Genomic classification of intrapulmonary metastasis and multiple primary lung cancer

•MeTel is a novel Bayesian algorithm that accurately classifies multifocal lung cancers. •It outperforms previous methods and is robust across different sequencing platforms. •Its clinical utility was validated by correctly reclassifying diagnostically challenging cases. Dear Editor, The most critical decision point in the diagnosis of multiple lung cancers is distinguishing between intrapulmonary metastasis (IPM) and multiple primary lung cancer (MPLC), as this differentiation significantly affects staging, treatment, and prognosis.1, 2 With growing emphasis on molecular profiling, the latest guidelines highlight its role in distinguishing IPM from MPLC, aiming to overcome the limitations of classic histological criteria.3, 4 However, both conventional histology-based5, 6 and recently proposed genomic approaches7, 8 have shown limited reliability in clinical decision-making. In this study, we developed MeTel (Metastasis Teller), a probabilistic Bayesian model that robustly classifies multifocal lung cancers as either IPM or MPLC. MeTel is designed for non-small cell lung cancer (NSCLC), which represents the majority of multifocal lung cancer cases and for which somatic mutation data are widely available through routine clinical sequencing. Unlike previous approaches, MeTel does not rely on empirically defined thresholds or data-driven approaches, thereby enhancing generalisability across sequencing platforms and patient populations. The overall workflow of MeTel evaluates the profile of somatic mutations from different tumour sites (Figures 1 and S1). Tumours are first evaluated for key driver mutations (Table S1); if discordant, MPLC is immediately assigned. For concordant cases, MeTel calculates the probabilities of IPM and MPLC classification, providing two outputs: a classification score and a confidence level (‘Confident’ or ‘Likely’). For ‘Likely’ cases, defined by low variant counts, MeTel allows for post-adjustment using Comprehensive Histological Assessment (CHA).6 Detailed descriptions of the MeTel algorithm are provided in the Supplementary Methods. We compared MeTel with four previously described methods using independent test datasets (Table S2). Of the 635 samples in the test set, MeTel accurately classified 623 cases, with an error rate of 1.89%, outperforming all four other methods (Figure 2A, Table S3, and Note S1). Moreover, MeTel achieved the highest Cohen κ score of.95 (Figure 2B and Table S3), indicating a high inter-rater reliability. The F1 scores for IPM (.97) and MPLC (.99) were also the best for MeTel (Figure 2C and Table S4). Notably, MeTel maintained high accuracy across the entire range of gene panel sizes (93.75%–100% for 4–808 genes and whole-exome sequencing (WES)), whereas the four previous algorithms showed inconsistent accuracy across different panel sizes and platforms (Figure 2D). This performance stability indicates the strength of MeTel's probabilistic model, which avoids fixed cutoffs and training dependencies. Its robustness across sequencing platforms and panel sizes supports MeTel's applicability in diverse clinical settings. The confidence level assigned by MeTel provided practical guidance in clinical decision-making. We found that almost all miscalls (10/12) by MeTel were included in the 336 ‘Likely’ calls (accuracy = 97.02%), whereas the 299 ‘Confident’ calls were 99.33% accurate (2/12 miscalls) (Figure 2E). As low confidence arises from limited genomic data, MeTel enables the integration of histological information for ‘Likely’ calls. When this optional process was applied, the overall accuracy of MeTel was increased to 99.37% (Figure 2F), with F1 scores of.99 and 1.00 for IPM and MPLC, respectively (Figure 2G). This demonstrates that MeTel provides highly accurate classifications based solely on mutation profiles and can achieve near-perfect accuracy when supplemented with histologic information. Moreover, this result shows the algorithm's flexibility, enabling integrative diagnosis that combines mutation-based classification with pathological assessments, particularly in ambiguous cases. We further applied MeTel to 12 in-house patients with multifocal NSCLC from Yonsei University Severance Hospital in Seoul, Korea (Tables S5 and S6), whose initial classifications were equivocal under standard procedures by pathologists, as described in the Supplementary Methods (section: In-House Cohort for Clinical Application). We reviewed the 12 cases with additional follow-up information, including disease progression, metastasis, and sequencing data (Table S7). In eight cases, MeTel's classifications aligned with the initial histopathologic diagnoses and were further supported by prognostic and follow-up data (Figures S2 and S3). In contrast, four cases (Patients 2, 6, 7 and 11) exhibited discordance between histologic diagnosis and MeTel, and underwent in-depth reassessment. Patient 2 had three adenocarcinomas with acinar-predominant morphology. Although the second and third tumours exhibited micropapillary and lepidic features suggestive of MPLC, MeTel predicted IPM. This prediction was subsequently confirmed by CT imaging and WES-based clonality analysis, leading to a revised diagnosis (Figure 3A and Note S2). Patient 6, originally diagnosed with IPM due to similar morphology of two squamous cell carcinomas, was reclassified as MPLC based on MeTel's prediction. This was supported by distinct clonal architectures (Figure 3B and Note S3). Patient 7 had two lepidic-predominant adenocarcinomas separated by 8 years and was initially diagnosed as MPLC. However, MeTel predicted IPM, which was substantiated by copy number variation (CNV) similarity and clinical course (Figure 3C, Table S8 and Note S4). Patient 11 presented with two histologically dissimilar adenocarcinomas diagnosed 1 year apart. Although originally diagnosed as MPLC, MeTel predicted IPM, supported by shared rare variants and similar CNV patterns (Figure 3D, Table S9 and Note S5). These cases show the advantage of using genomic information in distinguishing IPM from MPLC, particularly when the clinical diagnosis is uncertain. They not only validate MeTel's utility in guiding reassessment but also reveal the limitations of conventional criteria – such as time interval or histologic similarity – in accurately classifying multifocal lung cancers. Supporting discussion is provided in Note S6. In conclusion, MeTel provides a statistically principled, training-free, and platform-independent approach for distinguishing IPM from MPLC using somatic mutation data. Its Bayesian framework computes posterior probabilities based on the number and frequency of shared variants, without relying on arbitrary cutoffs or large labelled datasets, thereby ensuring consistent performance across sequencing platforms and patient populations. In cases with low variant counts – and therefore lower-confidence predictions – MeTel also allows the optional integration of histopathologic information to support clinical decision-making. The robustness of the current model provides a strong foundation for future enhancements. Our preliminary explorations into CNV integration and ethnicity-specific modelling confirm the feasibility of these extensions and underscore the critical need for larger genomic databases to further advance the field (Notes S7 and S8). Ultimately, applying MeTel to larger prospective cohorts will be essential to further establish its clinical utility. Jeongsoo Won: conceptualisation, methodology, software, validation, formal analysis, investigation, data curation, writing – original draft, visualisation. Yeon Seung Chung: conceptualisation, validation, formal analysis, investigation, writing – original draft, visualisation. Se-Young Jo: methodology, software, formal analysis. Jiho Park: methodology, software, formal analysis. Hyo Sup Shim: conceptualisation, resources, writing – review and editing, supervision, project administration. Sangwoo Kim: conceptualisation, methodology, formal analysis, resources, writing – original draft, writing – review and editing, supervision, project administration. The authors have nothing to report. S.K. is a cofounder of AIMA Inc., which seeks to develop techniques for early cancer diagnosis based on circulating tumour DNA. This study was supported by the Bio&Medical Technology Development Program of the National Research Foundation (NRF) funded by the Korean government (MSIT) (RS-2023-00261820) to S.K. and the National Research Foundation of Korea (NRF) grant funded by the Korean Government (MSIT) (RS-2025-00518452) to H.S.S. This study involving human participants was reviewed and approved by the Ethics and Human Subjects Committee of Severance Hospital, Korea (Approval No. 4-2017-1149). The raw sequencing in-house data generated from this study are available under NCBI Sequence Read Archive (SRA) (SRA code: PRJNA1070282), which can be accessed at https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1070282. The source codes for MeTel algorithm are available at https://github.com/JeongsooWon/MeTel. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.