Machine Learning Pinpoints 12 Candidate Biomarkers for Mesothelioma

Researchers have combined gene co-expression network analysis with machine learning to nominate a panel of twelve candidate biomarkers for malignant pleural mesothelioma, an aggressive cancer of the lining of the lung that is primarily linked to asbestos exposure. The study, published in Cancer Reports, is entirely computational: it mined a single public microarray dataset, built gene co-expression modules, extracted highly connected hub genes, and then tested how well those genes could distinguish tumor from normal tissue within the same cohort. The authors are explicit that the findings are computational nominations that require independent biological and clinical validation before any diagnostic or therapeutic use.

Malignant pleural mesothelioma remains a difficult target. Because of the disease’s long latency period and continued asbestos exposure in some regions, its incidence persists despite legislation restricting the mineral. Patients typically face a dismal prognosis, with median survival durations ranging from 9 to 17 months after diagnosis, largely because early detection is difficult and effective treatment options are scarce. Existing protein-based biomarker candidates, such as soluble mesothelin-related peptides, fibulin-3, and osteopontin, have shown limited sensitivity and specificity for early identification or for distinguishing mesothelioma from other pleural diseases, which motivated the search for gene-level markers.

The team drew its data from the Gene Expression Omnibus repository under accession number GSE51024, a dataset measured on the Affymetrix Human Genome U133 Plus 2.0 Array containing 96 samples: 55 mesothelioma tumor samples and 41 paired normal lung samples. After normalizing the 54,675 probe sets with robust multiarray averaging, the researchers identified differentially expressed genes using two statistical approaches in parallel. The Limma package, with a false discovery rate threshold below 0.01, yielded 13,305 differentially expressed genes, while the Wilcoxon rank test, with an adjusted p-value of 0.01 or lower, yielded 14,491. Taking the intersection of the two lists left 12,649 genes for downstream network analysis.

Those genes were then fed into weighted gene co-expression network analysis, a method that groups genes with correlated expression patterns into modules rather than treating each gene in isolation. Using a soft-thresholding power of 9, which produced a scale-free fit index of 0.90, the analysis identified several co-expression modules. A module preservation analysis was then used to assess how stable and reproducible each module was, with modules scoring above 10 on the Z-summary statistic considered strongly preserved. Based on the combined preservation metrics, the black, pink, purple, magenta, and red modules were identified as moderately stable and well-preserved, and these five modules were carried forward.

Functional enrichment analysis of the shared genes pointed to biological processes relevant to cancer behavior. Gene Ontology analysis highlighted cilium assembly, chromosomal segregation, mitotic nuclear division, and microtubule-based motility, processes essential for maintaining cytoskeletal integrity and controlling cell division that malignant cells commonly disrupt. The Kyoto Encyclopedia of Genes and Genomes pathway analysis showed marked enrichment of the cell cycle, cellular senescence, and FoxO signaling pathways, consistent with dysregulated apoptosis and cell growth control, along with enrichment in endocytosis, focal adhesion, and tight junction pathways, which relate to cellular adhesion and intracellular transport. The Wnt signaling pathway and axon guidance also appeared, which the authors note may relate to immune evasion and tumor microenvironment remodeling.

Within the five preserved modules, the researchers defined hub genes as those with an absolute module membership correlation above 0.8, indicating a strong association with the module’s overall expression profile. This yielded twelve hub genes: ZNF638, RAB18, SOS1, TTC17, and RANBP2 from the black module; FNDC3B and QSER1 from the magenta module; CLDN15 from the purple module; and PSTPIP1, SEPTIN6, ADAP2, and PVRIG from the red module. No hub genes met the criterion in the pink module. The authors then examined each gene’s prognostic relevance using the GEPIA2 web platform and the TCGA mesothelioma cohort, stratifying patients into high- and low-expression groups at the median cutoff and comparing overall survival with log-rank tests.

Eleven of the twelve hub genes were associated with overall survival in the TCGA mesothelioma cohort. Higher expression of RAB18, TTC17, RANBP2, FNDC3B, and QSER1 was associated with poorer overall survival, while lower expression of ZNF638, CLDN15, and PVRIG was associated with poorer survival. The authors caution that these are associations from a public cohort analysis, not established prognostic tests. Prior literature reviewed in the paper offers context for individual genes: ZNF638 has been described as a potential epigenetic repressor whose reduction can trigger innate immune responses and viral mimicry in other work, RAB18 was identified as a target of miR-30b/c in nonsmall cell lung cancer cells, and PVRIG is an inhibitory immune checkpoint receptor expressed on T cells and natural killer cells whose blockade improved T cell responses in preclinical mouse models.

To test whether the twelve-gene panel could discriminate tumor from normal samples, the team trained five supervised classifiers on the same cohort: support vector machine with a radial basis kernel, linear discriminant analysis, logistic regression, extreme gradient boosting, and K-nearest neighbor. Performance was evaluated with repeated stratified fivefold cross-validation using 25 repeats, with hyperparameters optimized within the resampling scheme. Extreme gradient boosting achieved the best result, with a mean area under the receiver operating characteristic curve of 97.53 percent, accuracy of 90.04 percent, sensitivity of 90.69 percent, and specificity of 89.12 percent. The support vector machine followed with an AUC of 96.90 percent and accuracy of 92.99 percent, K-nearest neighbor reached an AUC of 95.42 percent, logistic regression 93.17 percent, and linear discriminant analysis 92.39 percent, the weakest of the five.

The authors emphasize that these classification figures describe within-cohort performance, meaning the models were trained and evaluated on resamples of the same dataset rather than validated on independent patient groups. The study’s limitations section is candid on this point: the analysis rested on a single public Affymetrix dataset with a modest sample size, complete batch-level metadata were unavailable so residual batch effects cannot be excluded, and the comparator samples were normal lung tissues rather than normal pleural mesothelial tissues. Clinical variables such as age, sex, histological subtype, smoking history, asbestos exposure, tumor stage, and treatment history were not fully available.

What the study delivers, then, is a ranked shortlist rather than a validated test. The twelve hub genes and the enriched pathways they implicate, spanning cell cycle regulation, cytoskeletal dynamics, adhesion, and immune signaling, provide a systems-level view of the molecular landscape of mesothelioma and a starting point for laboratory follow-up. The authors state that future work should validate the findings in independent, larger patient cohorts spanning diverse clinical stages and treatment histories, and should pursue in vitro and in vivo experiments, including qRT-PCR, immunohistochemistry, and functional assays, to confirm the genes’ functional roles and assess their potential as therapeutic targets. Until such validation is complete, the panel remains a computationally nominated set of candidates for one of oncology’s most stubborn diseases.

Subject of Research: Integrative bioinformatics and machine learning analysis of gene expression data to identify candidate biomarkers and pathways in malignant pleural mesothelioma

Article Title: Integrative Bioinformatics and Machine Learning Approaches for Biomarker and Pathway Discovery in Malignant Pleural Mesothelioma

Article References: Rohman, M. L., & Biswas, S. (2026). Integrative Bioinformatics and Machine Learning Approaches for Biomarker and Pathway Discovery in Malignant Pleural Mesothelioma. Cancer Reports, 9(10), Article e70731. https://doi.org/10.1002/cnr2.70731

Image Credits: AI Generated

DOI: 10.1002/cnr2.70731

Keywords: malignant pleural mesothelioma, biomarkers, WGCNA, machine learning, gene expression, hub genes, GSE51024, survival analysis, XGBoost, pathway enrichment, asbestos exposure, computational biology