Lung cancer remains the leading cause of cancer-related death worldwide, and lung adenocarcinoma is its most common subtype. A new study in Discover Oncology reports the identification of a 40-long non-coding RNA signature associated with overall survival in lung adenocarcinoma, built by applying machine learning feature selection to gene expression data from 349 patients. The work, led by Sina Farazmandi, Seyed Alireza Khanghahi, Hadi Kamkar and colleagues including Parviz Abdolmaleki, combined two public cohorts, TCGA-LUAD and OncoSG-LCCS, and classified patients into longer survival, defined as more than three years, versus shorter survival, defined as less than two years.
Long non-coding RNAs, or lncRNAs, are RNA molecules that are transcribed from the genome but are not translated into proteins. Emerging evidence suggests they are integral to cancer biology and hold promise as prognostic biomarkers and therapeutic targets. In this study, the researchers pre-processed the combined expression dataset and then applied recursive feature elimination with support vector machines and a ridge classifier to select the lncRNA features most relevant to the survival phenotypes. This process narrowed a large field of candidate transcripts down to a novel signature of 40 lncRNAs.
Survival analysis demonstrated the prognostic value of this 40-lncRNA signature with statistically significant results in both favorable and unfavorable subsets. In the favorable subset the signature yielded a p value of 0.00018 with a hazard ratio of 0.63, and in the unfavorable subset it yielded a p value of 0.00011 with a hazard ratio of 1.6. These hazard ratios indicate that the signature’s expression pattern was associated with different risks of death across patient groups, though the study reports associations rather than a validated clinical test.
To understand what these lncRNAs might be doing biologically, the team performed enrichment analysis, which maps genes onto known functional pathways to detect overrepresented biological themes. The analysis revealed immune-related and cell cycle-related pathways associated with these lncRNAs, highlighting their biological relevance in lung adenocarcinoma. Both immune regulation and cell cycle control are central processes in tumor behavior, which makes the association plausible as a starting point for further mechanistic work rather than proof of function.
The researchers also tested whether the 40-lncRNA signature could improve survival classification compared with using the full gene expression dataset. Classification modeling showed improved accuracy with the signature compared to the full dataset, suggesting that the feature selection step removed noise and retained the most informative transcripts. Feature selection of this kind is a standard strategy in high-dimensional biological data, where the number of measured genes far exceeds the number of samples and overfitting is a constant risk.
Among the classification models tested, artificial neural networks demonstrated the highest performance, with an accuracy of 0.805 plus or minus 0.02 and an area under the curve of 0.85 plus or minus 0.03. The area under the curve, or AUC, measures a model’s ability to distinguish between the two survival classes across all classification thresholds, with 0.5 representing chance-level performance and 1.0 representing perfect separation. An AUC of 0.85 indicates moderately strong discriminative ability within this dataset, and the reported standard deviations reflect variability across the model’s evaluation runs.
Notably, the key features driving the neural network’s predictions included DPP10-AS1 and FOXD3-AS1, two lncRNAs that are recognized for their significant contributions to lung adenocarcinoma progression according to prior literature. Their prominence in the model suggests that the machine learning approach converged on transcripts with previously documented biological links to the disease, which strengthens confidence that the signature captures meaningful biology rather than statistical artifacts. However, the authors position the signature as a basis for future research on discovering lncRNA biomarkers, not as an immediately deployable prognostic tool.
The study introduces the signature with potential applications in personalized prognosis and therapeutic development for lung adenocarcinoma. Because the analysis relied on a combined 349-case dataset split by survival duration, external validation in independent cohorts and prospective clinical settings would be needed before any clinical use. The open access article, published on 30 October 2026, is shared early as a peer-reviewed accepted version that is citable and carries a permanent DOI, with a final Version of Record to follow. For now, the work adds to a growing body of evidence that non-coding portions of the genome, read through machine learning filters, may help stratify patients with lung adenocarcinoma by survival outlook.
Subject of Research: Machine learning identification of a long non-coding RNA prognostic signature for lung adenocarcinoma survival.
Article Title: Identification of a novel lncRNA prognostic signature for lung adenocarcinoma via ML-based feature selection and artificial neural networks
Article References: Farazmandi, S., Khanghahi, S. A., Kamkar, H., Shojaeilangari, S., Allahverdi, A., & Abdolmaleki, P. (2026). Identification of a novel lncRNA prognostic signature for lung adenocarcinoma via ML-based feature selection and artificial neural networks. Discover Oncology. https://doi.org/10.1007/s12672-026-06226-z
Image Credits: AI Generated
DOI: 10.1007/s12672-026-06226-z
Keywords: lung adenocarcinoma, lncRNA, prognostic signature, machine learning, artificial neural networks, TCGA-LUAD, OncoSG-LCCS, recursive feature elimination, survival analysis, enrichment analysis, biomarkers, DPP10-AS1
News Source: Nathaniel Bowman. (October 30, 2026). Machine learning pinpoints a 40-lncRNA signature linked to lung adenocarcinoma survival. Scienmag.
