Lung cancer remains the leading cause of cancer-related death worldwide. Non-small cell lung cancer accounts for roughly 85% of all cases, and of those, about 40% are the adenocarcinoma subtype (LUAD). Patients with advanced LUAD generally face a poor prognosis, making early and accurate risk classification critical for guiding treatment decisions.
Traditional prognostic tools fall short for LUAD. Clinicopathological staging, standard laboratory tests, and CT imaging often fail to capture the full biological picture, especially for early diagnosis. Researchers are therefore searching for non-invasive molecular markers that can reliably signal a patient's likely outcome before disease progression.
ALOX5 (arachidonic acid 5-lipoxygenase) is a gene that has emerged as a candidate biomarker across several cancer types. Prior studies have connected ALOX5 activity to bladder cancer progression, melanoma prognosis, and low-grade glioma outcomes. This paper investigates whether ALOX5 expression levels carry similar prognostic weight in lung adenocarcinoma.
Measuring ALOX5 in clinical practice has historically been difficult. Methods like immunohistochemistry require specific antibodies, fresh tissue, and skilled technicians, all of which add cost and variability. The researchers proposed using routine hematoxylin and eosin (H&E) stained slides combined with machine learning - a far more accessible approach - to predict ALOX5 expression indirectly.
Data for this study came from The Cancer Genome Atlas (TCGA). After applying strict inclusion criteria - excluding non-primary LUAD, missing follow-up data, survival under 30 days, and poor-quality images - 327 patients were retained. For each patient, the dataset included H&E-stained pathological images, RNA sequencing gene expression data, and clinical information.
Pathomics is the science of extracting quantitative features from pathology images using software. The researchers used the OTSU algorithm to identify tissue regions and the PyRadiomics package to extract a total of 1,488 numerical features from image tiles. Features ranged from first-order statistics (like average pixel intensity) to complex texture patterns captured by gray-level matrices.
Feature selection narrowed 1,488 candidates down to just seven key features. Two algorithms worked in tandem: minimum-redundancy-maximum-relevance (mRMR) selected the top 30 most informative and non-redundant features, and recursive feature elimination (RFE) then refined this to the best seven. These included measures of gray-level non-uniformity, dependence entropy, gray-level variance, inverse variance, intensity mean, low gray-level run emphasis, and kurtosis.
A gradient boosting machine (GBM) algorithm was trained on 70% of the cohort to predict whether a patient's ALOX5 expression was high or low. The remaining 30% served as an independent validation set. The model output a continuous score - the pathomics score (PS) - reflecting the likelihood of high ALOX5 expression from image features alone.
ALOX5 expression was notably lower in tumor tissue compared to normal lung tissue (p less than 0.001). This finding aligned with the hypothesis that ALOX5 plays a protective role in the tumor microenvironment, and that its loss may contribute to disease progression.
Patients with high ALOX5 expression lived significantly longer. The median overall survival was 53.3 months for the high-expression group versus 41.0 months for the low-expression group (p = 0.012). This survival difference persisted after adjusting for age, sex, stage, smoking history, radiotherapy, and other clinical variables, confirming ALOX5 as an independent prognostic factor (HR = 0.575; 95% CI, 0.384-0.861; p = 0.007).
Subgroup analysis reinforced the finding in younger patients. Among patients aged 65 or younger, high ALOX5 expression was especially strongly linked to better outcomes (HR = 0.501; p = 0.011). Importantly, the effect was consistent across patients who did and did not receive radiotherapy or chemotherapy, suggesting ALOX5 reflects underlying biology rather than treatment response.
The pathomics score (PS) closely tracked ALOX5 expression groups. Patients classified as high-PS had a median survival of 51.03 months compared to 39.9 months for low-PS patients, with a statistically significant difference (p = 0.003). The PS remained an independent predictor after multivariate adjustment (HR = 0.612; p = 0.02), validating the imaging model's clinical relevance.
The GBM pathomics model achieved solid predictive accuracy in both training and independent validation. On the training set of 230 patients, the area under the ROC curve (AUC) was 0.786 - a meaningful level of discrimination for a model predicting a molecular biomarker from image texture alone. On the held-out validation set of 97 patients, the AUC remained at 0.741.
Calibration curves and Hosmer-Lemeshow goodness-of-fit tests confirmed the model's reliability. A well-calibrated model's predicted probabilities match actual outcomes - the training set passed this test (p = 0.061) and the validation set also passed comfortably (p = 0.441). This means the model's confidence levels are trustworthy, not just its rank-ordering of patients.
Decision curve analysis (DCA) demonstrated that the model provides real clinical benefit. DCA evaluates whether using a model to guide decisions results in better outcomes than treating all patients or treating none. The pathomics model showed positive net benefit across a wide range of clinical decision thresholds in both training and validation sets.
Among the seven selected features, gray-level non-uniformity and gradient kurtosis carried the greatest predictive weight. These texture features capture how uniformly or chaotically pixel intensities are distributed across tissue regions - patterns that, while visually subtle to the human eye, encode information about underlying cellular organization and, indirectly, gene expression.
The tumor immune microenvironment differed significantly between high and low PS groups. Using CIBERSORTx - a computational tool that estimates immune cell populations from gene expression data - the researchers found that the high-PS group had greater infiltration of M0, M1, and M2 macrophages compared to the low-PS group (p less than 0.05 for each type).
Gene set enrichment analysis (GSEA) identified two major pathways linked to PS variation. The Hallmark gene sets showed that genes differentially expressed between high and low PS groups were significantly enriched in apoptosis (cell death) and inflammatory response pathways. This suggests that patients with high ALOX5 expression have tumors characterized by active immune and cell-death programs.
Apoptosis-related genes were strongly and positively correlated with PS values. Genes including PIK3R5, FAS, and PIK3CG showed correlation coefficients significantly above zero (p less than 0.001 for all). These genes are known mediators of programmed cell death, suggesting that high-ALOX5 tumors may be more susceptible to natural and therapy-induced apoptosis.
Weighted gene co-expression network analysis (WGCNA) identified 63 hub genes linked to inflammatory response and prognosis. These genes clustered into a network module strongly correlated with patient outcomes. GO enrichment analysis showed they were concentrated in cytokine-mediated signaling pathways, pointing to ALOX5's role in regulating the immune communication network within tumors.
To confirm that ALOX5 actually influences cancer cell behavior, the team conducted laboratory experiments. They first mapped ALOX5 expression across nine cell lines including normal bronchial epithelial cells and eight lung cancer lines. Two high-expressing lines (A549 and PC9) and two low-expressing lines (H1975 and H23) were selected for manipulation.
Knocking down ALOX5 using shRNA made cancer cells more aggressive. In A549 and PC9 cells where ALOX5 was reduced, CCK-8 proliferation assays showed faster cell growth. Colony formation assays confirmed that ALOX5-depleted cells formed more colonies - a hallmark of increased tumorigenic potential. Scratch wound healing assays revealed significantly faster migration in ALOX5-knockdown cells.
Overexpressing ALOX5 had the opposite effect - it slowed tumor cells down. In H1975 and H23 cells engineered to produce more ALOX5, proliferation decreased and colony formation was reduced. The wound-healing assay showed markedly slower migration in cells with elevated ALOX5, consistent with a tumor-suppressive role for this gene in LUAD.
These results were verified at both protein and mRNA levels. Western blot and quantitative PCR confirmed that the knockdown and overexpression constructs successfully changed ALOX5 levels. The consistency across two independent cell lines for each condition, validated by multiple methods, substantially strengthens confidence in the functional conclusions.
This study is the first to use histopathological imaging to predict ALOX5 expression levels in lung cancer. Prior work required molecular assays - expensive, technically demanding, and tissue-destructive. The pathomics approach replaces these with analysis of routine H&E slides, potentially enabling ALOX5-based prognostication without additional laboratory procedures.
The authors propose that ALOX5 promotes better outcomes partly through its role in ferroptosis and inflammation. Ferroptosis is a form of iron-dependent cell death. ALOX5 is a key enzyme in this process. The macrophage infiltration observed in high-ALOX5 tumors may reflect ALOX5-driven iron metabolism changes that shape the immune microenvironment in a favorable direction.
GSEA pointed to enrichment in the MAPK signaling pathway, apoptosis, and inflammatory response. Prior research links lipid peroxidation products - which ALOX5 generates - to activation of the MAPK cascade and downstream caspase signaling, both of which drive apoptosis. This mechanistic chain provides a plausible explanation for why high ALOX5 correlates with improved survival.
The relationship between ALOX5 and prognosis appears cancer-type specific. High ALOX5 is protective in bladder cancer and LUAD, but is associated with worse outcomes in colon cancer. This variability likely reflects differences in tumor cell biology and the molecular context in which ALOX5 operates, underscoring the importance of studying it within each cancer type independently.
The pathomics-based model offers a practical and low-cost tool for clinical decision support. Because it works from standard H&E slides rather than specialized molecular assays, it could be integrated into existing pathology workflows without major additional cost. The ability to predict ALOX5 status from images alone could help stratify patients into risk groups and guide personalized treatment planning.
Several limitations temper the conclusions. The data came entirely from the TCGA public database, meaning no external validation on an independent clinical cohort has yet been performed. The study also focused exclusively on LUAD and does not generalize to other lung cancer subtypes. Subimage selection for feature extraction may introduce sampling bias.
The biological mechanisms linking ALOX5 expression to outcomes remain incompletely defined. The authors propose that ALOX5 may modulate the tumor microenvironment or reflect a less invasive differentiation state, but these are working hypotheses rather than established facts. Dedicated mechanistic studies are needed to confirm and extend these observations.
Future work will focus on prospective multicenter validation. The research team is establishing a multicenter cohort with standardized specimen processing to enable external validation. Plans include expanding the clinical annotation to include detailed treatment histories, comorbidity data, and additional molecular markers to better control for confounding factors and test potential interactions.