Machine Learning Links a Rare Protein Modification and Mast Cells to Duchenne Muscular Dystrophy

Duchenne muscular dystrophy (DMD) is one of the most devastating inherited muscle diseases, caused by mutations in the X-linked DMD gene that eliminate the structural protein dystrophin. Without dystrophin, skeletal and cardiac muscle fibers degenerate progressively, leaving children and young adults facing relentless weakness and, ultimately, heart and respiratory failure. Yet the destruction of muscle in DMD is not purely mechanical. A growing body of evidence points to two parallel contributors: chemical alterations of proteins after they are made, known as post-translational modifications, and aberrant activation of immune cells within damaged tissue. A new study published in Frontiers in Molecular Biosciences brings these two threads together in an unusually ambitious computational and experimental framework, and in doing so identifies a single gene—USP18—as a central hub connecting a little-known protein modification to mast cell-driven inflammation in dystrophic muscle.

The modification in question is isonicotinylation, abbreviated Kinic, a recently discovered post-translational mark in which an isonicotinyl group is attached to proteins. It was first described in liver cancer, where it modifies histones to boost oncogenic protein expression and activate the PI3K–AKT signaling pathway. Subsequent work showed that Kinic can also modify non-histone proteins, driving epithelial–mesenchymal transition and invasion in liver cells, and an independent artificial intelligence study suggested that Kinic signatures could support risk stratification and drug design in hepatocellular carcinoma. What remained unclear was whether this modification plays any role in the immune system—and, by extension, in immune-mediated diseases such as DMD.

Mast cells, meanwhile, have long been suspected participants in muscular dystrophy. These inflammatory cells are recruited to damaged muscle, where they can promote progressive myofiber necrosis and sustain chronic inflammation. Earlier studies in dystrophin-deficient humans, dogs, and mice, including double-mutant mdx mice with exaggerated mast cell activity, correlated mast cell burden with disease severity. But whether the balance of mast cell activity could serve as a measurable indicator for DMD patients had never been systematically tested. The research team, led by investigators from hospitals in Zhejiang and Jiangsu, China, set out to close both gaps at once by asking how Kinic-related biology and mast cell behavior intersect in dystrophic muscle.

Methodologically, the study is a tour de force of modern bioinformatics. The researchers drew on four publicly available bulk gene expression datasets from the Gene Expression Omnibus—GSE38417, GSE109178, GSE6011, and GSE32720—covering a total of 98 muscle samples from DMD patients and healthy controls. After normalizing the data and correcting for batch effects using the sva and limma packages, they applied gene set variation analysis to quantify Kinic pathway activity, the CIBERSORT algorithm to estimate immune cell infiltration, and weighted gene co-expression network analysis to identify gene modules correlated with each feature. In GSE38417, the black module correlated most strongly with Kinic activity (R = 0.49) and negatively with DMD pathology (R = −0.86); in GSE109178, the green module tracked mast cell signatures (R = −0.63) and disease status (R = 0.91). Intersecting these modules yielded nine shared genes, which the team designated KM-related differentially expressed genes—KM standing for the combined Kinic and mast cell axis.

Functional enrichment revealed that these nine genes were principally involved in maintaining intracellular homeostasis, a finding that hints at how a protein-modification program and an immune cell population might converge on a shared cellular vulnerability. Using consensus clustering on the GSE6011 training cohort, the researchers then split DMD patients into two molecular subgroups, C1 and C2, based on the expression patterns of these KM-related genes. The subgroups differed in molecular function, gene expression, and immune infiltration profiles, suggesting that the KM axis captures biologically meaningful variation among patients rather than statistical noise.

The heart of the diagnostic effort lay in an ensemble machine learning pipeline. The team systematically combined twelve machine learning algorithms with multiple feature selection and preprocessing strategies, generating 132 candidate models, each trained on GSE6011 with ten-fold cross-validation and then tested on the independent validation set GSE32720. The winning combination—random forest paired with stepwise generalized linear modeling, abbreviated RF + Stepglm—achieved favorable discrimination in both cohorts, as assessed by receiver operating characteristic and precision–recall analyses, and its accuracy was further confirmed through nomogram and calibration plots. To interpret which genes drove the model’s decisions, the researchers applied SHAP analysis, a game-theoretic approach for quantifying each feature’s contribution to predictions. The verdict was unambiguous: USP18, a gene encoding a ubiquitin-specific protease, emerged as the dominant contributor.

USP18’s credentials made the finding compelling. The enzyme is known to regulate ubiquitination and has been implicated in muscular disorders, including as an ISG15 activator promoting inflammation in dermatomyositis muscle with perifascicular atrophy and in juvenile inflammatory myopathies. In the new study, USP18 expression was consistently upregulated in DMD samples across all four bulk datasets, and single-gene pathway analysis in GSE6011 tied it to inflammatory signaling. Crucially, the team also found that high USP18 expression was negatively correlated with mast cell activation, suggesting a regulatory relationship between the two. To validate these patterns experimentally, the researchers engineered human mast cell lines using CRISPR–Cas9: one line (DMDI) carried a deletion of exon 52 of the DMD gene, and another (DMDII) carried a stop codon introduced within exon 52. Quantitative PCR and Western blotting confirmed the USP18 expression patterns observed computationally, anchoring the bioinformatic predictions in wet-lab measurements.

To place USP18 in its cellular context, the team turned to single-cell transcriptomics. They analyzed a DMD mouse muscle dataset (GSE213925) comprising 382,392 cells, which resolved into 24 clusters and 10 cell types after quality control and dimensionality reduction. Mast cells represented a relatively small fraction of this landscape, but AUCELL scoring showed that Kinic activity was comparatively enriched within them, and Usp18 was distributed in mast cells specifically. Using the scTenifoldKnk package, the researchers performed an in silico knockout: they constructed a gene regulatory network from the single-cell data via tensor decomposition, virtually set Usp18 expression to zero, and recomputed the network to identify genes whose regulatory connections changed significantly. The perturbed targets were enriched for molecules involved in maintaining muscular function and linked to muscular disease. Pseudotime trajectory analysis with monocle2 further revealed five differentiation trajectories along which Usp18 expression correlated with mast cell maturation, and cell–cell communication and metabolic profiling rounded out the atlas.

The therapeutic arm of the study was equally computational. Using the DrugReflector platform—a multi-omics drug repositioning database integrating transcriptomic, proteomic, and protein–protein interaction data—the team scored compounds against the GSE6011 signature using a Drug–Target Association Score that weighs gene expression similarity, chemical structure similarity, known drug–target interactions, and network topology. The deep learning pipeline flagged BRD-K10482608 as the top candidate. In parallel, network pharmacology based on the molecular perturbations caused by virtual USP18 knockout identified finafloxacin, an antibiotic, as a USP18-targeted agent. Molecular docking using AlphaFold2-predicted protein structures, AutoDock Vina simulations, and visualization in PyMOL and MOE showed favorable binding affinity between finafloxacin and USP18, with defined hydrogen bond and hydrophobic interactions at the binding interface. Together, the two compounds represent computationally derived starting points for DMD drug development.

The authors are careful to frame their conclusions within clear limitations. The Kinic signal was inferred from a transcriptomic gene signature rather than measured directly by mass spectrometry or antibody-based detection, so the findings should be read as signature-based associations rather than direct biochemical proof. The in vitro work quantified USP18 expression but did not dissect how Kinic modification of USP18 might function mechanistically. The single-cell cohort comprised only two DMD mouse samples, which may not capture the full heterogeneity of mast cell states across the disease spectrum, and the bulk datasets, despite batch correction, may retain unmodeled technical and biological variability. The drug candidates, identified entirely in silico, will require rigorous preclinical testing to confirm efficacy, dosing, and muscle tissue penetration. Still, the study offers something DMD research has lacked: a unified, data-driven framework linking a novel protein modification, a specific immune cell type, and a druggable target, with a validated diagnostic model that could support molecular stratification of patients. If future multi-center trials bear out the model’s predictive power and the therapeutic candidates’ promise, the convergence of artificial intelligence, multi-omics, and targeted validation demonstrated here could reshape how this relentless disease is diagnosed and treated.

Subject of Research: Isonicotinylation and mast cell-associated gene signatures in Duchenne muscular dystrophy identified through ensemble machine learning and in vitro validation

Article Title: Deciphering isonicotinylation and mast cell-associated pathogenic patterns in Duchenne muscular dystrophy: evidence from ensemble learning-driven multi-omics and in vitro validation

Article References: Yu, S., Xu, D., Hua, Y., Mo, H., & Liu, Y. (2026). Deciphering isonicotinylation and mast cell-associated pathogenic patterns in Duchenne muscular dystrophy: evidence from ensemble learning-driven multi-omics and in vitro validation. Frontiers in Molecular Biosciences, 13, Article 1853169. https://doi.org/10.3389/fmolb.2026.1853169

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.1853169

Keywords: Duchenne muscular dystrophy, isonicotinylation, mast cells, USP18, machine learning, multi-omics, single-cell RNA sequencing, drug repositioning, post-translational modification, biomarkers, CRISPR, inflammation