Spatial transcriptomics reveals molecular heterogeneity and subtype-specific therapeutic targets in small cell lung cancer.

Introduction
Small cell lung cancer (SCLC) is a highly invasive and poorly prognostic malignant tumor with neuroendocrine (NE) characteristics, accounting for ~15% of all lung cancers1–3. SCLC stands out as one of the most aggressive forms of cancer, notably associated with smoking, which accounts for over 95% of SCLC cases. Consequently, there exists a higher prevalence of SCLC among males. Nevertheless, the gender-based incidence gap has progressively diminished over the last three decades4–6. The 5-year survival rate for SCLC is only 5–10%, and at the time of initial diagnosis, 60–70% of SCLC patients are diagnosed with extensive-stage (ES-SCLC)6,7.
Current clinical management of SCLC frequently relies on disease stage, with previous treatments often lacking a pre-defined patient subgroup delineation. One distinctive feature of SCLC is its characteristic high sensitivity to platinum-based chemotherapy, often accompanied by swift recurrence. Moreover, second-line treatment with alternative chemotherapeutic agents or immunotherapy is generally only marginally effective. However, the evolutionary processes are unknown6,8–10. Consequently, several investigations have emerged aiming to identify and delineate biologically distinct SCLC subtypes. A growing body of studies utilizing patient-derived tissues, cancer cell lines, and murine models has tentatively confirmed the existence of predominant subtypes in SCLC, alongside intratumoral subtype heterogeneity—findings that collectively support the occurrence of dynamic subtype evolution within patient tumors. For instance, studies have shown that MYC drives the dedifferentiation of tumor cells by activating the Notch signaling pathway, thereby facilitating a sequential temporal transition of SCLC from the ASCL1+ state to the NEUROD1+ state, and ultimately to the YAP1+ state11–14. Furthermore, the tissue samples of patients with SCLC appear more heterogeneous than expected due to the high biological plasticity of this malignancy and its ability to adapt to different growth conditions.
In this study, we establish a comprehensive spatial transcriptomic (ST) landscape of SCLC through the integration of two key methodological innovations: the Edgeindex metric for quantitative assessment of tumor spatial architecture, and a specialized artificial neural network (ANN) model for precise tumor annotation. This analytical framework enables systematic resolution of SCLC heterogeneity across clinical, spatial, functional, and temporal dimensions, and further pathway enrichment analysis was performed to explore the underlying mechanism.

Results

Patient characteristics
Table S1 summarizes the baseline characteristics of the 21 patients involved in this study. All patients were diagnosed with limited-stage SCLC and had undergone surgical resection without receiving any prior neoadjuvant therapy. The cohort had a higher proportion of males (~71.4%) compared to females. To quantitatively assess the tumor-stroma interface, we developed a spatial metric termed the “Edgeindex” (see “Methods” for detailed calculation). The average Edgeindex across patients was about 0.08. Based on the transcriptional characteristics of tumors and non-tumor, patients can be categorized into 3 clusters. We analyzed the expression patterns of the transcription factors ASCL1, NEUROD1, POU2F3, and YAP1 across all samples, a framework widely utilized in contemporary SCLC research to dissect molecular heterogeneity15–17. The ASCL1 (47.6%) subtype was the most common across the cohort. The median intratumoral (tumor spot), microenvironment (non-tumor spot) and total heterogeneity value based on the number of spot cluster was 3, 3 and 6, respectively. The median cell-cell communication count and weight was 10,910 and 117, respectively.

Gender heterogeneity in SCLC
We hypothesized that gender-related biological differences might contribute to variations in tumor phenotype. To explore this possibility, we compared the transcriptomic and pathway profiles between male and female tumors. Although the sample size is limited, we observed several gene expression and pathway trends that may provide preliminary clues regarding potential sex-associated molecular features in SCLC. Among tumor spots, the top five female-related marker genes were PAEP, SCGB3A1, DRGX, FGG, and SST, while the top five male-related marker genes were EIF1AY, HIST1H1A, DEFA5, CXCL14, and VSTM2B (Fig. 1A). For other spots, the top five female-related marker genes were SST, CGA, SCGN, PCSK1, and SLC38A11, and the top five male-related marker genes were EIF1AY, PRB3, BPIFB1, PRB4, and DDX3Y (Fig. 1A).
Gene set enrichment analysis (GSEA) results for gender in tumor areas were shown in Fig. S1A. Figure 1B highlights only the gene sets with statistical significance, revealing that the “Extracellular matrix (ECM) & metastasis” gene sets were enriched in females. For non-tumor areas, Fig. S1B presents the GSEA results, while Fig. 1C displays those with statistical significance, indicating that the “Immunity” gene sets were enriched in males. Detailed GSEA results for all 1350 gene sets in both tumor and non-tumor areas are provided in Table S2.

Spatial structure heterogeneity in SCLC
Infiltrative growth refers to the spread of tumor cells into and mingling with surrounding other tissues. This growth pattern associated with invasion and metastasis. To quantify this, we applied the Edgeindex, a spatial metric (as defined in the Methods).
In our study, we observed 21 samples displaying distinct histopathological growth patterns (Fig. 2A and Fig. S2A). The Edgeindex quantifies the degree of tumor-stromal mixing, calculated as the proportion of non-tumor spots among the neighbors of each tumor spot, then averaged across all tumor spots in a sample. A higher Edgeindex indicates a more infiltrative growth pattern, characterized by scattered tumor areas and irregular, elongated tumor margins. This variation in growth patterns may reflect differing underlying biological mechanisms. To explore this further, we calculated the correlations between gene expression and the Edgeindex. Figure 2B shows the top five genes that are positively and negatively correlated with the Edgeindex in both tumor and non-tumor areas. Among tumor spots, the top five Edgeindex-related marker genes were PLEKHG6, LTBR, GATA6, CD9, and IL20RA. In contrast, in other spots, the top five Edgeindex-related marker genes were PGM3, GORASP2, ERGIC2, PLOD2, and ARF1.
GSEA results for Edgeindex in tumor areas were shown in Figs. S2B and 2C. Figure 2C highlights only the gene sets with statistical significance, revealing more “ECM & metastasis”, “Immunity”, and “Cell death” pathways were associated with high Edgeindex, while more “Cell Cycle” pathways were associated with low Edgeindex. For non-tumor areas, Figs. S2C and 2D present the GSEA results, with Fig. 2D showing those with statistical significance, indicating more “Cell Cycle”, “Cell Death” and “Genetic and epigenetic information” pathways were associated with high Edgeindex, while more “ECM & metastasis” pathways were associated with low Edgeindex. Detailed GSEA results for all 1350 gene sets related to the Edgeindex are provided in Table S3.
Analysis of the spatial infiltration pattern, as quantified by the Edgeindex, revealed a notable association with the ANPY molecular subtypes of SCLC (Fig. 2E). The high-NE subtypes (ASCL1 and NEUROD1) exhibited significantly lower Edgeindex values, consistent with compact tumor growth, whereas the low-NE subtypes (POU2F3 and YAP1) showed higher values, indicative of a more dispersed or infiltrative growth pattern.
Our analysis revealed that a higher Edgeindex was significantly associated with a distinct immune contexture in the tumor microenvironment (TME). The specific statistically significant correlations (P < 0.05) are listed below: Positive correlations: CD56dim natural killer cell_ssGSEA (r = 0.574, P = 0.007), Monocyte_ssGSEA (r = 0.526, P = 0.016), Central memory CD4 T cell_ssGSEA (r = 0.516, P = 0.018), and Macrophages M0_CIBERSORT (r = 0.564, P = 0.008). Negative correlations: Eosinophil_ssGSEA (r = −0.465, P = 0.035), Mast cell_ssGSEA (r = −0.644, P = 0.002), Neutrophil_ssGSEA (r = −0.696, P = 0.0006), Type 1 T helper cell_ssGSEA (r = −0.449, P = 0.042), Type 17 T helper cell_ssGSEA (r = −0.626, P = 0.003), Type 2 T helper cell_ssGSEA (r = −0.491, P = 0.025), Dendritic cells activated_CIBERSORT (r = −0.607, P = 0.004), and NK cell_quanTIseq (r = −0.487, P = 0.027).
In addition, to identify tumor spots for ST data of SCLC, an ANN model was constructed (Figs. 2F and S2D). Ten of the 21 samples were randomly divided into the discovery cohort for training and testing the ANN model, and the remaining samples were the validation cohort for blind test. Our analysis shows that the ANN-predicted tumor regions largely correspond to those identified by histopathology, with concordance rates ranging from 69.1% to 97.9% across samples. 15 in 21 samples had the concordance rates over 85%. This demonstrates a high level of agreement between the ANN model and the pathologist-defined tumor boundaries.

ST heterogeneity in SCLC
We conducted an unsupervised clustering analysis to categorize all samples into distinct clusters. The clustering outcomes on the basis of ST data from the tumor areas are depicted in Fig. 3A, while those on the basis of ST data from the non-tumor areas are presented in Fig. 3B. The samples were classified into 3 and 3 clusters, based on the ST data of the tumor areas and non-tumor areas, respectively.
To further investigate the 3 clusters identified from the tumor and non-tumor areas, we compared the gene expression patterns among them. Figure 3C illustrates the top five differentially expressed genes in each pairwise cluster comparison.
GSEA was performed for the 3 clusters from the tumor areas, as displayed in Figs. S3A, S3B, and 3D. Figure 3D specifically emphasizes gene sets with statistically significant enrichment uncovered almost similar trends between results for spots in the tumor and non-tumor areas. For the 3 clusters based on non-tumor areas, Figs. S3C, S3D, and 3E display the GSEA results, with Fig. 3E highlighting those with statistical significance, which showed similar trends between results for spots in the tumor and non-tumor areas. Detailed GSEA results associated with these clusters for all 1350 gene sets are detailed in Table S4.

Classical transcription factor defined subtypes (ANPY) in SCLC
We analyzed the expression patterns of the transcription factors ASCL1, NEUROD1, POU2F3, and YAP1 across all samples. Figure 4A shows the expression distribution of these genes. The initial 10 samples, characterized by high ASCL1 expression, were classified as the ASCL1 subtype. Subsequently, the next four samples with elevated NEUROD1 expression were categorized as the NEUROD1 subtype. The following three samples with prominent POU2F3 expression and four samples with significant YAP1 expression were designated as the POU2F3 and YAP1 subtypes, respectively. To delve deeper into these four subtypes, we conducted a comparative analysis of their gene expression profiles. Figure 4B presents the top five differentially expressed genes in each pairwise subtype comparison, revealing distinct expression patterns among the subtypes.
GSEA was conducted for the four subtypes within the tumor areas, as shown in Figs. S4A and 4C. Figure 4C, in particular, highlights gene sets with statistically significant enrichment. It reveals that the ASCL1 and NEUROD1 subtype are similar and associated with “Cell Cycle” pathways. The POU2F3 subtype showed enrichment in “Immunity” pathways, while the YAP1 subtype was linked to pathways involving “Metabolism and energy”, and “ECM & Metastasis”. GSEA was also performed for the four subtypes in non-tumor areas, as depicted in Figs. S4B and 4D, with Fig. 4D emphasizing those with statistical significance. The ASCL1 subtype showed connections to pathways related to “Immunity”, “Cell Death”, and “ECM & Metastasis” when compared with the NEUROD1 subtype. The ASCL1 and NEUROD1 subtypes showed similar results when compared with POU2F3 and YAP1 subtypes. The POU2F3 subtype exhibited enrichment in pathways concerning “Immunity”, “Cell Cycle” and “Genetic and Epigenetic Information”. Lastly, the YAP1 subtype was associated with “ECM & Metastasis” pathways. A comprehensive set of GSEA results related to the four subtypes within the tumor areas and non-tumor areas is provided in Table S5.

Intratumoral and microenvironment heterogeneity in SCLC
We conducted separate clustering analyses on tumor and non-tumor regions within each sample, categorizing them into distinct clusters based on their expression patterns as shown in Figs. 5A and S5A. To delve deeper into the expression patterns associated with intratumoral (number of tumor clusters) and microenvironment heterogeneity (number of non-tumor clusters), we analyzed the correlation between these heterogeneity values and gene expression profiles. Figure 5B highlights the top five genes correlated with intratumoral heterogeneity value. GSEA was applied for intratumoral heterogeneity in the tumor areas, as depicted in Figs. S5B and 5C, revealing more “Immunity” pathways were associated with high heterogeneity while more “Cell Cycle” pathways were associated with low heterogeneity. GSEA results in the non-tumor areas, as depicted in Figs. S5C and 5D, revealing more “Cell Cycle” pathways were associated with low heterogeneity.
Subsequently, we analyzed the microenvironment heterogeneity of SCLC. Figure 5E identifies the top five genes correlated with microenvironment heterogeneity value. GSEA results for microenvironment heterogeneity in the tumor areas, as shown in Figs. S5D and 5F, uncovered more “Cell Cycle” pathways were associated with high heterogeneity, while more “Immunity”, “Cell Death”, and “ECM & Metastasis” pathways were associated with low heterogeneity. GSEA results in the non-tumor areas, as depicted in Figs. S5E and 5G, didn’t reveal significant bias of pathway enrichment among microenvironment heterogeneity.
Lastly, we calculated the sum of clusters from tumor and non-tumor spots as the total heterogeneity. Figure 5H presents the top five genes correlated with total heterogeneity value. GSEA was performed for the total heterogeneity in the tumor areas, as illustrated in Figs. S5F and 5I, didn’t reveal significant bias of pathway enrichment. GSEA results in the non-tumor areas, as depicted in Figs. S5G and 5J, revealing more “Genetic and epigenetic information” pathways were associated with low heterogeneity. A comprehensive compilation of GSEA results related to heterogeneity of SCLC is detailed in Table S6.

Spatial signal communication heterogeneity in SCLC
We analyzed the patterns of cell-cell communication across various clusters within each sample, with Figs. 6A and S6A illustrating the communication counts within each cluster. To further explore the expression patterns associated with communication count, a comparative gene expression analysis was conducted. Figure 6B identifies the top five genes correlated the count of cell-cell communications. GSEA was then applied for the cell-cell communication count in the tumor areas, as shown in Figs. S6B and 6C, uncovering more “Immunity” pathways were associated with high cell-cell communication count, while more “Genetic and epigenetic information” and “Cell Cycle” pathways were associated with low cell-cell communication count. GSEA results in the non-tumor areas, as depicted in Figs. S6C and 6D, revealing similar results to the tumor areas.
Subsequently, we calculated the cell-cell communication weight among different clusters within each sample (Fig. 6E and S6D). Comparative analysis of gene expression profiles with the cell-cell communication weight was performed. Figure 6F presents the top five correlated with the cell-cell communication weight. GSEA was conducted for the cell-cell communication weight in the tumor areas, as depicted in Figs. S6E and 6G, revealing more “Immunity”, “Cell Death” and “ECM & Metastasis” pathways were associated with high cell-cell communication weight. GSEA results in the non-tumor areas, as depicted in Figs. S6F and 6H, revealing more “Cell Cycle” pathways were associated with high cell-cell communication weight, while more “ECM & Metastasis” pathways were associated with low cell-cell communication weight. A comprehensive compilation of GSEA results related to communication count and weight is detailed in Table S7.

Temporal developmental heterogeneity in SCLC
We conducted a pseudotime analysis to trace the developmental trajectories across each sample, with Fig. 7A depicting the key genes that define the pseudotime trajectory. A gene of particular significance, UCHL1, was identified for its prominent role in the pseudotime trajectory across all samples. Figures 7B and S7A visualizes the pseudotime trajectory of UCHL1 across the samples. To delve deeper into the expression patterns between UCHL1 expression, a comparative gene expression analysis was performed. Figure 7C pinpoints the top five genes correlated with UCHL1 expression. GSEA was subsequently applied for UCHL1 expression levels in the tumor areas, as illustrated in Figs. S7B and 7D, revealing more “Cell Cycle” pathways were associated with high UCHL1 levels, while more “Immunity”, “Cell Death” and “ECM & Metastasis” pathways were associated with low UCHL1 levels. GSEA results from the non-tumor areas, as depicted in Figs. S7C and 7E, revealing more “Immunity” and “Cell Cycle” pathways were associated with high UCHL1 levels, while more “ECM & Metastasis” pathways were associated with low UCHL1 levels. Additionally, Fig. 7F presents an analysis of the expression patterns of four transcription factors—ASCL1, NEUROD1, POU2F3, and YAP1—alongside UCHL1 across all samples, providing further insights into the regulatory mechanisms that govern the pseudotime trajectory in SCLC. A comprehensive compilation of GSEA results related to UCHL1 is detailed in Table S8.

Discussion
SCLC is an aggressive malignancy characterized by a poor overall prognosis. There remain limited effective molecular targets or targeted therapies available for SCLC. Most of our current understanding of SCLC has been derived from studies utilizing model-centric approaches, contributing to a translational gap between basic research findings and clinical outcomes. Genomic analyses have significantly expanded our knowledge of the molecular mechanisms underlying this aggressive disease11,14,17–21. To bridge this gap and provide a more clinically relevant interpretation of SCLC biology, we focused our analysis on cancer-related pathways. Specifically, we have presented the GSEA results of 1350 pathways across six functionally defined categories; the complete enrichment results for all comparisons are provided in Table S9.
In this study, we utilized ST data to dissect the molecular heterogeneity of SCLC, revealing distinct subtypes and their associated biological pathways. Our work contributes to a rapidly evolving landscape of SCLC research that leverages spatial technologies to decode tumor complexity. For instance, recent studies have powerfully illustrated how the TME shapes NE cell states, with CAFs being a key driver of plasticity and immune exclusion22, and how spatial crosstalk between epithelial subtypes and myeloid cells via the MIF-SPP1 axis underlies immunotherapy resistance23. In contrast to these TME-centric perspectives, our study provides a complementary, tumor-centric view by rigorously defining the intrinsic molecular features and spatial organization of the core SCLC transcriptional subtypes (ANPY). Our findings, consistent with the established molecular taxonomy of SCLC11,20, highlight the pivotal role of four key transcription factors—ASCL1, NEUROD1, POU2F3, and YAP1—in defining biologically distinct subtypes. The clinical significance of this classification lies in its ability to pinpoint subtype-specific therapeutic vulnerabilities, thereby transforming these subtypes into actionable frameworks for precision oncology24. For instance, the ASCL1 (SCLC-A) subtype, which predominated in our cohort, is intrinsically linked to high expression of druggable targets such as DLL3 (amenable to antibody-drug conjugates like Rova-T) and BCL-2 (vulnerable to BCL-2 inhibitors). Conversely, the NEUROD1 (SCLC-N) subtype exhibits susceptibility to AURKA inhibitors, while the POU2F3 (SCLC-P) and YAP1 (SCLC-Y) subtypes show dependencies on IGF-1R/PARP inhibition and immune checkpoint blockade, respectively. A key advantage of our ST approach is its capacity to resolve the dominant driver subtype within a tumor specimen, overcoming the limitations of bulk analyses, which can be obscured by intra-tumoral heterogeneity. This precise subtyping is a critical prerequisite for the future application of these promising subtype-specific therapeutic strategies. This functional spatial profiling of the ANPY subtypes reveals subtype-intrinsic biological programs, such as cell cycle regulation in ASCL1, which extends beyond previous classification schemes and offers a deeper mechanistic understanding. The identification of elevated Edgeindex scores in infiltrative tumors underscores the aggressive nature of SCLC and its propensity for metastasis. The development of this quantitative spatial metric provides a novel tool to assess tumor-stroma interplay and invasive potential, offering a distinct angle from the specific TME-driven mechanisms of progression. Furthermore, our ANN model demonstrated high accuracy in identifying tumor regions, underscoring the potential of machine learning in ST analysis.
The exploration of tumor-immune interactions and cell-cell communication dynamics provided critical insights into the TME. Our observation that increased cell-cell communication is broadly associated with immune responses aligns with the central role of the TME in SCLC. This complements and contrasts with other spatial studies that have identified highly specific resistance mechanisms, such as the spatial cascade of epithelial-myelial crosstalk driving the Epi-Ⅱ to Epi-I conversion23, by highlighting that heightened communication is a pervasive feature across SCLC spatial contexts. Pseudotime analysis further revealed the pivotal role of UCHL1 in tumor development, implicating it in cell cycle regulation. UCHL1 represents a promising cell-intrinsic therapeutic target within the tumor cells themselves, a focus that is orthogonal to the extrinsic, TME-focused targets or the immune niche components identified in other studies22,25.
Our results align with prior studies identifying ASCL1 as the most prevalent subtype in SCLC12,20,26. However, our ST analysis revealed that the ASCL1 subtype is uniquely associated with pathways involved in cell cycle regulation, findings not previously reported. The NEUROD1 subtype, also associated with cell cycle pathways, corroborates earlier findings suggesting a role for NEUROD1 in tumor invasion and metastasis27. The YAP1 subtype, linked to immunity, cell death, and ECM remodeling, further emphasizes the biological diversity of SCLC and underscores the need for subtype-specific therapeutic strategies.
Through ST analysis, we observed that the boundaries between tumor and non-tumor regions exhibited an infiltrative growth pattern. Subsequently, we developed a novel scoring system, termed Edgeindex, to quantitatively evaluate the intensity of tumor infiltration and assess its metastatic and invasive potential into adjacent tissues. Among tumor regions, the top five marker genes most strongly associated with a high Edgeindex score were identified as PLEKHG6, LTBR, GATA6, CD9, and IL20RA. Significant correlations were observed between the Edgeindex and gene sets related to ECM remodeling and metastasis, immune responses, and cell death pathways.
LTBR, a member of the tumor necrosis factor receptor superfamily, plays critical roles in secondary lymphoid organ development, host defense, chemokine production, and apoptosis28,29. LTBR signaling has been implicated in tumor progression, and its inhibition has been shown to exert anti-tumor effects30. GATA6 regulates key signaling pathways and influences development, differentiation, and carcinogenesis in the pancreas, lung, and other organs31. CD9 is a cell surface glycoprotein belonging to the tetraspanin superfamily, playing crucial roles in cellular growth, motility, and signal transduction32. Several studies have reported a positive correlation between CD9 expression and metastatic potential33–35. IL20RA, a member of the type Ⅱ cytokine receptor family, has been shown to regulate cancer progression by promoting an immunosuppressive TME through the reduction of CD8+ T-cell infiltration36,37. In addition, the functional annotation and pathway enrichment analyses of these top marker genes further supported the association between high Edgeindex scores and tumor invasiveness, suggesting their potential utility as biomarkers for predicting metastatic behavior in SCLC.
Based on the relationships among scores, the microenvironment, and pathways, we propose a novel model of “Immunological Niche Remodeling and Functional Compensation.” We interpret the concurrent positive correlation with CD56dim NK cells and negative correlations with multiple T-helper subsets as evidence of a fundamental shift in the immune landscape. The infiltrative phenotype (high Edgeindex) appears to foster an environment that is specifically suppressive towards the adaptive T-cell response, effectively creating an “immunological vacuum.” Into this vacuum, we observe a relative enrichment of certain innate immune cells, particularly the cytotoxic CD56dim NK cell subset. This suggests a potential, albeit likely dysfunctional, compensatory engagement of the innate immune system. This model is powerfully supported by our complementary GSEA results: the enrichment of “Cell Death” pathways in high-Edgeindex tumor spots points to a source of antigen and stress ligands that could simultaneously drive both T-cell exhaustion and NK cell recruitment. Therefore, the Edgeindex captures a spatial ecosystem where the tumor has remodeled the immune milieu, switching from an adaptive-dominated to an innate-dominated, yet still ineffective, state of immune engagement, providing a profound mechanistic insight into the barriers to effective immunotherapy.
The observed association between increased cell-cell communication, immune responses, and tumor progression underscores the importance of targeting the TME in SCLC treatment. Our findings further highlight the potential of ST to elucidate the molecular mechanisms underlying SCLC heterogeneity. The pronounced heterogeneity of SCLC reflects the complexity of this disease13,38,39.
SCLC is an aggressive NE tumor. Our pseudotime analysis demonstrates that UCHL1 plays a prominent role in the developmental trajectories of SCLC. UCHL1 is a NE cell-specific deubiquitinating enzyme that removes ubiquitin from ubiquitinated proteins and is specifically expressed in neurons and cells of the diffuse NE system40,41. Several reports suggest that UCHL1 is upregulated and serves as a critical regulator in the progression of lung cancers, implicating it as a key molecule in tumor cell invasion42,43. Thus, UCHL1 may represent a promising therapeutic target for SCLC intervention. Its elevated expression in SCLC suggests it may represent a promising therapeutic target for future intervention.
Our study establishes a comprehensive ST landscape of SCLC through two key innovations: the Edgeindex metric quantifying tumor spatial architecture, and a specialized ANN model for tumor annotation. This framework systematically resolves SCLC heterogeneity across clinical, spatial, functional, and temporal dimensions, revealing novel phenotypes and identifying UCHL1 as a pivotal differentiation regulator. These findings provide a multi-dimensional resource and conceptual advance for developing personalized SCLC therapeutics.

Methods

Patient samples
Formalin-fixed paraffin-embedded (FFPE) tissue blocks were collected from patients who had undergone lung resection and were pathologically diagnosed with SCLC at the Cancer Hospital, Chinese Academy of Medical Sciences (CHCAMS) in Beijing, China, in compliance with institutional ethical guidelines and following informed consent from the patients. The study protocol received approval from the Ethics Committee of Cancer Hospital, CAMS (No. 23/262-4004). This research was conducted in accordance with all pertinent ethical regulations, including the principles outlined in the Declaration of Helsinki.

ST sequencing
FFPE blocks were obtained from SCLC patients, and 5 μm sections of these samples were mounted onto slides. These slides were then incubated at 42 °C for 2 h and allowed to air dry at room temperature. Subsequently, they were further dried for 3 h at 60 °C. Hematoxylin (Dako, Part number S330930-2) and eosin (Sigma-Aldrich, Product number HT110216) were employed for H&E staining, with the staining times adjusted based on the tissue type. Following staining, ~100 µL of 85% Glycerol (Thermofisher, Catalog number 15514011) was added, coverslips were applied, and tissue imaging was performed. The coverslips were removed using a beaker filled with Milli-Q water.
The Visium slide was then placed into a cassette, and 100 µL of 0.1 N HCl (Sigma-Aldrich, Product number H1758) was added to each well. The slide was incubated at 42 °C for 15 min, after which the HCl was removed, and decrosslinking buffer was added. Subsequently, the slide was incubated at 95 °C for 1 h. The pre-hybridization step was carried out in accordance with the Visium Spatial Gene Expression for FFPE reagent kit protocol (10× Genomics, User Guide CG000407 Rev C, human transcriptome Product number 1000338). Specifically, 100 µL of Pre-hybridization mix was added to each well and incubated at room temperature for 15 min. Following incubation, the Pre-hybridization mix was aspirated, and 100 µL of Hybridization mix was added. The Visium slide was then incubated with the Hybridization mix overnight at 50 °C.
Subsequent steps of library preparation, including probe ligation, probe release and extension, probe elution, and FFPE library construction, were conducted following the user guide of the “Visium Spatial Gene Expression for FFPE reagent kit” (10× Genomics, User Guide CG000407 Rev C, mouse transcriptome Product number 1000339, human transcriptome Product number 1000338). Finally, the finished libraries were sequenced on Novaseq6000 (Illumina), with read 1 and read 2 lengths of 28 base pairs and 91 base pairs, respectively.
The raw ST data were normalized by the “Seurat”44 package for the following analysis. The identification of “tumor spots” was performed by integrating tumor purity scores derived from the “ESTIMATE”45 algorithm with morphological assessment of H&E-stained sections. Spots classified as tumor-associated were required to exhibit both high computational purity and concordant histopathological features, while all remaining spots were designated as “other spots”.

Characters identification
Characters of clinical, spatial, and transcriptomics information were identified. Gender information was obtained from electronic medical records of the CHCAMS. To quantitatively evaluate the infiltrative growth pattern of tumors, we developed a spatial metric termed the “Edgeindex”. This index is calculated by first determining the local microenvironment for each tumor spot: the ratio of adjacent non-tumor (“other”) spots to all neighboring spots is computed, yielding a local infiltration score. Tumor spots are then functionally classified as “Core” (score = 0, fully embedded), “Border” (0 < score ≤ 1, at the interface), or “Island” (no neighbors, highly isolated). The sample-level Edgeindex is derived by summing the local infiltration scores of all “Border” spots and normalizing this sum by the total number of spots in the sample. This final value, ranging from 0 to 1, provides a continuous measure of tumor-stroma interaction, where a higher Edgeindex indicates a more irregular, dispersed tumor boundary and greater infiltrative potential.
For characters of transcriptomics information, two calculating methods were used. In bulk calculating method, average expression values of each gene among both tumor and other spots were used respectively. Samples were grouped on the basis of all genes expression among tumor spots, all genes expression among other spots, and four transcription factors including ASCL1, NEUROD1, POU2F3, and YAP1 expression among tumor spots, respectively. In spatial calculating method, tumor and other spots were clustered on the basis of all genes expression, respectively. Samples were grouped on the basis of heterogeneity (cluster number), communication (cell-cell communication count and weight), and differentiation (UCHL1 expression) levels, respectively.

Clustering analysis
In both bulk and spatial calculating method, ST data were scaled and used to perform principal component analysis. Then, the principal components were used to perform uniform manifold approximation and projection to further dimension reduction. K-means method was used to group samples or spots into different clusters, the optimal number of clusters was determined by the “NbClust”46package. To quantitatively assess transcriptional heterogeneity, we defined intratumoral coupled with microenvironment heterogeneity as the number of robust transcriptional clusters identified among tumor spots and other spots, respectively. Moreover, we also calculated the sum of cluster in tumor and other spots as the total heterogeneity. A higher cluster count reflects a greater diversity of gene expression programs within the respective tissue compartment.

Cell-cell communication analysis
CellChat was applied to the STs data primarily to infer potential signaling networks among transcriptionally defined cell populations, consistent with established analytical approaches in recent STs studies. Cell-cell communication analysis was performed by the “CellChat”47 package with the “CellChatDB.human” database as a reference. The interaction counts and weights derived from CellChat were used to quantify signaling strength between transcriptional clusters in the TME. For each sample, the overall communication level was defined as the sum of interaction counts or weights among all tumor and non-tumor clusters. This communication metric served as a basis for defining distinct functional phenotypes, enabling the classification of samples according to dominant inter-cluster signaling patterns.

Pseudotime analysis
Pseudotime analysis was performed by “monocle”48 package. To identify genes associated with SCLC differentiation trajectories, we analyzed results from the “orderCells” function under both directional assumptions (parameter “reverse” set to “TRUE” and “FALSE”). Genes with an adjusted P < 1 × 10⁻⁸ in pseudotime-associated differential expression tests were initially selected. We further refined this set by retaining genes that recurred in over 30 of the 42 pseudotime analyses (21 samples × 2 directions) and exhibited statistically significant correlation (P < 0.05) with pseudotime values across all samples. Final candidate selection was based on recurrence and biological plausibility. This multi-step filtering strategy ensured the identification of robust, recurrence-based marker genes linked to SCLC differentiation, independent of arbitrary trajectory orientation.

Character related marker genes and GSEA
Average gene expression values among tumor and other spots of each sample were used to find character related marker genes, respectively. For categorical characters including Gender, “Tumor Cluster”, “Other Cluster”, and SCLC ANPY subtype, log2FC between two different levels was used. For continuous characters including “Edgeindex”, heterogeneity level, communication level, and differentiation level, Spearman correlation coefficient was used. GSEA was performed by “clusterProfiler” package49, the gene lists were ranked by log2FC of categorical characters or by Spearman correlation coefficient of continuous characters. Information of gene sets in “HALLMARK”, “Kyoto Encyclopedia of Genes and Genomes (KEGG)” and “Gene Ontology (GO)” was obtained from www.gsea-msigdb.org/gsea/msigdb/index.jsp. A total of 1350 gene sets associated tumor and TME were analyzed as a point of importance. The 1350 gene sets were summarized into six classes by us according to the information of these gene sets provided on the molecular signatures database (MSigDB), including 421 “Immunity”, 127 “Cell cycle”, 315 “Metabolism & energy”, 286 “Genetic and epigenetic information”, 129 “ECM & metastasis” and 72 “Cell death” gene sets. The gene sets with raw P < 0.05 & false discovery rate (FDR) < 0.25 were considered as statistically significant.

Tumor immune microenvironment analysis
To quantify immune cell abundance, we applied three independent algorithms to the gene expression data from non-tumor (“other”) spots: CIBERSORT, ssGSEA and quanTIseq. Subsequently, for each sample, we calculated the median abundance (score or fraction) for each immune cell type across all its “other” spots. Finally, we performed Spearman’s rank correlation analysis between these sample-level median immune cell abundances and the sample-level Edgeindex to assess their association.

ANN model construction
To identify tumor spots for ST data of SCLC, an ANN model was constructed. Ten of the 21 samples were randomly divided into the discovery cohort for training and testing the ANN model, and the other samples were the validation cohort for blind test. In the discovery cohort, 2000 spots were randomly divided into the training cohort, the remaining spots of the 10 samples were into the test cohort. Receiver operating characteristic (ROC) analysis was used in the training cohort to select genes related to the type of spot, and the significant (P < 0.05, area under the curve [AUC] of ROC curve > 0.8 or <0.2) genes were used to construct ANN model. The “neuralnet” package (https://CRAN.R-project.org/package=neuralnet) was used to construed ANN model. The number of hidden neurons was based on Nh = (4n2 + 3)/(n2 − 8) (Nh, the number of hidden neurons; n, the number of input neurons). Finally, the accuracy of the ANN model was tested in the test and validation cohort. The best cut-off value of this ANN model was determined by the results of ROC analysis for all spots from 21 samples.

Statistical analysis
R software (version 4.2.1, https://www.r-project.org) was employed for all the statistical analysis. Loupe Browser (version 6.5.0, https://support.10xgenomics.com) was used to assist tumor spots identification. Spearman correlation analysis was performed to assess the correlations between two continuous variables. In this study, raw P < 0.05 & FDR < 0.25 were considered statistically significant.

Ethics approval and consent to participate
The study protocol received approval from the Ethics Committee of Cancer Hospital, CAMS (No. 23/262-4004). The study was performed in full accordance with the guidelines for Good Clinical Practice and the Declaration of Helsinki.

Supplementary information