Molecular models show promise for predicting lung cancer recurrence

Molecular biomarkers such as gene expression profiles and DNA methylation markers can predict recurrence and survival in early-stage non-small cell lung cancer (NSCLC) with moderate to good accuracy, but most of the models built around them have never been tested outside the patient population where they were developed, according to a systematic review and meta-analysis published in Thoracic Cancer.

Researchers led by Aaron Ling, of Latrobe Regional Hospital, and Evangeline Samuel, of Monash University, both in Victoria, Australia, searched MEDLINE, Embase and the Cochrane Library for studies published between January 2000 and December 2023, screening 2,447 records before identifying five eligible studies of peer-reviewed, validated models predicting recurrence-free, cancer-specific or overall survival in patients with stage I or II NSCLC. All five used Cox proportional hazards regression and drew on molecular predictors including mRNA expression profiles, long noncoding RNAs and DNA methylation biomarkers, spanning four retrospective cohorts and one prospective cohort.

Up to 55% of patients with NSCLC relapse after surgery or other curative-intent treatment, and treatment failure is the leading cause of death among these patients, the authors explained. Yet oncology societies disagree on post-treatment surveillance.

The National Comprehensive Cancer Network recommends chest CT scans every six months for two to three years, the American Association for Thoracic Surgery advises twice-yearly scans for four years, and the European Society for Medical Oncology recommends at least annual scans, or twice yearly if salvage therapy may be needed.

“There is no agreed-upon standard for the optimal frequency or method of post-treatment surveillance imaging, resulting in differing guidelines from various oncology organizations,” the authors wrote. “These recommendations are often based on lower-quality evidence and expert opinions.”

Across the four studies with internal validation, AUC values ranged from 0.66 to 0.89, yielding a pooled AUC of 0.77, though heterogeneity was high. The two externally validated models performed somewhat less well, with a pooled AUC of 0.72 and low heterogeneity. The best-performing model, developed by researchers in China using a two-gene signature (ACADM and RPS8), achieved an AUC of 0.89 for overall survival in 243 patients with stage I lung adenocarcinoma. A 50-gene signature developed at the University of Michigan and validated in 483 patients with squamous cell lung carcinoma posted a lower but still respectable external C-index of 0.683.

Using the PROBAST tool, reviewers rated two of the five studies at high risk of bias, primarily because of how they handled missing data. The authors also noted that none of the mRNA-based gene signatures overlapped between studies, and that all five originated in just two countries, the United States and China.

"The small number of included studies, as well as the limited external validation, does impact the generalizability of the performance models," the authors wrote, adding that future models should account for the growing use of neoadjuvant immunotherapy, chemotherapy and targeted therapy for actionable mutations.

The review has limitations of its own. Several included studies did not report confidence intervals, the restriction to English-language publications may have excluded relevant research, and the predominance of studies from the U.S. and China limits how well the findings translate to other regions or health systems.

"Molecular-based prediction models demonstrate moderate-to-good discriminative performance for recurrence and survival in early-stage NSCLC," the authors concluded. "However, broader external validation and improved methodological rigor are required before routine clinical implementation."