electrophilic species. These metabolites can form covalent adducts with hepatic proteins and DNA, disrupting critical cellular functions and triggering downstream immune responses or direct mitochondrial toxicity. Key structural alerts include anilines, thiophenes, and acyl glucuronides. Predicting the formation of these species is a central goal of in silico DILI models.

A critical sub-mechanism of DILI involving the impairment of hepatocyte mitochondria. Drugs can directly uncouple oxidative phosphorylation, inhibit the electron transport chain (Complexes I-V), or open the mitochondrial permeability transition pore (mPTP). This leads to ATP depletion, release of pro-apoptotic factors like cytochrome c, and excessive generation of reactive oxygen species (ROS). In vitro assays like the Seahorse XF Analyzer measure oxygen consumption rate (OCR) to quantify this liability.

BSEP, encoded by the ABCB11 gene, is the primary transporter responsible for excreting bile salts from hepatocytes into the bile canaliculi. Inhibition of BSEP by a drug or its metabolite causes intracellular accumulation of cytotoxic bile acids, leading to cholestatic liver injury. This is a well-validated mechanism for DILI, and in silico models often include a specific BSEP inhibition classifier alongside in vitro vesicular transport assays to flag this risk early in discovery.

No single assay predicts DILI. A holistic risk assessment relies on a panel of in vitro tests that probe different mechanisms:

• Cytotoxicity: High-content imaging (HCI) in HepG2 or primary human hepatocytes to measure cell loss.
• Mitochondrial Toxicity: Seahorse assay for OCR and JC-1 staining for membrane potential.
• BSEP Inhibition: Vesicular transport assay using taurocholate.
• Reactive Metabolites: Trapping studies with glutathione (GSH) or dansyl glutathione.
• 3D Spheroid Models: Long-term culture systems that better recapitulate in vivo physiology.

A simple, interpretable clinical chemistry heuristic for identifying drugs with high DILI risk. The Rule-of-Two states that a drug is at elevated risk if it has both:

• Daily dose ≥ 100 mg
• LogP ≥ 3 This rule captures the interplay between dose (total body burden) and lipophilicity (propensity for metabolic bioactivation and off-target binding). While highly specific, it lacks sensitivity, and modern machine learning models significantly outperform it by incorporating structural and mechanistic features.

Curated reference datasets are the foundation for training supervised DILI prediction models. DILIst (FDA) provides a binary classification of drugs based on their clinical hepatotoxicity potential. DILIrank expands on this by providing a ranked severity score. These datasets compile data from FDA drug labels, post-market surveillance, and published case reports. A key challenge in using them is the inherent label noise and class imbalance, as severe hepatotoxins are rare compared to safe drugs.