Amiodarone has saved countless lives as one of the most effective antiarrhythmic drugs available to cardiologists, yet its long-term use carries a shadow: a potentially irreversible scarring of the lungs known as amiodarone-induced pulmonary fibrosis, or AIPF. The condition can emerge months or even years into therapy, and clinicians currently lack reliable early biomarkers that would allow them to detect the damage before it becomes permanent. A new study published in BMC Pharmacology and Toxicology by Hua Sheng, Xingang Lu, and YunTao Lu set out to change that, deploying an unusually ambitious combination of computational biology and laboratory experiments to identify the genes that may drive the disease process.
The team’s starting point was network toxicology, a discipline that treats the interaction between a drug and the body as a web of molecular relationships rather than a single linear pathway. By mapping the known protein targets of amiodarone against genes already associated with pulmonary fibrosis, the researchers identified eighty-four overlapping genes that sit at the intersection of the drug’s pharmacology and the pathology of scarring. When these genes were subjected to enrichment analysis, two biological themes emerged with striking clarity: necroptosis, a form of programmed inflammatory cell death, and the wingless/integrated, or Wnt, signaling pathway, a developmental cascade long implicated in fibrotic remodeling of tissue. Both processes are plausible mechanistic bridges between a drug accumulating in lung tissue and the progressive deposition of scar collagen.
What distinguishes this study from earlier network-based analyses of AIPF is the sheer scale of the machine learning framework the authors built on top of that gene list. Rather than relying on a single algorithm, they integrated 127 different algorithms into a consensus pipeline, then cross-validated the results against independent datasets drawn from the Gene Expression Omnibus, a public repository of gene expression data. The logic is straightforward: a candidate biomarker that survives scrutiny across many analytical approaches and multiple patient cohorts is far less likely to be a statistical artifact than one identified by a single method. This kind of ensemble strategy, borrowed from the competitive machine learning world, is increasingly seen as the gold standard for extracting trustworthy signals from the noisy, small-sample datasets that dominate clinical genomics.
The consensus winner was SERPING1, a gene encoding the C1 inhibitor protein best known for its role in regulating the complement and contact cascades of the immune system. The pipeline elevated SERPING1 to the status of primary core biomarker, while a second gene, lipocalin 2 (LCN2), emerged as a secondary candidate. LCN2, which encodes a protein involved in iron trafficking and inflammatory responses, showed a more complicated behavior across datasets, exhibiting prominent inter-cohort spatial heterogeneity — in plain terms, its expression patterns varied noticeably from one patient cohort to another, a warning sign the authors were careful not to gloss over.
To test whether these computational candidates could physically interact with the drug, the researchers turned to molecular docking and molecular dynamics simulation. Docking algorithms predict how a small molecule like amiodarone might fit into the structural pockets of a protein, while dynamics simulations then let that predicted complex flex and move over time, revealing whether the binding is stable or falls apart under thermal motion. Both SERPING1 and LCN2 proteins were found to bind stably to amiodarone in these simulations, a result that lends biophysical plausibility to the idea that the drug could directly perturb the function of these proteins in lung tissue, not merely alter their expression indirectly through cellular stress.
The next layer of evidence came from single-cell RNA sequencing and spatial transcriptomics, two of the most powerful tools in modern genomics. Single-cell sequencing breaks a tissue down into its constituent cell types and measures gene activity in each one, while spatial transcriptomics preserves information about where in the tissue architecture those genes are active. The analyses showed that both SERPING1 and LCN2 were expressed predominantly in cell populations associated with lung injury, placing the candidate biomarkers in exactly the cellular neighborhoods where fibrotic damage unfolds. This kind of multiomics localization matters because a biomarker measured in bulk tissue can be misleading if the signal actually originates from a minor cell population, such as infiltrating immune cells, rather than from the epithelial cells that orchestrate fibrosis.
The spatial data also delivered one of the study’s most nuanced findings. SERPING1 expression showed a weak positive correlation with the extent of fibrosis across tissue samples, a modest but consistent relationship. LCN2, by contrast, displayed variable, cohort-dependent correlation patterns — its relationship to fibrosis shifted depending on which dataset was examined. The authors interpret this honestly: LCN2 remains an interesting candidate, but its inconsistency across cohorts means it cannot yet be considered a dependable marker, and it awaits further experimental validation before it can be pursued clinically.
Crucially, the study did not stop at computation. The team constructed an in vitro model of AIPF by exposing A549 cells, a widely used human lung epithelial cell line, to amiodarone in the laboratory. When they measured gene expression in these drug-exposed cells, they confirmed a dose-dependent downregulation of SERPING1 — the higher the amiodarone concentration, the lower the gene’s activity. This laboratory confirmation is the study’s strongest single piece of evidence, because it demonstrates that a biologically relevant concentration of the drug can directly suppress SERPING1 expression in the very epithelial cells that line the lung alveoli and are central to the fibrotic response.
The authors are careful to frame their contribution with appropriate modesty. They note that the findings do not establish entirely new drivers of the disease; rather, they extend previous network-based studies of AIPF by adding multiomics localization evidence and in vitro support, particularly for the downregulation of SERPING1 in amiodarone-exposed epithelial cells. That distinction is important in a field where computational screens sometimes generate long lists of speculative targets that never survive contact with experimental reality. By anchoring their top candidate in molecular simulation, tissue-level spatial data, and a living-cell model, the researchers have given SERPING1 a far more solid evidentiary foundation than most computationally nominated biomarkers enjoy.
Even so, the road from candidate gene to clinical biomarker is long. The authors explicitly acknowledge that validation in animal models and in clinical samples from actual AIPF patients is still needed before SERPING1 could inform the care of patients taking amiodarone. If that validation succeeds, the implications could be substantial: a simple measure of SERPING1 expression or C1 inhibitor levels might one day allow physicians to monitor patients on long-term amiodarone therapy and intervene before irreversible scarring takes hold. In the meantime, the study stands as a compelling demonstration of how network toxicology, ensemble machine learning, molecular simulation, single-cell and spatial transcriptomics, and classical cell biology can be woven into a single pipeline — a template that other researchers hunting for drug-toxicity biomarkers are likely to follow.
Subject of Research: Identification of key genes involved in amiodarone-induced pulmonary fibrosis using integrated multiomics and in vitro validation
Article Title: Integrated multiomics methodology and in vitro experiments for the detection of key genes involved in amiodarone-induced pulmonary fibrosis
Article References: Sheng, H., Lu, X., & Lu, Y. (2026). Integrated multiomics methodology and in vitro experiments for the detection of key genes involved in amiodarone-induced pulmonary fibrosis. BMC Pharmacology and Toxicology. https://doi.org/10.1186/s40360-026-01238-5
Image Credits: AI Generated
DOI: 10.1186/s40360-026-01238-5
Keywords: amiodarone, pulmonary fibrosis, SERPING1, LCN2, network toxicology, machine learning, molecular docking, molecular dynamics simulation, single-cell RNA sequencing, spatial transcriptomics, biomarkers, drug-induced lung injury
Cite Scienmag News
APA MLA Chicago
Juliet Wilcox. (September 26, 2026). Machine Learning Hunt Pinpoints Gene Linked to a Heart Drug’s Lung Damage. Scienmag. https://scienmag.com/machine-learning-hunt-pinpoints-gene-linked-to-a-heart-drugs-lung-damage/
Juliet Wilcox. “Machine Learning Hunt Pinpoints Gene Linked to a Heart Drug’s Lung Damage.” Scienmag, 26 September 2026, https://scienmag.com/machine-learning-hunt-pinpoints-gene-linked-to-a-heart-drugs-lung-damage/. Accessed 26 September 2026.
Juliet Wilcox. “Machine Learning Hunt Pinpoints Gene Linked to a Heart Drug’s Lung Damage.” Scienmag. September 26, 2026. https://scienmag.com/machine-learning-hunt-pinpoints-gene-linked-to-a-heart-drugs-lung-damage/
Copy citation Download RIS
Tags: amiodaroneamiodarone-induced lung damagebioinformatics for adverse drug reactionsBiomarkerscomputational biology in cardiologydrug-induced lung injuryearly detection of drug-related lung injurygene biomarkers for pulmonary fibrosisgene-environment interactions in drug toxicitygenetic factors in drug-induced fibrosislaboratory validation of toxicity-related genesLCN2Machine learningmachine learning in drug toxicity predictionmolecular dockingmolecular dynamics simulationmolecular pathways in pulmonary scarringnetwork toxicologynetwork toxicology for drug safetypulmonary fibrosisresearch on antiarrhythmic drug side effectsSERPING1Single-Cell RNA SequencingSpatial transcriptomics
