What a liver flag is actually worth: tested on recent clinical discontinuations

Any hepatotoxicity model can look impressive if it calls everything dangerous. The only way to learn whether a liver flag is worth anything is to make it stay quiet on safe drugs and still fire on the ones that failed. So we ran our liver model in both directions at once: forwards, on three investigational drugs discontinued after liver findings, and backwards, on eight widely prescribed medicines with clean liver records. The backwards half is the one worth reading.

Why specificity is the hard half

Drug-induced liver injury remains one of the most common safety reasons programmes stop, and the idiosyncratic kind often appears only at larger exposures or longer dosing. Experienced teams with full preclinical packages still meet it in the clinic. That makes it a fair test for a structure-based model: not whether anyone missed something, but whether a fast, design-stage read adds an independent signal while the molecule is still cheap to change.

The trap in publishing a test like this is choosing only the positives. A single binary hepatotoxicity classifier tuned for sensitivity will flag a large share of ordinary drug-like molecules, and will therefore score three for three on any set of failures you hand it while being useless in a real programme. So this note reports both halves of the confusion matrix, including the approved drugs we get wrong.

The positives: three recent discontinuations, scored from structure alone

We picked recent discontinuations where the liver finding is publicly documented and the molecule is structurally unfamiliar to the model, measured as nearest-neighbour Tanimoto similarity to the training set.

• Lotiglipron, an oral GLP-1 receptor agonist for obesity and type 2 diabetes. Pfizer ended development in June 2023 based on Phase 1 drug-drug interaction data together with transaminase elevations also seen in Phase 2, with no liver failure reported. Merged liver model: HIGH. Single heads: hepatic 0.95, DILI 0.56, BSEP 0.57. Nearest-neighbour Tanimoto 0.47, applicability domain borderline.
• Zelnecirnon, an oral CCR4 antagonist for atopic dermatitis and asthma. The FDA placed the Phase 2 trials on clinical hold in February 2024 after a single serious liver event in one participant; the sponsor cited confounding factors, no toxicity in any other participant or in nonclinical studies, and stopped the programme in November 2024. Causality was never established. Merged liver model: HIGH. Single heads: hepatic 0.18, DILI 0.79, BSEP 0.50. Nearest-neighbour Tanimoto 0.34, out of distribution.
• TNG348, an oral USP1 inhibitor for HRD-positive solid tumours. The Phase 1/2 trial was terminated in May 2024 after grade 3 and 4 liver function abnormalities in patients who remained on study beyond eight weeks. Merged liver model: FLAG FOR ASSAY, elevated but below the HIGH threshold. Single heads: hepatic 0.65, DILI 0.99, BSEP 0.52. Nearest-neighbour Tanimoto 0.58, applicability domain borderline.

These outcomes postdate the model. The liver model’s training data was frozen before June 2023, ahead of all three discontinuations, the earliest of which was announced that month. Neither the clinical findings nor the decisions to stop these programmes could have reached the training labels, so the structural distance above measures what it appears to measure.

The negatives: eight approved drugs with clean liver records

A flag is only worth having if it stays quiet on safe chemistry. We ran eight widely prescribed medicines with clean or no-concern liver records in the DILIrank and LiverTox references through exactly the same endpoints: metformin, atenolol, lisinopril, cetirizine, famotidine, amlodipine, sitagliptin and hydrochlorothiazide. These sit in or near the training distribution, so this panel measures specificity, not held-out recall.

On any single head, specificity is poor, and we will not hide it. Six of the eight safe drugs cross the 0.5 flag line on at least one individual head. Famotidine reads 1.00 on the hepatic head and 0.98 on the DILI head; amlodipine reads 0.95 and 0.94. Both score as high as the three discontinued candidates. A high reading on any one head is therefore not, by itself, evidence of a liability. An individual binary head is sensitive and over-calls ordinary drug-like molecules. This is a known and documented property of single hepatotoxicity classifiers, not a quirk of ours, and it is the reason a per-head screenshot from any vendor should be read sceptically.

The result: the number we stand behind is the merged one

Our production liver model, part of the Gnosis III safety suite, unifies the hepatic and DILI reads into a single model, gates it by applicability domain, and requires concurring mechanisms rather than one high head. On the panel above it returns LOW for seven of the eight safe drugs, with one honest false positive, sitagliptin. On the same model, the three discontinued candidates read HIGH for lotiglipron and zelnecirnon and flag-for-assay for TNG348.

• Seven of eight approved drugs are correctly LOW on the merged model.
• Six of eight approved drugs trip at least one single head.
• Three of three discontinued candidates are elevated: two HIGH, one flag-for-assay.

Read the three numbers together. The middle one is what makes the other two mean anything. Eleven compounds is a small panel, the negatives are in-distribution and the positives are not, and a single false positive out of eight is not a specificity estimate. It is an illustration. We are not claiming a validated error rate.

The harder test: a safe drug from the same novel chemistry

The eight approved drugs above sit inside the training distribution. The sterner test is a drug that is both structurally novel and liver-clean, in the same out-of-distribution regime as the three discontinuations but with no liver finding. Orforglipron, Eli Lilly’s oral GLP-1 receptor agonist, is exactly that: a different chemotype from lotiglipron and danuglipron, out of distribution to the model at nearest-neighbour Tanimoto 0.20, and still advancing in Phase 3 with no hepatic safety signal reported to date.

Its individual heads are as loud as the failures: DILI 0.84, BSEP 0.77, hepatic 0.65. The merged model does not follow them. In the out-of-distribution regime it reserves a HIGH call for two or more concurring mechanisms, and orforglipron raises only one, high lipophilicity, so it returns FLAG FOR ASSAY: elevated, not cleared, not condemned.

Read against the failures, HIGH fired on the two discontinuations that actually failed for liver, lotiglipron and zelnecirnon, each with two concurring mechanisms, and not on the advancing safe comparator. Orforglipron lands in the same flag-for-assay bucket as TNG348, which did fail, so the model is not claiming to tell a safe novel drug from a toxic one on a single signal. In the regime where it cannot be sure, it defers to the assay rather than guess, and keeps its confident HIGH call off the drug that is, so far, fine.

The caveat matters: no signal to date is not the same as safe, orforglipron is still in trials, and flag-for-assay is not a clean pass. This is one compound, shown because it is the hardest fair comparator available, a novel-scaffold drug from the same class as the positives.

From a verdict to a place on the molecule

A verdict is more useful to a chemist when it points somewhere. Gnosis Lens traces the structural hepatic signal back onto the atoms that most influence it: a spatial, testable hypothesis about where a liability might sit, not a claim about the mechanism in patients.

One honest distinction holds this together. The attribution is of a single structural head, the same kind of head we just cautioned against trusting as a standalone verdict. A head that over-calls as a global classifier can still usefully localise which substructure moves its own score, so the map is a design-stage hypothesis about where a signal concentrates, not the verdict and not a mechanism. The merged verdict for lotiglipron is set elsewhere, by two concurring out-of-distribution mechanisms, lipophilicity and a reactive-metabolite signal, not by this head’s probability. What reaches the chemist is therefore two complementary reads, kept separate: a region on the molecule to work on, and the mechanism concurrence that sets the verdict we stand behind.

Limits, stated plainly

• Small, retrospective and chosen by us. Three positives and eight negatives, selected after the outcomes were known. This illustrates behaviour; it is not a validation study and it yields no error bars.
• The two halves are not comparable. The discontinued candidates are held out by structure, Tanimoto 0.34 to 0.58; the approved drugs are mostly in-distribution, up to 1.00. The panel measures specificity on familiar chemistry and recall on unfamiliar chemistry, and neither number transfers to the other setting.
• One out-of-distribution safe point is not a specificity estimate. Orforglipron shows the merged model withholding HIGH on unfamiliar safe chemistry, but one compound illustrates behaviour; it does not measure out-of-distribution specificity.
• We left out an easier case. Danuglipron, a better-known oral GLP-1 discontinuation, is excluded because its scaffold is close enough to the training set that a correct call would show recall rather than prediction.
• The verdicts are uncalibrated. The merged model is decision-only, without a calibration gate, so HIGH is a triage decision and not a probability you should put in a risk model.
• Structure cannot see everything. Idiosyncratic injury depends on dose, exposure, duration and patient factors that a structure-only model has no access to. TNG348’s findings appeared beyond eight weeks of dosing, and nothing in a structure predicts that timing.
• No hindsight criticism is intended. Each sponsor ran extensive preclinical safety work and stopped responsibly when signals appeared. Our interest is in adding an inexpensive, early, independent read, not in second-guessing those decisions.

Why it matters

Earlier is cheaper. A structure-level read costs minutes, while a halted clinical programme costs years and exposes participants to risk. The read also gives a direction, not just a number: mechanism-specific heads and atom-level attribution tell chemists where to look. And it reports what it does not know, because every call carries an applicability-domain reading, so novel chemistry is treated as novel and a confident-looking single head is treated as exactly what it is.

A hepatotoxicity model earns its place not by catching bad drugs, which is easy, but by staying quiet on the good ones. The number we stand behind is the one that manages both, and says so when it cannot.

Vecentra Research

Sources

• Pfizer. Update on the oral GLP-1 receptor agonist clinical development programme for adults with obesity and type 2 diabetes. Press release, 26 June 2023.
• Cronenberger et al. Assessment of the drug-drug interaction potential of lotiglipron (PF-07081532). Journal of Clinical Pharmacology, 2026.
• RAPT Therapeutics. Clinical hold on studies evaluating zelnecirnon, February 2024, and discontinuation of the programme, 11 November 2024.
• Tango Therapeutics. Discontinuation of the TNG348 programme, 23 May 2024.
• Pfizer. Discontinuation of danuglipron development, 14 April 2025.
• DILIrank dataset, US FDA National Center for Toxicological Research.
• LiverTox: Clinical and Research Information on Drug-Induced Liver Injury, NIDDK.

Structures and clinical outcomes are taken from public company disclosures and peer-reviewed publications. Compound names are used for identification only, and Vecentra has no affiliation with Pfizer, RAPT Therapeutics, Eli Lilly or Tango Therapeutics. Where individual heads and the merged model disagree, the merged model is the read we stand behind. No Vecentra pipeline compounds, generated structures or implementation details are disclosed in this note.