S807
Clinical - Lung
ESTRO 2026
Digital Poster 3053
Prediction and related genes of distant metastasis of lung adenocarcinoma based on deep learning ChunMei Xia, Pei Wang, Juan Deng Department of Oncology, Chongqing Traditional Chinese Medicine Hospital, Chongqing, China Purpose/Objective: Lung adenocarcinoma (LUAD) is one of the leading causes of cancer-related deaths, with distant metastasis significantly affecting prognosis. Early detection of metastatic disease is crucial, but existing clinical tools such as TNM staging and imaging are limited in predictive power. This study aims to develop a deep learning model using transcriptomic data to predict distant metastasis in LUAD, offering a molecular-level tool for early detection and personalized treatment strategies. Material/Methods: Gene expression data and clinical information from 293 LUAD patients, including 21 with distant metastasis, were downloaded from TCGA. Differentially expressed genes (DEGs) between metastatic and non-metastatic samples were identified using the edgeR package, yielding 63 DEGs. Subsequently, the Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to select the top 6 genes with the highest predictive value. A Gene-based Deep Neural Network (GDNN) model was trained using these 6 genes, employing Focal Loss and weighted sampling to handle class imbalance. To ensure robustness, experiments were repeated 5 times with different random training/validation splits (80%/20%), and performance metrics including area under the ROC curve (AUC), accuracy, and recall were averaged across runs. The MDNN was compared with traditional models, such as logistic regression and random forest. Results: The 6 selected genes were significantly associated with distant metastasis in LUAD, and functional enrichment analysis revealed their involvement in key pathways related to tumor progression, metastasis, and cell adhesion. The deep learning model built on these 6 genes demonstrated excellent predictive performance, achieving an AUC of 0.936 and an accuracy of 0.939 . Compared to traditional models like logistic regression and random forest, the deep learning model significantly outperformed these methods, offering superior predictive capabilities. Fig 1. displays the LASSO feature selection process, and Fig 2. illustrates the model performance comparison.
Conclusion: While the PACIFIC study remains the only trial to date that has changed the standard of care for stage III NSCLC, a rapidly expanding portfolio of ongoing trials is exploring novel RT–IO strategies that may further redefine the management of this challenging clinical setting. References: 1. doi: 10.1200/JCO.21.013082. doi: 10.1016/S1470- 2045(21)00630-63. https://doi.org/10.1093/annonc/mdz438.0214. Dziadziuszko R, et al. SKYSCRAPER-03: Phase 3, open- label, randomised study of atezolizumab (atezo) + tiragolumab (tira) vs durvalumab (durva) in locally advanced, unresectable, stage III non-small cell lung cancer (NSCLC) after platinum-based concurrent chemoradiation (CCRT). ESMO Congress 2025 - LBA695. . doi: 10.1016/j.cllc.2021.07.0056. D. Bradley , S. Sugawara , K.H.H. Lee Durvalumab in combination with chemoradiotherapy for patients with unresectable stage III NSCLC: Final results from PACIFIC-2. ESMO 24 congress.7. ECOG-ACRIN EA5181: phase 3 trial. In 2025 World Conference on Lung Cancer, Abstract PL-3.04. Keywords: Radiotherapy, Immunotherapy, NSCLC
Made with FlippingBook - Share PDF online