Machine Learning Models Based on Molecular Features for Distinguishing Multiple Primary Lung Cancers from Intrapulmonary Metastases

Transl Lung Cancer Res 2025 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Telling Two Lung Tumors Apart Changes Everything for Treatment

The Clinical Problem When a lung cancer patient develops a second lung lesion, clinicians face a critical diagnostic question: is this a separate primary tumor (multiple primary lung cancer, MPLC) or a metastasis of the original cancer (intrapulmonary metastasis, IPM)? The answer completely changes the treatment strategy and prognosis.

Why It Matters MPLC patients can potentially benefit from curative-intent surgery on both lesions, while IPM patients are staged as metastatic disease requiring systemic therapy. Misclassification can lead to either undertreating a curable patient or overtreating one whose disease has already spread.

Study Approach The researchers performed whole-exome sequencing (WES) on 157 tumor lesions from 72 patients, extracted four molecular features from the genomic data, and trained three machine learning models to distinguish MPLC from IPM.

Published in Translational Lung Cancer Research This multicenter study was published in 2025 in a high-impact thoracic oncology journal, underscoring its clinical relevance for the growing population of patients with synchronous or metachronous lung lesions.

TL;DR: This study develops machine learning models using whole-exome sequencing to distinguish multiple primary lung cancers (MPLC) from intrapulmonary metastases (IPM), a clinically critical distinction that guides treatment strategy.
Pages 2-3
Four Genomic Features That Capture Tumor Relatedness

Genetic Divergence (delta-SI) The delta Similarity Index measures how genetically dissimilar two tumors are based on their mutational profiles. High divergence suggests independent origins (MPLC), while low divergence suggests a shared lineage (IPM).

Shared Mutations The number of identical somatic mutations found in both lesions is directly counted. A high overlap in driver mutations strongly implies that both tumors arose from the same clone, supporting an IPM diagnosis.

Pearson Correlation The Pearson correlation coefficient between the two tumors' complete mutation profiles quantifies overall genomic similarity. IPM pairs show high correlation because the metastatic lesion inherits most of its mutations from the primary tumor.

Early Mutation Number This feature captures the number of mutations that likely occurred early in tumorigenesis, before any divergence between lesions. Tumors with a large shared pool of early mutations were more likely to be clonally related (IPM).

TL;DR: Four WES-derived molecular features - genetic divergence, shared mutation count, genomic correlation, and early mutation burden - were engineered to quantify the clonal relationship between lung lesion pairs.
Pages 3-4
Study Design and Machine Learning Approach

Patient Cohort Seventy-two patients with two or more lung lesions underwent WES on all tumor samples, yielding 157 individual tumor lesions. Expert pathologists and genomic analysis established ground-truth MPLC vs. IPM labels for model training.

Three ML Classifiers Decision tree (DT), random forest (RF), and gradient boosted decision tree (GBDT) classifiers were trained on the four molecular features using cross-validation, selected for their ability to handle small sample sizes without overfitting.

Validation Strategy The dataset was split into training and validation sets, with performance assessed using area under the ROC curve (AUC) and other standard metrics. The relatively small cohort made cross-validation essential for reliable performance estimates.

Clinical Integration Beyond just classification accuracy, the researchers analyzed whether the ML-predicted MPLC designation correlated with actual disease-free survival (DFS) outcomes, testing whether the model captured clinically meaningful biological differences.

TL;DR: WES was performed on 157 tumor lesions from 72 patients; three ML classifiers were trained on four molecular features and validated against ground-truth clinical labels.
Pages 5-6
High Accuracy Across All Three Classifiers

Random Forest and GBDT Performance Both the random forest and GBDT classifiers achieved near-perfect AUC scores of 1.00 in validation, indicating that the four molecular features provide near-perfect discrimination between MPLC and IPM in this cohort.

Decision Tree Performance The simpler decision tree also performed impressively with an AUC of 0.94, demonstrating that even an interpretable single-tree model can capture the essence of tumor clonal relationships from these four features.

Survival Validation Patients classified as MPLC had significantly prolonged disease-free survival compared to IPM patients, with a hazard ratio of 0.21 (95% CI 0.04-1.0, p=0.04), confirming that the molecular classification captures real biological and prognostic differences.

Feature Importance Across all three models, genetic divergence (delta-SI) and shared mutations were the two most important discriminating features, consistent with their direct measurement of clonal relatedness between tumor pairs.

TL;DR: Random forest and GBDT achieved AUC of 1.00 in validation; MPLC classification correlated with significantly better disease-free survival (HR=0.21), confirming the model's prognostic value.
Pages 6-7
Changing Treatment Decisions with Molecular Evidence

Upgrading Staging Decisions Current clinical practice relies on imaging, histology, and sometimes clonal analysis, but lacks a standardized molecular framework. These ML models could provide objective, reproducible molecular evidence to guide staging decisions.

Surgical Candidacy Patients whose second lesion is reclassified from presumed IPM to MPLC using this approach may become candidates for resection of both lesions with curative intent, potentially converting what appeared to be metastatic disease into a curable scenario.

Genomic Testing Integration As WES becomes more cost-effective and clinically accessible, incorporating these four molecular features into routine genomic profiling panels could make ML-aided MPLC/IPM distinction a standard part of the diagnostic workup.

Avoiding Over- and Under-Treatment The survival data confirm that correct classification has real consequences - MPLC patients unnecessarily treated as metastatic receive systemic therapy and forgo curative surgery, while IPM patients incorrectly treated with surgery may relapse quickly.

TL;DR: This molecular classification tool could upgrade the accuracy of staging decisions, potentially enabling curative surgical treatment for patients who would otherwise be incorrectly classified as having metastatic disease.
Pages 7-8
Expanding and Validating the Approach

Small Cohort Size Seventy-two patients is a limited training set for machine learning, particularly given the complexity of genomic data. Larger prospective multicenter validation cohorts are needed before clinical implementation.

WES Cost and Accessibility Whole-exome sequencing remains expensive for routine clinical use. Future work should investigate whether targeted sequencing panels covering fewer genes can reproduce the four key features at a fraction of the cost.

Metachronous Lesions Most patients in this cohort had synchronous lesions discovered simultaneously. The model's performance on metachronous lesions developing years apart, where clonal evolution has had more time to occur, warrants separate validation.

Integration with Liquid Biopsy Future studies should explore whether circulating tumor DNA from blood draws can provide sufficient genomic data to calculate these four features non-invasively, eliminating the need for separate biopsies of each lesion.

TL;DR: Future validation requires larger multicenter cohorts and investigation of whether targeted panels or liquid biopsy can replace expensive whole-exome sequencing for practical clinical implementation.
Citation: Open Access, 2025. Available at: PMC12082235.