Can THC-Related Liver-Cell Gene Changes Prove Cannabis-Induced Hepatitis? A 2025 Primary Study

Quick answer: No. A 2025 experimental paper by Zou and colleagues identified 22 candidate proteins through database overlap, calculated potential THC interactions with 17 of them, and observed altered messenger-RNA expression for ten genes in a cultured human liver cell line exposed to nominal THC concentrations of 1 or 10 micromolar for 48 hours. Those are meaningful mechanistic research observations, not evidence that a particular person developed hepatitis, liver failure, or an elevated liver enzyme because of cannabis. The investigators did not administer THC to human participants, collect patient blood or liver specimens, measure AST or ALT responses, demonstrate histologic liver injury, or establish an individual concentration–effect threshold. Their own limitations expressly identify the absence of protein experiments and animal validation.

The question is especially important when a forensic file places a positive cannabinoid test next to abnormal liver chemistry. A molecular signal in cultured cells and a patient’s clinical disease are not interchangeable endpoints. The original study is Zou T, Zhang R, Hu L, and colleagues, Uncovering mechanism of hepatotoxicity diseases caused by tetrahydrocannabinol based on novel network toxicology and experimental verification, Scientific Reports 15, 13712 (2025), DOI 10.1038/s41598-025-97523-0.

What did the investigators actually test?

The study combined three distinct levels of evidence. First, computer-assisted target databases proposed which proteins might interact with THC and which proteins appear in hepatotoxicity-associated disease records. Second, software predicted how THC might fit within selected protein structures. Third, the investigators treated one commercially obtained human liver-derived cell line, L-02, with THC and used reverse-transcription quantitative polymerase chain reaction (RT-qPCR) to measure relative RNA expression. Each level answers a different scientific question. A database match proposes a candidate; a docking score estimates a possible molecular configuration; altered RNA abundance records a laboratory response under a chosen cell-culture condition.

None of those steps is a controlled patient-outcome experiment. No persons were randomized to THC exposure, no clinical diagnosis was independently adjudicated, and no human intoxication, driving, injury, or work-performance endpoint was observed. A title containing the term “hepatotoxicity” should not be mistaken for a demonstrated clinical incidence rate or for proof of a liver-disease mechanism in people.

How were 22 candidate liver-associated THC targets identified?

The researchers assembled 41 THC-related targets from DrugBank and 54 from SwissTargetPrediction. After removing four duplicates they retained 91 candidate THC targets. Separately, hepatotoxicity-related database searches produced 414 DisGeNET entries, 356 GeneCards entries, and six Therapeutic Target Database entries; deduplication left 689 candidate disease targets. The intersection was 22 proposed common targets. These are counts of database records and proteins, not 91 subjects, 689 liver patients, or 22 demonstrated THC injuries.

A subsequent STRING protein–protein interaction network used 20 nodes and 73 edges, with average node degree reported as 6.95. The fact that 22 candidates reduced to a 20-node network is a feature of that specific model, not a loss of two human outcomes. Network degree identifies connectivity within the chosen candidate set; it is not a clinical effect size. MAPK14, PPARG, PTGS2, EGFR, and other proteins ranked within the modeled network because of the database and computational relationships selected by the authors.

Which source results were computed, and which were physically measured?

The most reliable way to read the paper is to keep the data source attached to each reported number. The following table preserves the researchers’ principal denominators while separating predicted targets, model structures, and measured cellular RNA.

Did the toxicology prediction programs agree that THC causes drug-induced liver injury?

No. An important detail in the original Table 2 is that the endpoints did not all point in the same direction. The ADMETlab 2.0 output labeled human hepatotoxicity with “+++,” while its separate drug-induced liver injury output was “−.” Respiratory toxicity was also labeled “+++,” while carcinogenicity was “−.” The table’s legend maps “−” to a score band of 0.3–0.5 and “+++” to 0.9–1.0. These are program-specific prediction categories, not observed rates of human liver disease or probabilities that any individual event occurred.

This internal divergence matters. One should not quote the most alarming predicted category and omit the negative prediction on a closely related endpoint. Equally, the negative computational category is not clinical proof of no possible risk. Different endpoints, training data, and algorithms may yield different predictions. The measured human-cell results must be assessed separately from either prediction.

What happened to cultured liver cells exposed to THC?

Investigators obtained L-02 human liver cells, maintained them in standard supplemented growth medium, and exposed cultures to THC dissolved in dimethyl sulfoxide (DMSO). The stated final THC treatment levels were 1 and 10 µM, with a 48-hour incubation and a 0-µM reference. The paper reports that samples were performed three times in parallel, but it does not provide a clearly documented independent donor population or independent biological replicate count that can be treated as human n. The repeated qPCR measurements cannot be turned into three separate people or three separate liver-injury events.

RNA was extracted with TRIzol and quantified with a QuantStudio 5 RT-qPCR workflow. ACTB was used as the reference gene; the authors reported expression via the 2^−ΔΔCq calculation. That relative expression readout compares target transcripts with a reference under experimental conditions. It is not a direct concentration of THC, a quantity of target protein, an activity assay for a protein, a validated clinical diagnosis, or a serum enzyme measurement. The paper does not give a human plasma concentration–time curve corresponding to those culture conditions.

Which ten genes changed, and which direction did they move?

The experimental Figure 6 indicates that nine evaluated targets had lower relative messenger-RNA expression after THC treatment and that KDR had higher expression. The caption uses threshold symbols rather than furnishing an exact numerical p value and confidence interval for each comparison. Therefore a careful account preserves the direction and significance-threshold convention without inventing exact per-gene statistics or claiming the plotted bar heights are independently validated fold changes.

For Figure 6 the legend denotes * for p<0.05, ** for p<0.01, and *** for p<0.001. The figure provides separate 0-, 1-, and 10-µM bars, but these significance thresholds do not produce a validated dose–response function for clinical liver disease. No confidence intervals for a patient injury risk were reported. An expert should also avoid representing all evaluated target genes as significant merely because a docking analysis selected them.

What does molecular docking contribute?

SwissDock evaluated computationally constructed interactions between THC and 17 target protein structures. The reported estimated free-energy values included approximately −7.985 kcal/mol for CFTR, −7.898 kcal/mol for CXCR2, −7.845 kcal/mol for TYMS, and −7.759 kcal/mol for PPARG. More negative modeled energy in this computational comparison suggests a more favorable predicted pose; it is not a measured dissociation constant in living tissue, a confirmed THC–protein bond inside a patient, or an estimate of the chance of hepatitis. Five additional candidate targets yielded no usable docking site in the stated workflow.

The paper’s docking section describes hydrogen-bond patterns for ERBB2, GPX1, MAPK14, NR1H4, and SOD1. That is a computational structural result, not a biochemical demonstration that those five proteins were bound by THC in the L-02 cultures. A docking model may be useful for prioritizing assays, but direct target engagement requires separate experimental confirmation. A person’s exposure history and liver-disease diagnosis cannot be reconstructed from a SwissDock energy table.

Does a significant gene-expression result prove damaged hepatocytes?

It does not. Transcript changes can be compatible with cellular response, stress, adaptation, altered signaling, or other mechanisms; interpretation depends on complementary endpoints. In this experiment there was no reported direct assay of liver-cell necrosis, histopathologic grading, organ-level cholestasis, human hepatitis, or biopsy-confirmed fibrosis. The paper’s own stated limitations acknowledge that the proposed targets were evaluated with simulations and in-vitro messenger-RNA analysis, without protein experimentation or in-vivo validation. This absence is not a minor technicality: transcription, translation, enzyme activity, tissue injury, and clinical organ dysfunction are successive biological levels that cannot simply be substituted for one another.

For example, SOD1 is involved in oxidative-stress biology, but a decrease in SOD1 transcript abundance in a cultured cell does not automatically measure loss of SOD1 protein activity or tissue oxidative damage. NR1H4 may be relevant to bile-acid signaling, but lower NR1H4 RNA is not a directly observed bile duct obstruction or clinical cholestasis. KDR RNA rising does not quantify VEGF-dependent pathology. The study provides hypotheses about these pathways, not completed proof of the corresponding diseases in humans.

Why does the nominal 1–10 micromolar exposure not translate into a DUI blood result?

The authors intentionally placed nominal concentrations of THC into a culture medium. Their units (µM) describe a molecular concentration in that artificial culture environment. Forensic blood results, when available, are produced from a collected and analyzed specimen, ordinarily expressed in mass-per-volume units such as ng/mL for THC. Directly equating nominal culture concentration to venous whole-blood, serum, or plasma concentration would require a justified chemical conversion and, more importantly, a validated biological distribution, time course, protein binding, free-versus-total fraction, and exposure equivalence. This paper supplies none of those human pharmacokinetic bridges.

The study also does not establish how inhaled cannabis smoke, vaporized cannabis, an oral edible, or a prescribed preparation would generate the same intracellular exposure. Metabolism, liver first pass, transport, extracellular binding, repeated dosing, and duration could matter. Those questions cannot be answered by a 48-hour cell-culture concentration alone. Neither the nominal dose nor the gene-expression outcome is a validated behavioral impairment test.

What was absent from the actual analytical specimen record?

There was no human venous whole-blood THC measurement; no paired human serum or plasma THC measurement; no breath alcohol or breath cannabinoid measurement; and no tested human urine, oral fluid, hair, or vitreous specimen establishing drug exposure in an individual. The experimentally analyzed material was cellular RNA extracted from treated cultures. Neither a virtual protein–ligand complex nor a qPCR expression value is a forensic sample result. The study did not chemically quantitate intracellular THC or verify the nominal THC exposure over all 48 hours with a reported validated chromatographic assay.

The distinction is consistent with the site’s independently published overview of biological matrices in forensic toxicology: a specimen and its assay determine which claims can be made. Specific drug identity and quantitation require their own method validation, as explained in confirmatory drug testing. Pharmacokinetic translation is a different exercise, discussed in pharmacokinetics and interpretation. Those live contextual guides do not supply experimental data missing from Zou’s original primary study.

Could this study explain a patient’s elevated AST, ALT, or bilirubin?

Not by itself. Alanine aminotransferase (ALT), aspartate aminotransferase (AST), bilirubin, alkaline phosphatase, and other clinical markers have distinct interpretations. None was tested as a human outcome in this paper. A claim that cannabis caused a specific patient’s abnormal test would require a medically supportable temporal relationship, careful assessment of other plausible causes, the relevant laboratory record, medication and substance coexposures, clinical examination, and suitable diagnostic evidence. This experiment provides no individual likelihood ratio, attributable fraction, sensitivity, specificity, positive predictive value, or human dose–response threshold for that inference.

Even if an actual patient had both documented cannabinoid exposure and abnormal liver chemistry, temporal coincidence would not uniquely identify THC as the cause. Other drugs, alcohol exposure, viral or metabolic disease, underlying hepatobiliary conditions, ischemia, muscle injury affecting AST, and assay interpretation could complicate attribution. Those possibilities require patient-specific evidence; they are not ruled out or quantified by the paper. The work should neither be used to assert that cannabis cannot affect liver biology nor to certify that THC caused a particular hepatitis episode.

What do the negative and missing findings mean for scientific certainty?

The model’s drug-induced-liver-injury category was “−” despite a “+++” human-hepatotoxicity category. Seven of the 17 docked genes were not described as significantly altered in the qPCR results. The laboratory did not report a protein assay, direct cell-injury measure, animal confirmation, or human clinical outcome. These are distinct categories: a computational negative prediction, a statistical nonfinding for an evaluated endpoint, and an endpoint never measured at all. They must not be collapsed into one blanket assertion that either harm or safety was conclusively demonstrated.

The qPCR star thresholds address comparisons in the tested culture. They do not supply precision around a population-level hazard. Multiple candidate targets were filtered through interconnected data sources; selected pathway enrichment tests can generate attractive biological narratives, but pathway membership is not a causal mechanism by itself. Reporting every selected pathway as a proven operative mechanism would overstate both the design and the independent validation.

How should a forensic scientist explain the study to a court or decision-maker?

A concise evidence hierarchy is useful. The investigators generated computational hypotheses about THC and liver-associated proteins, calculated model docking energies for some proteins, and observed relative RNA changes in a single cultured human liver-cell system. These are laboratory and modeling findings. They do not show an observed human liver injury, prove that any measured cannabinoid concentration was the responsible dose, or distinguish THC causation from alternative clinical explanations in a real patient. The appropriate conclusion is a mechanistic research signal with substantial translational gaps.

Several cross-examination questions follow naturally from that hierarchy: Was a living person treated or sampled? What actual physical specimen contained THC, and was it analytically quantified? How many independent biological cultures, not merely technical PCR wells, were represented? Was cell death or liver-function injury assessed? Were target proteins or target engagement measured? Were predictions validated in animals or human clinical outcomes? Was a competing cause of actual hepatitis excluded? The published work answers some molecular-screening questions but cannot provide affirmative individual-causation answers to the latter questions.

What limitations did the authors themselves acknowledge?

The original Discussion identifies its work as preliminary and expressly calls for further western-blot protein experiments, in-vivo animal models, and evaluation of dose-dependent severity. That acknowledgment narrows the permissible interpretation of its mechanistic claims. The study also gives no independent person-level clinical denominator and no measured human exposure–response function. Its research value lies in suggesting which protein pathways warrant direct follow-up, not in offering a completed diagnostic method for cannabis-associated liver injury.

When this article is used in litigation, the full original figures and methods matter more than a generalized sentence in its abstract or title. The reported “22,” “17,” and “10” count different stages of evidence; the tested culture concentrations are not legal blood measurements; and significance symbols are not estimates of individual injury probability. An answer that retains those distinctions is both scientifically more useful and less vulnerable to overinterpretation.

Frequently asked questions

1. Did this study prove that smoking cannabis causes hepatitis in people?

No. It used target databases, computational docking, and laboratory L-02 liver-cell cultures. It did not expose human subjects, collect clinical liver outcomes, or compare smokers with an unexposed cohort.

2. Were the investigators testing genuine human whole-blood THC concentrations?

No. THC at nominal 1 and 10 µM was added to cell culture. There was no reported collected human venous whole-blood, serum, or plasma THC series and no validated equivalence to a person’s toxicology result.

3. What is the difference between 22 targets and ten changed genes?

Twenty-two represents an overlap of computationally assembled THC and hepatotoxicity candidate targets. Ten represents transcript targets with significant RT-qPCR changes in exposed L-02 cells. Neither number counts injured patients.

4. Was liver-cell death or enzyme leakage measured?

The reported confirmation centered on RNA expression, not a direct cell-death, ALT, AST, bilirubin, or histopathology endpoint. The authors called for further protein and in-vivo validation.

5. Does the KDR finding mean every predicted target was suppressed?

No. KDR messenger-RNA expression was higher, while nine named significant targets had lower expression. Other modeled targets lacked a statistically significant difference in the reported qPCR comparisons.

6. Does a favorable molecular docking score confirm a harmful THC–protein interaction inside a person?

No. Docking values are simulated interaction estimates. Direct target engagement, downstream protein function, actual exposure, tissue damage, and patient disease require separate empirical demonstrations.

7. Can a positive THC test plus abnormal liver enzymes establish individual causation?

No, not from this study. An actual case requires specimen-specific validated chemistry, clinical timelines, liver investigations, competing-cause assessment, and reasoned causation analysis. This original paper does not supply a clinically validated individual attribution rule.

Original primary source

Zou T, Zhang R, Hu L, Yang G, Chen X, Wang M, Cheng T, Nie S, Liu L, Hong S. Uncovering mechanism of hepatotoxicity diseases caused by tetrahydrocannabinol based on novel network toxicology and experimental verification. Scientific Reports. 2025;15:13712. doi:10.1038/s41598-025-97523-0. The analysis above distinguishes the authors’ actual database and cell-culture observations from clinical extrapolations the experiment did not test.