Machine Learning Pinpoints Five Mitochondrial Genes Tied to Bone Infection Pathology

A computational and experimental investigation has mapped out a specialized five-gene mitochondrial expression profile associated with Staphylococcus aureus osteomyelitis, demonstrating how metabolic and organellar stress signatures track with severe bacterial bone infections. Published in Frontiers in Cellular and Infection Microbiology by a team from the Affiliated Hospital of Zunyi Medical University, the research integrates multi-cohort human peripheral blood transcriptomic datasets, supervised machine-learning algorithms, immune deconvolution, and experimental validations in both murine models and human skeletal tissues. Osteomyelitis caused by Staphylococcus aureus presents an enduring clinical challenge because the pathogen establishes stubborn local infections characterized by bone destruction, dynamic immunometabolic warfare, and intracellular evasion tactics. Standard clinical management often struggles with diagnostic delays, invasive biopsy requirements, and false negatives following empirical antibiotic exposure. By analyzing host transcriptional profiles through the lens of mitochondrial dysfunction, the investigators sought to identify robust candidate molecular markers capable of providing tissue-level adjunctive insight into local lesion biology and host metabolic remodeling.

To establish candidate disease signatures, the researchers acquired three publicly available human microarray datasets deposited in the Gene Expression Omnibus repository, specifically examining cohorts with clear clinical annotations for Staphylococcus aureus infection and healthy control states. The discovery training cohort was constructed by pooling peripheral blood mononuclear cell expression data from datasets GSE6269 and the GPL96 platform subset of GSE16129, both generated on the Affymetrix Human Genome U133A Array. Together, these two datasets contributed 48 osteomyelitis samples and 16 healthy control profiles from pediatric patients presenting with acute or invasive staphylococcal disease. Probe annotations were standardized to official gene symbols, duplicate probe signals were collapsed, and technical batch variations between the two merged discovery datasets were adjusted using parametric empirical Bayes methods in ComBat. A completely separate whole-blood dataset, GSE30119, assayed on the Illumina HumanHT-12 platform and containing 36 osteomyelitis and 40 healthy control specimens, was set aside as an external evaluation cohort to evaluate candidate model transferability across independent populations and distinct profiling technologies.

Differential expression analysis of the batch-corrected discovery cohort using the limma package identified 434 differentially expressed genes matching the screening thresholds of an adjusted significance level below 0.05 and an absolute log2 fold-change exceeding 1.0. Among these, 231 genes were upregulated and 203 were downregulated in osteomyelitis patients compared to healthy individuals. To focus specifically on metabolic and organelle-driven pathways, the authors intersected the 434 differentially expressed genes with the MitoCarta3.0 inventory of 1,136 curated human mitochondrial genes. This overlap isolated 20 mitochondria-related differentially expressed genes, most of which displayed significantly higher expression levels in infected patients. Functional enrichment analyses using Gene Ontology and Kyoto Encyclopedia of Genes and Genomes classifications revealed that these 20 candidate genes were strongly enriched in metabolic and inflammatory signaling networks, including ATP-binding cassette transporters, peroxisome proliferator-activated receptor cascades, peroxisome functions, nuclear factor kappa B activation, iron and hemoglobin handling, cytochrome c release, and mitochondrial-mediated apoptosis.

To condense the 20 mitochondrial candidates into a compact and clinically interpretable diagnostic panel, the investigators deployed three complementary feature-selection algorithms. Least Absolute Shrinkage and Selection Operator regression utilizing ten-fold cross-validation isolated nine predictive markers; Support Vector Machine Recursive Feature Elimination identified ten genes minimizing classification error; and a Random Forest ensemble ranked candidates by importance scores to nominate its top ten genes. Taking the exact intersection of these three machine-learning methods isolated five core mitochondria-related genes: QTRT1, ABCB6, PPIF, ACSL1, and GLRX5. Each core gene reflects essential aspects of cellular physiology and organellar homeostasis: QTRT1 facilitates queuosine modification of transfer RNAs; ABCB6 functions as an ATP-binding cassette transporter implicated in porphyrin and heme transport; PPIF encodes cyclophilin D, an indispensable regulator of the mitochondrial permeability transition pore; ACSL1 directs long-chain fatty acid activation; and GLRX5 mediates iron-sulfur cluster biogenesis and mitochondrial iron homeostasis.

Using the five identified core genes, the authors constructed and benchmarked eight distinct machine-learning models within the 64-sample training cohort, encompassing Extreme Gradient Boosting, Generalized Linear Models, K-Nearest Neighbors, Artificial Neural Networks, Decision Trees, C5.0 algorithms, LASSO regression, and Support Vector Machines. Because five algorithms achieved an apparent area under the receiver operating characteristic curve of 1.000 and the neural network scored 0.982 in the training set, training discrimination alone could not differentiate algorithm utility. Applying the candidate models directly to the external whole-blood cohort GSE30119 exposed marked performance differentiation across platforms. The LASSO model retained the strongest discriminatory ability in this external evaluation set, securing an area under the curve of 0.879. Subsequent SHapley Additive exPlanations, or SHAP, analyses demonstrated how each of the five core genes contributed to LASSO predictions across individual patient profiles, illustrating complex feature dependencies and confirming that the five genes collectively drove model discrimination.

To explore the broader immunometabolic environment linked to the five-gene signature, the authors conducted computational immune deconvolution and single-sample gene set enrichment analyses on the discovery transcriptomic data. Applying CIBERSORT deconvolution revealed marked shifts in estimated circulating immune cell fractions between osteomyelitis patients and healthy individuals, characterized by significant differences in relative proportions of monocytes, neutrophils, resting natural killer cells, naive CD4-positive T cells, and naive B cells. Correlation testing showed that core gene expression tracked with these inferred leukocyte shifts; notably, ACSL1 expression correlated positively with estimated neutrophil and monocyte fractions while correlating inversely with resting natural killer cells. In parallel, single-sample enrichment scoring demonstrated significant upregulation of transcriptional programs governing regulated cell death in osteomyelitis, with apoptosis, ferroptosis, and neutrophil extracellular trap formation, or NETosis, exhibiting prominent enrichment scores, though the authors emphasized that these computational estimates represent transcriptional associations rather than direct cellular abundance measurements.

The study then sought experimental confirmation of the computational findings by establishing a controlled surgical model of Staphylococcus aureus osteomyelitis in 20 seven-week-old female C57BL/6J mice. Under sterile surgical conditions, an L-shaped stainless steel needle pre-coated with Staphylococcus aureus strain ATCC 25923 at a concentration of 100 million colony-forming units per milliliter was implanted through a drilled cortical hole directly into the right tibial medullary cavity, while control animals received identical sterile pins. At three weeks post-implantation, six randomly selected mice per group underwent high-resolution micro-computed tomography imaging and histological staining. Three-dimensional micro-CT reconstructions revealed extensive cortical discontinuity, destructive osteolytic lesions, and pathological cavity formation in infected tibiae, corroborated by significant reductions in bone mineral density and bone volume fraction compared to control limbs. Histological evaluation via hematoxylin and eosin staining confirmed extensive inflammatory leukocyte infiltration and architectural destruction throughout the infected cortical and trabecular compartments.

Quantitative reverse-transcription polymerase chain reaction and Western blot assays performed on harvested mouse tibial tissues confirmed that transcriptional alterations identified in circulating blood translated directly to the infected bone microenvironment. In the murine tibial lesion tissue, mRNA transcripts for ABCB6, ACSL1, PPIF, and GLRX5 showed significant upregulation compared with uninfected controls, whereas QTRT1 exhibited marked downregulation. Western blot analyses conducted across three biological replicates per group demonstrated identical changes at the protein level, confirming heightened expression of ABCB6, ACSL1, PPIF, and GLRX5 alongside reduced QTRT1 abundance in the infected bone matrix. These consistent quantitative measurements demonstrated that local staphylococcal bone infection triggers significant perturbations in mitochondrial protein machinery, reflecting altered fatty acid metabolism, dysregulated iron-sulfur balance, and heightened susceptibility to mitochondrial permeability transition during active osteolytic inflammation.

To substantiate the translational relevance of the animal and bioinformatic findings, the researchers examined human bone specimens harvested from six adult orthopedic surgery patients, consisting of three individuals diagnosed with culture-confirmed post-traumatic or periprosthetic Staphylococcus aureus osteomyelitis and three non-infected fracture controls. Protein extraction and densitometric Western blot quantification showed concordant directional changes matching the animal data: human osteomyelitic bone tissue displayed significantly elevated levels of ABCB6, ACSL1, GLRX5, and PPIF proteins, together with a concurrent reduction in QTRT1 expression compared to uninfected fracture specimens. Although limited by the small human sample size, the concordant expression profile across peripheral blood transcriptomics, controlled murine tibial osteomyelitis, and human surgical specimens supports the disease-associated relevance of this five-gene mitochondrial panel at the direct site of skeletal infection.

The authors were careful to outline key constraints governing their conclusions, underscoring that these five genes represent tissue-associated markers of disease rather than standalone, non-invasive early diagnostic screening tests. Because the discovery data relied on retrospective pediatric blood datasets and testing was validated in bulk bone tissue without matched blood-bone pairs from individual patients, the exact relationship between circulating leukocyte signals and focal bone alterations requires prospective clarification. Furthermore, because the primary comparisons contrasted Staphylococcus aureus infection with healthy or uninfected controls rather than sterile osteolysis or infections caused by unrelated pathogens, pathogen specificity remains unproven. The researchers also highlighted that computational immune and cell death scores cannot substitute for direct functional assays, and because gene knockout or overexpression was not conducted, causal roles in pathogenesis cannot be assumed. Future studies utilizing single-cell transcriptomics, prospective cohorts, and direct mechanistic perturbations will be essential to establish whether targeting these mitochondrial pathways can ultimately refine surgical debridement margins or therapeutic intervention in osteomyelitis.

Subject of Research: Mitochondria-related transcriptomic biomarkers and machine-learning modeling in Staphylococcus aureus osteomyelitis

Article Title: Machine learning identifies mitochondria-related biomarkers for Staphylococcus aureus osteomyelitis with validation in mouse and human bone tissues

Article References: Deng, X., Yang, F., Li, J., Peng, J., Peng, C., Yang, L., Qian, H., & Ao, J. (2026). Machine learning identifies mitochondria-related biomarkers for Staphylococcus aureus osteomyelitis with validation in mouse and human bone tissues. Frontiers in Cellular and Infection Microbiology, 16, Article 1923149. https://doi.org/10.3389/fcimb.2026.1923149

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1923149

Keywords: Osteomyelitis, Staphylococcus aureus, Machine Learning, Mitochondria, Biomarkers, Transcriptomics, Immunometabolism, Bone Infection, Machine, learning, identifies, mitochondria-related