Predicting Disease Progression in Inoperable Localized NSCLC Patients Using ctDNA Machine Learning Model

Abstract

Introduction: There is an urgent clinical need to accurately predict the risk for disease progression in post-treatment NSCLC patients, yet current ctDNA mutation profiling approaches are limited by low sensitivity. We represent a non-invasive liquid biopsy assay utilizing cfDNA neomer profiling for predicting disease progression in 44 inoperable localized NSCLC patients.

Methods: A total of 97 plasma samples were collected at various time points during or post-treatments (TP1: 39, TP2: 33, TP3: 25). cfDNA neomer profiling, generated based on target sequencing data, was used to fit survival support vector machine models for each time point. Leave-one-out cross-validation (LOOCV) was performed to evaluate the models' predictive performances.

Results: Our cfDNA neomer profiling assay showed excellent performance in detecting patients with a high risk for disease progression. At TP1, the high-risk patients detected by our model showed an increased risk of 3.62 times (hazard ratio [HR] = 3.62, p = 0.0026) for disease progression, compared to 3.91 times (HR = 3.91, p = 0.0022) and 4.00 times (HR = 4.00, p = 0.019) for TP2 and TP3. These neomer profiling determined HRs were higher than the ctDNA mutation-based results (HR = 2.08, p = 0.074; HR = 1.49, p = 0.61) at TP1 and TP3. At TP1, the predictive model reached 40% sensitivity at 92.9% specificity, outperforming the mutation-based method (40% sensitivity at 78.6% specificity), while the combination results reached a higher sensitivity (60%). Finally, the longitudinal analysis showed that the combination of neomer and ctDNA mutation-based results could predict disease progression with an excellent sensitivity of 88.9% at 80% specificity.

Conclusion: In conclusion, we developed a cfDNA neomer profiling assay for predicting disease progression in inoperable NSCLC patients. This assay showed increased predicting power during and post-treatment compared to the ctDNA mutation-based method, thus illustrating a great clinical potential to guide treatment decisions in inoperable NSCLC patients.

Trial registration: ClinicalTrials.gov: NCT04014465.

Keywords: MRD detection; NSCLC; machine learning; neomer; non‐invasive.