Deconvolution of evolutionary architecture unmasks a high-risk, subclonal-rich subtype in treatment-naive small cell lung cancer

Abstract

Background

Intratumoral heterogeneity (ITH) drives therapeutic resistance in small cell lung cancer (SCLC). However, conventional single-sample analysis has limited horizontal, cross-patient comparisons, leaving the overarching evolutionary architecture in treatment-naive tumors poorly understood. This study aims to deconvolve these architectures to identify clinically relevant evolutionary subtypes.

Methods

We analyzed whole-exome sequencing data from 41 treatment-naive SCLC patients. To overcome the cross-patient comparability bottleneck, we developed a novel probabilistic framework using a refined Gaussian Mixture Model (GMM). This standardized subclonal structures into four hierarchical strata, enabling the identification of evolutionary subtypes via unsupervised clustering. To address the scarcity of SCLC public data, prognostic concordance was robustly explored in The Cancer Genome Atlas (TCGA) lung squamous cell carcinoma (LUSC) based on shared smoking etiology, with lung adenocarcinoma (LUAD) serving as a negative control.

Results

The cohort robustly segregated into “Clonal-dominant” (Group 1, n=28) and “Subclonal-rich” (Group 2, n=13) subtypes. Group 1 evolution was primarily driven by tobacco signatures (SBS4). Conversely, Group 2 exhibited late-stage acquisition of a DNA mismatch repair deficiency (MMRd) signature (SBS15), fueling trace subclonal diversification. Clinically, Group 2 demonstrated a significantly lower objective response rate (ORR) to platinum-based regimens (25.0% vs. 81.3%, P=0.02). Furthermore, the Subclonal-rich architecture independently predicted inferior overall survival (OS) [adjusted hazard ratio (adj. HR) =2.93, P=0.02], driven predominantly by limited-stage disease. Cross-cancer analysis validated this histology-dependent, high-heterogeneity adverse pattern in early-stage LUSC but not in LUAD.

Conclusions

This hypothesis-generating study demonstrates that a “Subclonal-rich” architecture, driven by acquired MMRd, identifies high-risk, chemo-resistant SCLC. Our GMM approach suggests that pre-existing heterogeneity may serve as a potential, histology-dependent prognostic marker that warrants prospective validation for tailoring future therapeutic regimens.

Highlight box.

Key findings

• Using a novel Gaussian Mixture Model (GMM), we identified two evolutionary architectures in treatment-naive small cell lung cancer (SCLC): a “Clonal-dominant” subtype driven by early tobacco exposure, and a high-risk “Subclonal-rich” subtype driven by late-stage mismatch repair deficiency. The subclonal-rich phenotype is associated with an inferior response to platinum chemotherapy and shorter overall survival, a pattern validated in smoking-associated lung squamous cell carcinoma.

What is known and what is new?

• SCLC is traditionally viewed as a homogeneous disease dominated by a rapid clonal sweep. Furthermore, conventional subclonal reconstruction methods previously failed to allow horizontal comparisons across different patients.

• We developed a probabilistic framework that aligns subclonal architectures across patients. This approach unmasked that clinically relevant, targetable subclonal heterogeneity pre-exists in a subset of treatment-naive tumors, suggesting it may serve as a pre-existing substrate for therapeutic resistance prior to any therapy.

What is the implication, and what should change now?

• As a hypothesis-generating study, this evolutionary architecture provides a robust, histology-dependent biomarker to identify chemo-resistant, high-risk limited-stage SCLC.

• Identifying this pre-existing heterogeneity suggests that “Subclonal-rich” patients may benefit from intensified or alternative therapeutic strategies, such as poly(ADP-ribose) polymerase inhibitors or tailored immunotherapies, rather than standard chemoradiotherapy alone.

Introduction

Small cell lung cancer (SCLC) represents a persistent clinical paradox. Although the disease is initially exquisitely sensitive to chemotherapy, its trajectory is defined by rapid recurrence and high mortality. The foundational treatment paradigm based on platinum chemotherapy has barely shifted for decades. Even with the recent addition of immunotherapy, which has sparked a revolution in non-SCLC (NSCLC), it offers only modest and often transient survival benefits. This limitation underscores a profound lack of predictive biomarkers (1). Such a therapeutic impasse stems from a complex biology governed by extreme genomic instability and phenotypic plasticity.

The genomic landscape of SCLC was comprehensively mapped over the past decade. Landmark studies by Peifer et al. [2012] (2), Rudin et al. [2012] (3), and George et al. [2015] (4) established the defining hallmarks of the disease. These include the near-universal biallelic inactivation of TP53 and RB1 as very early truncal events (5), as well as a high tumor mutation burden (TMB) largely driven by tobacco exposure. This evidence led to a prevailing view of SCLC as a relatively homogeneous entity compared to NSCLC, seemingly dominated by a single “clonal sweep” established early in its life history. Consequently, efforts to stratify prognosis based on simple mutational presence or aggregate TMB have largely failed to guide clinical decision-making. This failure is particularly acute in limited-stage (LS) disease, in which curative intent is possible, but outcomes remain highly variable.

The field is now undergoing a paradigm shift from static genomic snapshots to the dynamic mapping of tumor evolution. This trend is illustrated by large-scale longitudinal studies in NSCLC, such as the TRACERx cohort (6,7). These efforts have revealed that subclonal selection and genomic heterogeneity are fundamental drivers of progression and metastasis. In SCLC, the landmark George et al. [2024] study elegantly demonstrated that chemotherapy acts as a profound evolutionary bottleneck that sculpts the genomic landscape and selects for rare, pre-existing subclones (8). However, a critical methodological gap persists: traditionally, subclonal reconstruction is performed entirely on a per-patient basis. Because each tumor yields a stochastic, sample-specific set of subclonal clusters, it has been mathematically intractable to perform a horizontal, cross-patient comparison of evolutionary architectures. Standard bulk sequencing often discards low-frequency variants as noise, thereby overlooking “trace” subclones that may harbor intrinsic engines of instability [e.g., DNA mismatch repair deficiency (MMRd)] before any drug is administered.

In this study, we hypothesize that the baseline structure of intratumoral heterogeneity (ITH) pre-encodes the capacity for lethal evolution. To overcome the cross-patient comparability bottleneck, we developed a novel probabilistic framework using a refined Gaussian Mixture Model (GMM). By anchoring population-level cancer cell fraction (CCF) distributions, we successfully standardized sample-specific subclonal noise into comparable evolutionary topologies. Furthermore, given the extreme scarcity of public SCLC datasets, we hypothesized that if this aggressive evolutionary architecture is fundamentally shaped by smoking-induced genomic instability, it should be biologically conserved in lung squamous cell carcinoma (LUSC)—a lineage sharing SCLC’s central airway origin and heavy tobacco etiology—but absent in lung adenocarcinoma (LUAD). Through this integrated approach, we unmask hidden evolutionary architectures that challenge the notion of SCLC as a purely homogeneous disease. We present this article in accordance with the REMARK reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1338/rc).

Methods

Patient cohort and study design

This retrospective study re-analyzed a cohort of 41 treatment-naive SCLC patients diagnosed at Shanghai Pulmonary Hospital from November 2018 to June 2019, as originally described by Wang et al. [2021] (9). Strict inclusion criteria ensured that all patients underwent biopsy prior to any anti-cancer treatment to characterize the baseline genomic landscape without therapy-induced selection pressures. The detailed patient screening and inclusion process is depicted in the REMARK flow diagram (Figure 1A), ensuring the transparency of our cohort assembly and minimizing selection bias. Patients were clinically classified into central-type or peripheral-type based on bronchoscopy and imaging results and were staged according to the Veterans Administration Lung Study Group criteria. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shanghai Pulmonary Hospital (Approval No. K18-203Y). Written informed consent was obtained from all patients participating in the study.

Figure 1.

Probabilistic refinement of clonal architecture identifies robust evolutionary subgroups. (A) Study workflow, integrating the REMARK patient selection flow diagram and the downstream bioinformatics analysis pipeline. (B) Probabilistic refinement of clonal architecture. The left panel shows BIC analysis (k=2−5) identifying k=4 as the optimal model complexity for the weighted GMM. The right panel displays faceted density plots deconvolving raw clusters into four hierarchical populations. The raw clusters (bottom row) are colored by their original PyClone cluster IDs to visualize the input data structure; however, these raw clusters are colored arbitrarily and lack biological comparability across samples due to the stochastic nature of upstream clustering. The weighted GMM organizes these scattered, disjointed inputs into coherent biological strata: Clonal (red, defined by biological prior CCF ≥0.9), major subclone (orange), minor subclone (green), and trace subclone (blue). The dashed line indicates the 0.9 threshold used for defining the clonal boundary. (C) Patient Stratification. The heatmap displays the relative fraction of the four GMM-defined clonal components across the 41-patient cohort. Each column represents a patient, and each row corresponds to a clonal component. Unsupervised hierarchical clustering of the patient profiles (dendrogram, top) robustly segregates the cohort into two primary super-groups, designated as Group 1 (cyan, Clonal-dominant) and Group 2 (red, Subclonal-rich) in the annotation track. The color scale indicates the fraction of each component within a tumor’s mutation profile, ranging from 0 (blue) to 1 (red). BIC, Bayesian information criterion; CCF, cancer cell fraction; CNV, copy number variation; ES, extensive-stage; GMM, Gaussian mixture model; ITH, intratumoral heterogeneity; LS, limited-stage; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; ORR, objective response rate; QC, quality control; SCLC, small cell lung cancer; TCGA, The Cancer Genome Atlas; WES, whole-exome sequencing; WGD, whole-genome duplication.

To rigorously exclude LUAD-to-SCLC transformation as a potential confounder of mutational signature analyses, we applied a three-tier verification framework. Clinically, all 41 patients were treatment-naive with no prior NSCLC diagnosis or exposure to EGFR tyrosine kinase inhibitors (TKIs); LUAD-to-SCLC transformation is an acquired resistance mechanism exclusive to TKI-treated patients and has not been reported in TKI-naïve individuals (10,11). Histomorphologically, all diagnoses were centrally confirmed as classical SCLC (high nuclear-to-cytoplasmic ratio, salt-and-pepper chromatin, nuclear molding, and high mitotic index), with no mixed or transitional adenocarcinomatous features. Genomically, whole-exome sequencing (WES) analysis confirmed the canonical de novo SCLC landscape: near-universal TP53 inactivation [any alteration in 98% of patients (40/41), comprising somatic mutations in 85% and 17p chromosomal deletion in 73%; 61% confirmed biallelic] and RB1 inactivation [any alteration in 83% (34/41), via somatic mutation in 44% and/or 13q14 deletion in 76%; 37% confirmed biallelic]. Critically, no EGFR activating mutations (exon 19 deletions or L858R) and no KRAS G12/G13 mutations were identified in any of the 41 patients. One patient (1/41, 2.4%) harbored a STK11 p.G163D missense variant (rs137853078); however, this patient demonstrated biallelic TP53 inactivation [somatic nonsense mutation p.E298 combined with copy number neutral loss of heterozygosity (CNLOH) at 17p] and RB1 inactivation (CNLOH at 13q14), with no co-occurring EGFR activating mutations, firmly excluding a LUAD transformation origin (10-12). STK11 loss-of-function variants are observed in a subset of de novo SCLC and do not independently indicate a LUAD lineage (13). One additional patient demonstrated a low-level EGFR copy number gain, representing a non-specific chromosomal instability event in the context of high overall genomic instability [median Weighted Genome Instability Index (wGII) =0.45], without any EGFR activating mutation, and carrying no implication for LUAD-driven transformation.

Genomic profiling and somatic variant calling

Whole-exome sequencing was performed on matched tumor-normal samples, yielding a mean sequencing depth of 491× for tumor samples and 256× for normal controls. Raw reads for the SCLC cohort were aligned to the GRCh38 reference genome. We implemented a bioinformatics pipeline designed to reconstruct subclonal architectures and identify low-abundance variants. Somatic single nucleotide variant (SNVs) and indels were identified using Strelka2 and Manta. To minimize false positives, we applied a filtering strategy where variants were retained only if they possessed robust tumor support of at least 5 reads and near-zero background noise in matched normal controls. Detailed parameters for alignment, quality control, and the Panel of Normals are provided in Appendix 1.

Hybrid clonal stratification

To establish a cross-patient comparable framework for clonal architecture, we utilized a hybrid stratification strategy. The process began on a per-patient basis using PyClone-VI to estimate the CCF of somatic mutations by integrating allele-specific copy number and tumor purity data from FACETS. Subsequently, inferred mutation CCFs from the entire cohort were pooled to enable cohort-level analysis. We performed a methodological benchmarking to identify the optimal statistical model, comparing a weighted GMM against alternative approaches, including Beta Mixture Models (BMM) and logit-transformed GMMs. The weighted GMM was selected based on the Bayesian Information Criterion and the biological coherence of the resulting clusters, as detailed in Appendix 1. This model was applied to automatically determine the optimal number of clusters, which were provisionally labeled according to their descending mean CCF.

These statistical labels were subsequently refined using a deterministic biological rule to enforce consistency. Specifically, any mutation with a CCF of 0.9 or higher was definitively re-classified as “Clonal” to ensure all truncal events were captured, regardless of the provisional GMM label. Conversely, mutations provisionally assigned to the “Clonal” cluster by the GMM but falling below this 0.9 threshold were systematically reassigned to the “Major Subclone” category. This approach partitioned all mutations into four evolutionary strata: clonal, major subclone, minor subclone, and trace subclone.

Internal validation of the GMM stratification framework

We emphasize that our population-level GMM framework is not an independent CCF deconvolution tool, but rather a population-level re-stratification layer built upon PyClone-VI CCF estimates, designed to resolve the cross-patient comparability bottleneck inherent to per-patient clustering. To demonstrate that this re-stratification is both biologically coherent and clinically informative, we conducted two lines of internal validation (Figure S1, Table S1).

First, we confirmed the biological coherence of the four-tier strata. The GMM-defined strata exhibited a strict, monotonically decreasing CCF gradient (median CCF: clonal =1.000, major subclone =0.785, minor subclone =0.334, trace subclone =0.043; Spearman ρ =−1.000, P<0.001; Figure S1A), consistent with a biologically ordered temporal hierarchy from truncal clonal mutations to late subclonal variants. This ordering was independently cross-validated by an orthogonal phylogenetic method: GMM-derived “Clonal” assignments achieved 100% concordance (40/40 patients) with trunk clusters identified by the pigeonhole principle.

Second, we assessed the clinical added value of this four-tier architecture relative to conventional TMB, a standard aggregate genomic heterogeneity metric. In LS-SCLC (n=18), the GMM evolutionary grouping (Group 1 vs. Group 2) achieved a Harrell’s C-index of 0.697 (log-rank P=0.009; Figure S1B), compared to a C-index of 0.543 (log-rank P=0.83) for TMB median-split stratification. This demonstrates that the GMM evolutionary architecture captures clinically relevant prognostic heterogeneity that aggregate mutation burden cannot detect.

Evolutionary subgrouping and genomic characterization

Patients were stratified into evolutionary subgroups via unsupervised hierarchical clustering based on the relative proportions of the four clonal strata. Cluster stability was validated using consensus clustering with 1,000 bootstrap iterations. The genomic landscape of these subgroups was characterized across multiple dimensions. We calculated standard genomic metrics following the protocols detailed in our previous study (9), including TMB, copy number alteration (CNA) burden, and the wGII. Additionally, we quantified the Mutant-Allele Tumor Heterogeneity (MATH) score, which measures the width of the variant allele frequency distribution. Driver dependency was evaluated by comparing the mutation frequencies of COSMIC Tier 1 driver genes (Table S2) and identifying high-confidence copy number drivers, including significant amplification and deletion peaks (Table S3). Evolutionary selection pressure was estimated using dNdScv to calculate non-synonymous to synonymous substitutions (dN/dS) ratios for missense and truncating mutations, with corrections for background mutation rates and sequence context. Methodological details regarding GISTIC2.0 parameters and dNdScv constraints are provided in Appendix 1.

Longitudinal modeling of mutational processes

Mutational signatures were extracted using SigProfilerAssignment (COSMIC v3.4; the full catalog is listed in Tables S4,S5), and their activities were mapped to the four evolutionary stages (clonal through trace) to reconstruct temporal dynamics. We applied Generalized Linear Mixed Models (GLMMs) to quantify evolutionary trends, accounting for the non-independence of repeated measures within the same patient. Specifically, relative contribution trends were modeled using Beta regression, while absolute burden trends were modeled using Negative Binomial regression. This modeling framework distinguished between early truncal drivers (such as tobacco exposure) and late-stage acquired mutational processes (such as MMRd). The mathematical formulation of these models is detailed in Appendix 1. To provide direct orthogonal verification of mutational timing, clonal evolutionary tree topologies were reconstructed for each patient using the pigeonhole principle. Somatic mutations were categorized into the early “clonal trunk” or late “subclonal branches” based on these hierarchies. Chronological timing was assessed by directly comparing the CCF distributions and trunk-versus-branch mutation proportions of MMRd-associated mutations (signatures SBS6, SBS15, SBS26) against background mutations.

Clinical statistical analysis and cross-cancer exploration

Clinical outcomes were assessed based on objective response rate (ORR) and overall survival (OS). Survival curves were generated using the Kaplan-Meier method with log-rank tests, and independent prognostic factors were identified using multivariable Cox proportional hazards models. Survival analyses were performed on the entire cohort and a sensitivity subcohort restricted to patients receiving systemic therapy.

To investigate lineage specificity, the analytical framework was applied to the The Cancer Genome Atlas (TCGA)-LUSC (n=486) and TCGA-LUAD (n=515) cohorts from the TCGA PanCancer Atlas. For the cross-cancer exploration, raw sequencing data were processed using a compatible hg19-based pipeline, whereas the primary SCLC cohort was aligned to GRCh38. Since our clonal stratification relies on mutational fractions and copy number profiles rather than absolute genomic coordinates, the difference in reference genomes does not impact the comparability of the evolutionary hierarchies. The validation analysis was restricted to early-stage (stage I–II) patients to parallel the LS-SCLC analysis. All statistical analyses were conducted in R (v4.5.1), with P values less than 0.05 considered statistically significant.

Results

High-fidelity genomic profiling and re-analysis framework

We performed a comprehensive re-analysis of a previously published cohort of 41 SCLC patients (Figure 1A). The high-depth whole-exome sequencing data (mean tumor coverage >400×; see “Methods” for details) provided the necessary resolution to accurately characterize the subclonal landscape. To this end, we employed an enhanced bioinformatics pipeline designed to maximize sensitivity for low-frequency variants while rigorously controlling for false positives. This pipeline incorporated robust corrections for tumor purity and CNAs to ensure the accurate estimation of CCF, thereby establishing a high-confidence mutation set for all subsequent analyses. Comprehensive cohort comparison statistics, genomic alterations, copy number driver peaks, and genomic instability indices are detailed in the Figures and Tables.

We implemented a stringent “clean normal” filtering strategy to distinguish genuine somatic mutations from sequencing artifacts. By mandating near-zero background noise in matched normal samples, we ensured that 88% of the identified variants with low variant allele frequencies (VAF <5%) had zero supporting reads in their corresponding normal controls, confirming their somatic origin. Following this rigorous filtering, we identified a total of 36,762 high-confidence somatic mutations (including SNVs and indels, with 34,106 SNVs), with a median TMB of 5.23 mutations/Mb. This high-fidelity mutation set successfully recapitulated the classic genomic landscape of SCLC, identifying widespread genomic inactivation of TP53 (85%) and RB1 (80%) through a combination of somatic mutations and copy number losses, and served as a robust foundation for the subsequent probabilistic deconvolution of clonal architectures.

Probabilistic refinement of clonal architecture via optimized mixture modeling

To establish a standardized framework for quantifying ITH, we performed a rigorous comparative analysis of three mixture modeling approaches: a BMM, a GMM on logit-transformed CCF, and a GMM on raw CCF. While per-patient CCF estimation tools such as PyClone-VI and SciClone generate stochastic, patient-specific cluster labels that are mathematically non-comparable across individuals, our population-level GMM framework addresses this critical limitation. By pooling mutation CCFs from the entire cohort (n=41), we define a shared, population-level CCF distribution landscape, into which each patient’s mutations are mapped to obtain universally comparable evolutionary strata. This design resolves the cross-patient comparability bottleneck that is inherent to all per-patient clustering tools. The detailed technical improvements (purity-corrected CCF estimation, BIC-based model order selection, and multi-start initialization) and the full model comparison rationale are provided in Appendix 1 and Figure S2.

As detailed in Figure S2, the BMM, despite being theoretically suitable for proportional data, exhibited mixture collapse and failed to resolve intermediate subclonal populations (effective k=2). The logit-transformed GMM, while improving statistical fit for specific clusters, introduced significant boundary distortions by compressing the clonal peak and artificially stretching the Trace Subclone tail. Quantitative comparison confirmed that the logit-transformed model severely underestimated the complexity of the Trace Subclone (identifying only 5 clusters versus 48 in the raw model), whereas the BMM failed to distinguish the clonal population entirely (Table S6).

In contrast, the GMM on raw CCF provided the superior balance between model complexity and biological resolution. Although goodness-of-fit tests indicated deviations from strict normality, this model robustly identified four biologically distinct evolutionary strata without the artifacts or underfitting observed in alternative methods.

Consequently, we adopted this raw-space GMM framework. To ensure biological rigor, we integrated the probabilistic modeling with a deterministic biological prior. While BIC analysis identified k=4 as the optimal complexity (Figure 1B, left), the final classification incorporated a two-step refinement process.

First, the “Clonal’” stratum was strictly defined as the primary cluster with a CCF ≥0.9, representing truncal mutations present in nearly all cancer cells.

Second, for the remaining three GMM-identified clusters in each tumor, we implemented a rank-based hierarchical assignment to define the subclonal strata. Specifically, after excluding the “Clonal” cluster, the remaining clusters were ranked based on their mean CCF values. The cluster with the highest mean CCF was designated as the “major subclone”, the cluster with the second-highest mean CCF was designated as the “minor subclone”, and the cluster with the lowest mean CCF was designated as the “trace subclone”. We chose this rank-based approach because it is adaptive to the unique subclonal distribution of each individual tumor, avoiding the imposition of arbitrary, universal CCF thresholds that may not be appropriate across a heterogeneous cohort. This method ensures a consistent and biologically meaningful hierarchical stratification, transforming scattered raw data into a coherent, standardized feature space (Figure 1B, right).

Patient stratification based on clonal composition reveals two distinct subgroups

Having established this standardized four-component model, we investigated whether distinct patterns in clonal composition could stratify the patient cohort. Unsupervised hierarchical clustering based on the relative proportions of the four clonal strata segregated the cohort into two robust super-groups (Figure 1C). The somatic alteration landscape and driver mutation profiles of these subgroups are presented in Figure 2A,2B. Group 1 (n=28), designated “Clonal-dominant”, is characterized by a clear predominance of the Clonal component. In stark contrast, Group 2 (n=13), designated “Subclonal-rich”, is characterized by a significantly lower clonal fraction and a heterogeneous expansion of subclonal populations (Figure 2C). This stratification was robust to resampling and provides a novel framework for exploring SCLC heterogeneity in this exploratory cohort.

Figure 2.

Distinct clonal architectures define the genomic landscapes of SCLC subgroups. (A) Somatic alteration landscape. The OncoPrint displays the distribution of non-synonymous SNVs in the top 25 most frequently altered genes (excluding TTN). Columns represent individual patients (n=41), segregated and ordered by their assigned genomic subgroup. Annotation tracks provide information on TMB, genomic subgroup, WGD status, clinical stage, smoking history, and gender. (B) Quantitative heatmap of mutation counts stratified by clonal status. This heatmap displays the counts of non-synonymous SNVs for a filtered set of frequently altered driver genes (present in ≥5 samples and classified as Tier 1 in the COSMIC Cancer Gene Census). Rows are stratified into four blocks representing the clonal strata, with genes clustered within each block. Columns represent patients, segregated and clustered within their respective subgroups. The color intensity reflects the mutation count, highlighting the clonally-dominant profile of Group 1 versus the subclonally-active profile of Group 2. (C) Comparison of clonal composition. The 100% stacked bar chart illustrates the average proportional contribution of each clonal stratum to the total burden of non-synonymous, Tier 1 CGC SNVs for Group 1 and Group 2. The significant difference in composition was confirmed by Fisher’s Exact Test (P<0.001). (D) Polarized evolutionary burden. Boxplots show the distribution of absolute non-synonymous, Tier 1 CGC SNV counts within each of the four clonal strata. P values from Wilcoxon Rank-Sum tests are displayed on each panel, revealing a polarized pattern where Group 1 has a higher burden in high-abundance strata, while Group 2 specifically exhibits a significantly higher burden in the lowest-abundance Trace Subclone stratum (P=0.03). Group 1: Clonal-dominant. Group 2: Subclonal-rich. F, female; M, male; SCLC, small cell lung cancer; SNV, single nucleotide variant; TMB, tumor mutation burden; WGD, whole-genome duplication.

Divergent evolutionary trajectories and late-stage acquisition of instability

We next sought to characterize the biological forces driving these distinct architectures. Baseline clinical features, including clinical stage, were comparable between groups (Table 1; Figure S3). To rigorously rule out treatment disparity as a confounder, we categorized the first-line systemic therapy into platinum-based regimens, non-platinum-based regimens, and no systemic therapy/unknown. The distribution of these clinical regimens was highly comparable between the two genomic groups (P=0.60; Table 1). Thus, the administration of standard first-line platinum-based therapy was statistically balanced, largely ruling out treatment disparity as a confounder for the observed prognostic differences. However, Group 2 displayed a significantly higher overall subclonal mutation burden and a specifically polarized burden in the trace subclone stratum (P=0.03; Figure 2D, Figure S3). We further verified that this subgrouping was not driven by simple differences in driver gene mutation frequencies, as no single COSMIC Tier 1 gene showed a statistically significant difference after multiple testing correction (Figure 2A,2B; Figure S4; Table S2).

Table 1. Baseline clinical and molecular characteristics of the SCLC cohort stratified by genomic subgroups.

Characteristic	Overall (N=41)	Group 1 (N=28)	Group 2 (N=13)	P value†
Age (years)	67 [64–69]	67 [64–71]	66 [64–69]	0.64
Gender				>0.99
Female	4 (9.8)	3 (10.7)	1 (7.7)
Male	37 (90.2)	25 (89.3)	12 (92.3)
Smoking history				0.40
Ever smoker	34 (82.9)	22 (78.6)	12 (92.3)
Never smoker	7 (17.1)	6 (21.4)	1 (7.7)
Clinical stage				>0.99
Limited-stage	18 (43.9)	12 (42.9)	6 (46.2)
Extensive-stage	23 (56.1)	16 (57.1)	7 (53.8)
First-line systemic therapy				0.60
Platinum-based regimens	32 (78.0)	23 (82.1)	9 (69.2)
Non-platinum-based regimens	4 (9.8)	2 (7.1)	2 (15.4)
No systemic therapy/unknown	5 (12.2)	3 (10.7)	2 (15.4)
Objective response rate (platinum-based cohort)‡	15/24 (62.5)	13/16 (81.3)	2/8 (25.0)	0.02
TMB (mut/Mb)	5.23 [3.53–6.90]	5.14 [3.80–6.90]	5.99 [3.09–6.74]	0.88
wGII score	0.45 [0.30–0.59]	0.44 [0.28–0.61]	0.51 [0.36–0.58]	0.97
MATH score	47.94 [26.37–65.38]	36.95 [22.40–53.47]	75.89 [52.76–93.91]	<0.001

Data are presented as median [Q1–Q3], n (%), or n/N (%). ^†, statistical tests: Fisher’s exact test for categorical variables; Wilcoxon rank-sum test for continuous variables. ^‡, ORR was evaluated exclusively in 24 patients who received first-line platinum-based therapy and had strictly evaluable radiographic baseline and follow-up assessments. The remaining 8 patients in the platinum-treated cohort were excluded from ORR calculation due to incomplete treatment cycles or missing imaging evaluations. Group 1: Clonal-dominant. Group 2: Subclonal-rich. MATH, Mutant-Allele Tumor Heterogeneity; ORR, objective response rate; SCLC, small cell lung cancer; TMB, tumor mutation burden; wGII, Weighted Genome Instability Index.

To delineate the evolutionary dynamics, we reconstructed mutational signatures (Tables S4,S5), revealing a “crossover” in evolutionary drivers (Figure 3). The clonal phase of Group 1 tumors was dominated by tobacco-associated signatures (SBS4), consistent with a classic smoking-induced etiology, whereas Group 2 showed a relative enrichment of endogenous clock-like processes (SBS5). Strikingly, this pattern reversed in later evolution, where we observed a specific acquisition of hyper-mutability in Group 2. The mutation burden in the Trace Subclone stratum was significantly higher in this group (P=0.002; Figure 3B). Signature deconvolution attributed this late-stage burst to the activity of signature SBS15, which is strongly associated with MMRd. The contribution of this MMRd-like signature showed a monotonic increase from clonal to trace stages in Group 2 (Figure 3C,3D), suggesting that while Group 1 is driven by early exogenous DNA damage, Group 2 appears to evolve via an acquired intrinsic mutator-like phenotype that fuels continuous subclonal diversification.

Figure 3.

Divergent evolutionary trajectories and acquired mutational instability. (A) Average mutational landscape. Stacked bar charts showing the relative contribution of mutational signatures across four evolutionary stages (clonal, major subclone, minor subclone, and trace subclone). Group 1 (left) is dominated by tobacco-associated signatures (SBS4) in the clonal phase, whereas Group 2 (right) shows a shift toward defective DNA repair signatures (e.g., SBS15) in later stages. (B) Evolution of mutational burden. Boxplots compare total mutation counts at each evolutionary stage. In the clonal phase, Group 1 exhibits significantly higher clonal TMB compared to Group 2 (P=0.003), consistent with smoking-induced carcinogenesis. In the trace phase, Group 2 displays a significantly higher burden in the trace subclone fraction compared to Group 1 (P=0.002), indicating a late-stage acquisition of hyper-mutability driven by intrinsic defects (e.g., MMRd) in a minor subpopulation. (C) Differential signatures. A balloon plot highlighting signatures with statistically significant differences (FDR <0.05) between groups at each stage. Note the persistence of SBS4/5 differences in the clonal stage and the emergence of SBS15/SBS86 in subclones. (D) Key evolutionary trajectories. Line plots (±SE) tracking the relative activity of specific signatures. SBS4 (Tobacco) functions as an initiating event (clonal) in Group 1 and decays over time. In contrast, SBS15 (MMRd) acts as a late-stage driver, showing a sharp increase in the trace subclones, particularly contributing to the high subclonal burden observed in Group 2. ns, not significant; **, P<0.01. Group 1: Clonal-dominant. Group 2: Subclonal-rich. FDR, false discovery rate; MMRd, mismatch repair deficiency; SE, standard error; TMB, tumor mutation burden.

To independently validate this chronological sequence, we performed cohort-wide clonal timing and hierarchical tree reconstruction (Figure S5). Mutations attributed to MMRd-associated signatures (SBS6/15/26) demonstrated a significantly lower median CCF compared to background clonal mutations (median CCF: 0.071 vs. 0.920; Mann-Whitney U test, P<0.001; Figure S5A). Furthermore, mapping these mutations to reconstructed phylogenetic hierarchies showed that while MMRd mutations account for only 7.8% of mutations on the early clonal trunk, they make up 22.2% of the mutations located on late subclonal branches (Figure S5B). GMM-derived “Clonal” category classifications showed a perfect 100% concordance (40/40 patients) with the trunk clusters derived from the pigeonhole principle, cross-validating the robustness of our subclonal stratification.

Genomic determinants of heterogeneity and driver dependency

We systematically interrogated the genomic landscape to identify genetic alterations underpinning these architectural differences. First, to assess the context of genomic stability, we evaluated whole-genome duplication (WGD) events and pathway-level mutation burdens, finding that neither the prevalence nor the timing of WGD differed between the groups (Figure S6). Similarly, the aggregate mutation burden in key signaling pathways also showed no difference (Figure S7A), reinforcing that the observed heterogeneity is driven by specific clonal architectures rather than broad genomic instability events.

Our analysis of copy number drivers revealed that Group 1 was enriched for MYC family amplifications, whereas Group 2 exhibited a nominally significant enrichment for broad FGFR1 gains (P=0.04, uncorrected; Figure 4A,4B). Detailed significant amplification and deletion peaks are listed in Table S3, although we note that differences in high-level MYC amplification did not reach statistical significance in this cohort (Figure S7B). Concurrently, we found that Group 2 tumors exhibited significantly higher MATH scores (P<0.001, Table 1 and Figure 4C). This elevated heterogeneity in Group 2 was intrinsically independent of TMB, in contrast to Group 1, where a negative correlation was observed (Figure S8). Despite these divergent trajectories, dN/dS analysis revealed that both groups operate under similar positive selection (w>1) for truncating mutations in TP53 and RB1 (Figure 4D), indicating a clear convergence on the inactivation of key tumor suppressors. Additionally, we evaluated the diagnostic immunohistochemical (IHC) marker expressions of synaptophysin (Syn), chromogranin A (CgA), and CD56 between the two evolutionary subgroups, which demonstrated no statistically significant differences (Figure S9, Table S7).

Figure 4.

Genomic determinants of heterogeneity and driver dependency. (A) Targeted copy number landscape heatmap. Rows represent key SCLC driver genes and tumor suppressors; columns represent individual patients sorted by genomic subgroup. The color intensity indicates copy number status derived from GISTIC2.0 [red: high amplification (+2); blue: deep deletion (−2)]. Group 1 shows a cluster of MYC amplifications, while Group 2 displays scattered FGFR1 and SOX2 alterations. (B) Driver amplification frequency. Bar chart displaying the percentage of samples with high-level amplification (GISTIC score ≥2). MYC high-level amplifications were predominantly found in Group 1. Notably, when considering broad copy number gains (GISTIC ≥1), Group 2 showed a nominally significant enrichment of FGFR1 (P=0.04), highlighting a distinct oncogenic dependency. (C) Intratumoral heterogeneity (MATH score). Boxplots comparing the distribution of MATH scores between Group 1 (blue) and Group 2 (red). Group 2 demonstrates significantly higher intrinsic heterogeneity (P<0.001, Wilcoxon rank-sum test). (D) Evolutionary selection pressure. Forest plot showing global dN/dS ratios (w) for all, missense, and truncating mutations calculated using dNdScv. Error bars represent 95% CI. “>1” indicates positive selection, “=1” indicates neutral evolution, “<1” indicates negative selection. Both groups exhibit significant positive selection (w>1), particularly in truncating mutations, indicating shared evolutionary pressure to inactivate tumor suppressors despite divergent clonal architectures. Group 1: Clonal-dominant. Group 2: Subclonal-rich. CI, confidence interval; dN/dS, ratio of non-synonymous to synonymous substitutions; GISTIC, Genomic Identification of Significant Targets in Cancer; MATH, Mutant-Allele Tumor Heterogeneity; MLE, maximum likelihood estimation; SCLC, small cell lung cancer.

Stage-dependent prognostic stratification by clonal architecture

The potential clinical relevance of these evolutionary subtypes was explored based on treatment response and survival. To minimize the confounding effect of heterogeneous treatments, we evaluated the ORR specifically within the evaluable patients who received standard first-line platinum-based regimens (n=24). Within this cohort, Group 1 patients showed a higher ORR compared to Group 2 (81.3% vs. 25.0%, P=0.02; Table 1 and Figure 5A). For patients receiving non-platinum-based regimens (n=4) or those with missing/untreated status (n=5), objective responses were either clinically not evaluable or the sample sizes were too small for meaningful statistical comparisons; thus, they were excluded from the formal ORR analysis.

Figure 5.

Clonal architecture is associated with therapeutic response and survival outcomes with a potential stage-dependent trend. (A) Chemotherapeutic response rates. Stacked bar chart showing the ORR within the sub-cohort of 24 patients who received first-line platinum-based regimens and had evaluable radiographic assessments (excluding 8 patients without valid post-baseline imaging). Group 1 exhibits a higher ORR compared to Group 2 (P=0.02, Fisher’s exact test). (B) Subgroup analysis and interaction test. Forest plot displaying the HRs (Group 2 vs. Group 1) stratified by clinical stage. A potential stage-dependent trend regarding the prognostic impact is observed (P=0.054). (C,D) Kaplan-Meier survival curves. Survival analysis was restricted to patients who received systemic therapy to minimize treatment confounding. In LS patients (C), Group 2 exhibits shorter OS compared to Group 1 (P=0.009), whereas in ES patients (D), this prognostic separation is not observed (P=0.81). P values were calculated using the log-rank test. (E) Permutation test. Histogram of P values from 1,000 random permutations of patient labels within the LS cohort. The red dashed line indicates the position of the original log-rank P value, which corresponds to an empirical P value of 0.009. This confirms the statistical robustness of the survival divergence observed in the LS subcohort. (F) Exploratory multivariate Cox regression analysis. Forest plot showing adjusted HRs for OS in the evaluable cohort. The model was adjusted for clinical stage, platinum-based treatment (Y vs. N), TMB, smoking history, and age. Group 2 status remains associated with an increased mortality risk (HR =2.93, P=0.02). Error bars represent 95% CIs. Group 1: Clonal-dominant. Group 2: Subclonal-rich. CI, confidence interval; ES, extensive-stage; HR, hazard ratio; LS, limited-stage; N, no; ORR, objective response rate; OS, overall survival; TMB, tumor mutation burden; Y, yes.

In the subgroup of patients with available survival data, subgroup analysis suggested a potential stage-dependent trend regarding the prognostic impact of the clonal architecture (interaction P=0.054; Figure 5B). Specifically, a sensitivity analysis restricted to patients who received systemic therapy indicated that in LS disease, Group 2 status was associated with shorter OS compared to Group 1 (P=0.009; Figure 5C). In contrast, this prognostic separation was not observed in extensive-stage (ES) disease, where outcomes were uniformly poor regardless of genomic subgroup (P=0.81; Figure 5D). To evaluate the statistical stability of the stratification observed in the LS cohort, we performed a permutation test with 1,000 iterations, suggesting that the observed survival difference is unlikely to be a purely random event (P=0.009; Figure 5E).

To evaluate the prognostic impact across the entire clinical spectrum, we constructed an exploratory multivariate Cox proportional hazards model encompassing the combined evaluable cohort (including both limited and extensive stages). After adjusting for clinical stage, platinum-based treatment usage, TMB, smoking history, and age, Group 2 status remained associated with an increased mortality risk [adjusted hazard ratio (adj. HR) =2.93; 95% confidence interval (CI): 1.21–7.10; P=0.02; Figure 5F]. However, given the retrospective design and limited sample size of this study, these findings should be interpreted cautiously and serve primarily as hypothesis-generating data for future validations.

Histology-dependent patterns in early-stage NSCLC cohorts

Finally, to explore the potential universality and lineage-specificity of our findings, we applied our framework to external cohorts from TCGA. We specifically chose LUSC as an exploratory cross-cancer cohort and LUAD as a negative control cohort. The rationale for selecting LUSC is its striking etiological and genomic resemblance to SCLC. Both are predominantly central-type tumors driven by heavy smoking exposure, which results in a similarly high TMB and frequent TP53 inactivation, distinguishing them from LUAD, which is more common in non-smokers and often driven by targetable oncogenes. These shared characteristics make LUSC a suitable context to explore whether our observed evolutionary architectures are specific to SCLC or represent a broader feature of smoking-induced lung cancers.

Applying our framework to the TCGA-LUSC cohort, we found that it identified similar Clonal-dominant vs. Subclonal-rich architectures (Figure 6A). Furthermore, in a parallel analysis restricted to early-stage (stage I–II) LUSC patients, those with high heterogeneity showed a marginal trend toward shorter OS (P=0.053, HR =1.39), partially reflecting our findings in LS-SCLC (Figure 6B). In contrast, this prognostic separation was not observed in the early-stage LUAD control cohort (P=0.85; Figure 6C). To rigorously assess these relationships, we performed multivariate Cox regression across the three stage-matched cohorts. Specifically, to parallel the early-stage (stage I–II) LUSC and LUAD validation cohorts, we restricted the SCLC discovery cohort exclusively to LS patients, filtering out all ES cases. Within this isolated LS-SCLC subset, after adjusting for available clinical confounders (age, gender/smoking history, and TMB status), high heterogeneity remained a significant independent adverse factor in LS-SCLC (adj. HR =9.65, P=0.006) and early-stage LUSC (adj. HR =1.40, P=0.04), but showed no prognostic association in early-stage LUAD (adj. HR =1.05, P=0.80; Figure 6D). This generates the hypothesis that the “heterogeneity-driven poor prognosis” phenotype is not a generic feature of high-TMB cancers but may be a specific histology-dependent pattern shared by smoking-associated squamous and neuroendocrine lineages.

Figure 6.

Exploratory cross-cancer analysis of clonal architecture and lineage-specific prognostic associations. (A) Hierarchical clustering of LUSC clonal profiles. Heatmap showing the clonal architecture of the TCGA-LUSC exploratory cohort. Rows represent the fractions of four evolutionary strata. The cohort can be grouped into two architectures resembling SCLC Group 1 (Clonal-dominant) and Group 2 (Subclonal-rich). (B) LUSC exploratory analysis (early stage). Kaplan-Meier curve for early-stage (stage I–II) LUSC patients. Similar to limited-stage SCLC, the high-heterogeneity Group 2 shows a marginal trend toward shorter OS (log-rank P=0.053). (C) LUAD negative control (early stage). Kaplan-Meier curve for early-stage (stage I–II) LUAD patients. The prognostic separation observed in SCLC and LUSC is not evident in this non-squamous/neuroendocrine cohort (log-rank P=0.85). (D) Prognostic concordance and specificity (multivariate analysis). Forest plot comparing the adj. HRs (high heterogeneity vs. low) across three distinct stage-matched cohorts. The top section shows the SCLC discovery cohort, which was strictly filtered to include only limited-stage patients (excluding all extensive-stage cases) to parallel the early-stage validation cohorts. This specific sub-model was adjusted for age, smoking history, and TMB status. The middle section shows the LUSC exploratory cohort (early stage), and the bottom section shows the LUAD control cohort (early stage), both adjusted for age, gender, smoking history, and TMB status. The clonal architecture model exhibits significant prognostic associations in SCLC and LUSC (HR >1) but shows no effect in LUAD (HR ≈1), suggesting a histology-dependent prognostic pattern, where high intratumoral heterogeneity serves as an adverse factor primarily in heavily smoking-associated phenotypes (SCLC and LUSC) rather than in LUAD. adj. HR, adjusted hazard ratio; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; OS, overall survival; SCLC, small cell lung cancer; TCGA, The Cancer Genome Atlas; TMB, tumor mutation burden.

Discussion

Our study advances the understanding of SCLC biology by shifting the analytical lens from a static catalog of driver mutations to a dynamic characterization of evolutionary architecture. A pivotal methodological innovation of this work is resolving the “cross-patient comparability” bottleneck. Standard clustering methods (e.g., Dirichlet processes) generate sample-specific subclonal fractions that cannot be horizontally aligned across a cohort (14,15). By applying an optimized population-level GMM—and empirically ruling out BMMs due to mixture collapse—we established a standardized, four-tier coordinate system for ITH. This probabilistic approach enabled the robust identification of recurrent evolutionary topologies that remain invisible in smaller, individual single-sample analyses. Our findings suggest that the specific pattern of heterogeneity, particularly the presence of an expanding, MMRd-driven subclone, may serve as an important contributor to adverse clinical outcomes.

To fully contextualize these findings within the broader landscape of lung malignancies, it is essential to re-evaluate the comparative biology of SCLC and LUSC. These subtypes share a striking convergence of etiological and anatomical features, including strong causal links to heavy tobacco exposure and central airway localization (16,17). Given this shared carcinogenic insult, parallel evolutionary trajectories might be expected. However, early genomic atlases historically portrayed SCLC as a relatively homogeneous entity characterized by a rapid “clonal sweep” of TP53 and RB1 inactivation (2,4), in sharp contrast to the immense subclonal diversity observed in LUSC by the TRACERx consortium (6,18). Our findings nuance this binary narrative by identifying a distinct architectural dichotomy within SCLC. We propose that the perceived homogeneity is not an immutable biological rule but rather a reflection of the Group 1 phenotype dominating previous studies. Conversely, our Group 2 tumors reveal an intrinsically unstable subtype that effectively mimics the chaotic evolutionary architecture of LUSC, suggesting that distinct histological subtypes may converge on similar evolutionary strategies to survive strong selection pressures (19,20).

The mechanistic basis for this convergence likely lies in the identification of MMRd signatures within the Trace Subclones of Group 2. While recent work by George et al. [2024] demonstrated that chemotherapy acts as a bottleneck selecting for resistant clones, our data elucidate the origin of this diversity. The presence of an intrinsic “mutator phenotype” suggests that Group 2 tumors generate a reservoir of subclonal variants prior to any therapeutic intervention. Indeed, emerging cohorts from 2025 and 2026 have increasingly substantiated that therapeutic bottlenecks select for pre-existing subclonal populations rather than inducing de novo mutations (8,21,22). This aligns with pan-cancer evolutionary principles where the acquisition of DNA repair defects acts as a catalyst for late-stage diversification (5,23). Consequently, the recurrence observed in these patients may represent a deterministic trajectory driven by their evolutionary architecture rather than a stochastic event. These tumors appear biologically primed for resistance, behaving less like classic, responsive SCLC and more like the refractory subclones described in chemo-resistant NSCLC (24,25). Notably, this explosive subclonal diversification is not merely a mathematical artifact of higher overall TMB. As our layer-by-layer modeling shows (Figure 3B), the mutation burden increase in Group 2 is not uniformly inflated but is entirely confined to the trace compartment and specifically driven by an acquired MMRd signature. This supports the view that the subclonal-rich architecture is a bona fide biological event rather than random statistical noise.

Translating these architectural insights into the clinic addresses a critical gap in LS-SCLC management, where transcription-based subtypes have shown limited prognostic utility (1,26). Our proposed evolutionary framework suggests a potential new dimension for risk assessment. For Group 2 patients, the standard concurrent chemoradiotherapy backbone may be insufficient given their inherent genomic instability, making them potential candidates for agents targeting replication stress, such as poly(ADP-ribose) polymerase (PARP) inhibitors (27,28), or consolidation immunotherapy (e.g., durvalumab) which has recently shown significant OS benefits in LS-SCLC (29). Clinically, the presence of MMRd necessitates a nuanced reconsideration of immunotherapy. Although SCLC is generally considered “cold” with modest checkpoint blockade benefits (30), the high neoantigen load implied by MMRd suggests a potential window of immunogenicity. Furthermore, the emerging landscape of bispecific T-cell engagers (such as tarlatamab) and predictive biomarkers for patient selection may offer tailored therapeutic opportunities for heterogeneous, refractory subgroups (31,32). However, as evidenced in NSCLC, high TMB does not guarantee immune recognition if the antigen presentation machinery is defective (33,34). Future studies must therefore investigate whether the genomic instability in Group 2 translates into genuine immune susceptibility or if these tumors employ additional evasion mechanisms like major histocompatibility complex (MHC) class I downregulation (35).

Beyond therapeutic implications, while ITH in other solid tumors is often initially evaluated via histopathology, SCLC presents a uniquely challenging paradigm. In our cohort, routine hematoxylin and eosin (H&E) staining exhibited classic, monotonous SCLC morphology across all samples, lacking distinct phenotypic sub-features. This morphological uniformity effectively masks the profound underlying genomic complexity. Standard histopathology cannot discern minor subclones harboring intrinsically unstable genomic engines, such as MMRd, prior to therapeutic pressure. Therefore, a sophisticated probabilistic deconvolution of deep sequencing data is strictly required to unmask these hidden evolutionary architectures that dictate clinical outcomes.

Relationship to established transcriptional subtypes

An important consideration is the relationship between the evolutionary subtypes defined here and the established transcriptional subtypes of SCLC (ASCL1, NEUROD1, POU2F3, and YAP1). Our cohort lacked RNA-sequencing data and systematic IHC profiling of these transcriptional markers, which we acknowledge as a limitation. However, we propose that our genomic evolutionary framework and the transcriptional subtype classification represent orthogonal, rather than redundant, dimensions of SCLC biology. Transcriptional subtypes capture the phenotypic differentiation state of the dominant tumor cell population (“who is expressing what transcription factor”), whereas our evolutionary framework captures the genomic architecture of clonal dynamics (“how the tumor evolved and diversified”). A tumor classified as ASCL1-positive, for instance, could harbor either a Clonal-dominant or a Subclonal-rich evolutionary architecture; these two axes are independently informative. This independence is reinforced by the observation that transcription-based subtypes have shown pronounced lineage plasticity and frequent subtype transitions (1,21,36), which limits their stability as fixed prognostic classifiers—precisely the clinical context where our genomic framework demonstrates the strongest prognostic signal. To empirically support this orthogonality, we analyzed diagnostic IHC data for classic neuroendocrine markers (Syn, CgA, and CD56) available for our cohort. Positivity rates for Syn (82.1% in Group 1 vs. 84.6% in Group 2; Fisher’s exact P>0.99), CgA (64.3% in Group 1 vs. 53.8% in Group 2; Fisher’s exact P=0.73), and CD56 (85.7% in Group 1 vs. 76.9% in Group 2; Fisher’s exact P=0.66) showed no significant differences between the two evolutionary subtypes (Figure S9, Table S7). This empirical evidence confirms that traditional clinical neuroendocrine markers cannot distinguish these subgroups, highlighting our evolutionary stratification as a new, independent prognostic dimension complementary to existing molecular classifications.

Limitation

This study has several limitations. First, given the retrospective design and the real-world setting during the coronavirus disease 2019 (COVID-19) pandemic, radiographic assessment intervals were not strictly standardized; thus, our clinical analyses primarily focused on objective response and OS to unveil the biological impact of clonal evolution. Second, the cohort size of 41 patients is relatively small, which is inherent to the clinical challenge of obtaining high-quality, treatment-naive tissue from early-stage SCLC patients before rapid systemic treatment initiation.

Most importantly, we acknowledge the lack of an independent, external real-world SCLC validation cohort. Currently, deep-sequenced, treatment-naive SCLC datasets with matched long-term clinical outcomes are virtually non-existent in the public domain. To circumvent this data scarcity, we leveraged TCGA-LUSC as a cross-cancer validation cohort based on the scientific rationale of their shared heavy-smoking etiology. While this successfully validates our lineage-specific hypothesis, we do not claim to have established a definitive clinical biomarker. Instead, our findings are primarily hypothesis-generating. Future efforts integrating larger, multi-center real-world SCLC cohorts and single-cell sequencing will be strictly required to prospectively validate this evolutionary architecture and its role in immune microenvironment interplay.

Conclusions

In summary, by integrating foundational genomic knowledge with a novel probabilistic analysis of subclonal architecture, we define a high-risk SCLC trajectory driven by intrinsic genomic instability. This evolutionary crossover from tobacco-driven initiation to MMRd-driven diversification provides a mechanistic framework for understanding intrinsic resistance. Our findings not only offer a hypothesis-generating framework for future biomarker development in this recalcitrant disease but also reframe the debate on SCLC heterogeneity, emphasizing that for the most aggressive tumors, the seeds of lethal recurrence are likely sown long before the initiation of therapy. It should be explicitly noted that our GMM-based evolutionary subtyping framework represents a hypothesis-generating approach derived from a single-center retrospective cohort. Prospective validation in larger, independent multi-center cohorts incorporating longitudinal sampling is required before any clinical implementation.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shanghai Pulmonary Hospital (Approval No. K18-203Y). Written informed consent was obtained from all patients participating in the study.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1338/dss

jtd-18-08-866-dss.pdf ^{(92.5KB, pdf)}

DOI: 10.21037/jtd-2026-1338