DILI: the hepatotoxicity liability you can't fully design away

Drug-induced liver injury is the single most common reason a drug gets a black-box warning or is pulled from the market. Unlike hERG, where one well-understood pharmacophore explains most of the risk, DILI is genuinely hard: part of it is predictable chemistry, and part of it depends on the patient. Knowing which part is which is the difference between catching a liability early and discovering it in Phase III.

Two kinds of liver injury

Intrinsic DILI is dose-dependent, reproducible, and predictable. Give enough of the drug and essentially everyone gets injured. Acetaminophen is the archetype: above a threshold dose, the reactive metabolite NAPQI overwhelms hepatic glutathione and the damage is mechanistic and consistent. Intrinsic toxicity is the easier problem because it shows up in standard preclinical tox.

Idiosyncratic DILI is the one that ends programs. It is rare, often delayed by weeks to months, and depends heavily on the individual patient. It frequently does not appear in animal studies or in the first few hundred patients, then surfaces once a drug reaches a wide population. Because it is host-dependent, you cannot fully design it out of a molecule the way you can dial down a hERG signal.

What actually causes idiosyncratic DILI

The mechanisms stack rather than compete. Most idiosyncratic DILI is thought to start with a reactive metabolite: the parent drug is bioactivated, usually by cytochrome P450 enzymes, into a species that covalently binds cellular proteins and forms drug-protein adducts. Those adducts plus other cell stress feed into a handful of downstream insults:

• Mitochondrial dysfunction — impaired ATP production, oxidative stress, and a hepatocyte that can no longer keep up with its energy demands.
• BSEP inhibition — blocking the bile salt export pump lets bile acids accumulate inside the hepatocyte, which is itself a stressor. Drugs that hit both BSEP and mitochondria tend to carry the more serious DILI risk.
• Adaptive immune activation — the drug-protein adduct is presented as a neoantigen, and in susceptible patients cytotoxic T cells do the actual tissue damage. This is why specific HLA haplotypes are associated with DILI from specific drugs, and why two patients on the same drug can have completely different outcomes.

That last point is the crux: the immune and host-genetic component is why no purely structural model will ever fully predict idiosyncratic DILI. The chemistry is necessary but not sufficient.

The rule of two, and why kinase inhibitors are exposed

The most useful first-pass heuristic comes from Chen, Borlak, and Tong’s 2013 analysis: a drug given at 100 mg/day or more AND with logP of 3 or higher carries a roughly 14-fold higher odds of hepatotoxicity. They called it the “rule of two.” The logic is intuitive once you see it: a high daily dose means more parent drug for the liver to bioactivate, and high lipophilicity means more of it partitions into hepatocytes and gets metabolized in the first place.

Oncology kinase inhibitors sit almost exactly in the danger zone. They are routinely dosed at hundreds of milligrams a day, and their chemistry pushes logP into the 3-5 range — basic amines for solubility, aromatic rings for the ATP-pocket hinge, the same features that make them rule-of-two positive. The hepatotoxicity warnings on lapatinib, pazopanib, regorafenib, ponatinib, and idelalisib are not bad luck; they are the predictable cost of that physicochemical profile. Sotorasib, the KRAS G12C inhibitor, ships with mandatory AST/ALT monitoring for the same reason.

What the prediction tools can and can't do

The reference dataset is the FDA’s DILIrank: roughly a thousand drugs sorted into Most-, Less-, No-, and Ambiguous-DILI-concern categories based on weighed clinical evidence. Nearly every machine-learning DILI predictor is trained on it or on datasets derived from it. Layered on top are structural alerts for reactive-metabolite-forming groups, and the rule-of-two physicochemistry, which you can read off any SMILES string for free.

ML predictors output a continuous DILI probability and do a respectable job recovering the chemistry-driven signal. What they cannot do is see the patient: the HLA association, the prior inflammatory state, the co-medication that induces the wrong P450. Treat a model’s DILI score as a flag for “this molecule has hepatotoxic chemistry,” not as a verdict, and pair it with the rule-of-two read on dose and logP. In a molecular docking workflow, that means running the ADMET screen on every pose you keep as a candidate, not just on the final pick.

Try the prediction yourself

Liganx’s ADMET panel runs the admet-ai ensemble on every compound after a successful molecular docking run, and DILI is one of the three risk dots in the violet ⚕ ADMET pill — emerald for low, amber for medium, rose for high — alongside hERG and CYP3A4. Open Studio on Liganx, the free molecular docking online platform, and run molecular docking on any candidate, then open the ADMET panel on a result row. Read the DILI probability together with the molecule’s logP: a high DILI score on a lipophilic compound you also intend to dose at hundreds of milligrams a day is the rule of two and the model agreeing with each other, and that is a redesign signal worth taking seriously before you commit synthesis budget.

Primary sources

• Chen M, Borlak J, Tong W. High lipophilicity and high daily dose of oral medications are associated with significant risk for drug-induced liver injury. Hepatology 58, 388-396 (2013). doi:10.1002/hep.26208
• Chen M, et al. DILIrank: the largest reference drug list ranked by the risk for developing drug-induced liver injury in humans. Drug Discov Today 21, 648-653 (2016). doi:10.1016/j.drudis.2016.02.015
• Kullak-Ublick GA, et al. Drug-induced liver injury: recent advances in diagnosis and risk assessment. Gut 66, 1154-1164 (2017). doi:10.1136/gutjnl-2016-313369