Graph convolutional network–derived pathomics and clinical integration for predicting chemoimmunotherapy response in advanced lung squamous cell carcinoma: a multicenter study

Abstract

Background

Chemoimmunotherapy (CIT) represents the standard first-line regimen for advanced lung squamous cell carcinoma (LUSC). However, treatment responses are heterogeneous, and conventional biomarkers are insufficient for precise patient stratification. This study aimed to develop and validate a deep learning-based model that integrated graph convolutional network (GCN)–derived pathomics signatures from whole-slide images (WSIs) with clinical features to predict the objective response rate (ORR) to CIT and achieve clinically meaningful risk stratification.

Methods

This retrospective study included 436 patients with pathologically confirmed LUSC and available WSIs from three centers, comprising 343 cases from the phase III AK105-302 trial and two external CIT cohorts (n = 63 and n = 30). Patch-level features were extracted using an Inception_v3-derived convolutional neural network (CNN) and aggregated at the patient level using a GCN to generate a pathomics signature. This signature was subsequently combined with clinical variables to construct the final predictive model. Model performance was assessed using the area under the curve (AUC), calibration, and decision curve analysis. Patients were stratified into high- and low-risk groups based on the Youden index derived from the combined model, and survival outcomes were evaluated using Kaplan-Meier and Cox regression analyses.

Results

The combined pathomics–clinical model demonstrated robust discrimination in predicting ORR across the training cohort (AUC 0.964), validation cohort (AUC 0.876), and two external cohorts (AUC 0.738 and 0.846, respectively), outperforming clinical-only, ensemble, and multiple-instance learning (MIL) models. In the risk stratification analysis, low-risk patients treated with CIT exhibited a significantly higher response rate than those treated with chemotherapy alone (95.6% vs. 53.7%; p < 0.001) and achieved superior progression-free survival (PFS; hazard ratio [HR] 0.35, 95% confidence interval [CI] 0.25–0.49; p < 0.001) and overall survival (OS; HR 0.35, 95% CI 0.22–0.55; p < 0.001). A significant treatment-by-risk interaction for PFS (interaction p = 0.019) and OS (interaction p < 0.001) indicated that the benefit of CIT was concentrated in the low-risk group. These findings were validated in both external CIT cohorts, while model performance and risk stratification were less effective in the chemotherapy-only and non-lung-origin cohorts.

Conclusions

The GCN-based pathomics–clinical model accurately predicted ORR to CIT and demonstrated robust, externally validated performance, suggesting its potential utility for response enrichment and risk stratification in advanced LUSC.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07962-0.

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide, with lung squamous cell carcinoma (LUSC) accounting for approximately 20–40% of all lung cancer cases [1]. Due to the relative scarcity of actionable driver mutations in this subtype, effective targeted therapies are limited [2, 3]. The KEYNOTE-407 trial established chemoimmunotherapy (CIT) as the first-line standard for advanced LUSC, offering the potential for durable clinical responses [4]. The objective response rate (ORR) serves as a critical early endpoint for assessing treatment efficacy. However, clinical responses are highly heterogeneous, with a substantial proportion of patients experiencing primary resistance or early progression [5, 6]. Accurate identification of patients most likely to achieve an objective response to CIT may optimize patient selection and risk stratification, potentially informing treatment planning and patient counseling.

Current predictive strategies, however, face significant challenges. Conventional biomarkers—including programmed death-ligand 1 (PD-L1) expression and tumor mutational burden (TMB)—exhibit suboptimal predictive value and lack consistent standardization for identifying patients who benefit from CIT, thereby limiting precise patient selection [7, 8]. For example, in KEYNOTE-407, improvements in ORR with CIT were observed across PD-L1 tumor proportion score (TPS) strata, underscoring the limited ability of PD-L1 alone to enrich for CIT responders [9]. These findings suggest that response determinants may extend beyond bulk molecular measures and involve the spatial architecture of the tumor microenvironment.

Accumulating evidence suggests that response to CIT depends not only on the presence of immune cells, but also on their spatial distribution and interactions with tumor nests [10, 11]. Biological phenomena such as “immune exclusion”—where dense stromal barriers physically restrict T cells from accessing the tumor core—and the proximity between lymphocytes and tumor cells are important contributors to therapeutic response [12]. Unlike bulk sequencing, hematoxylin and eosin (H&E) slides inherently preserve this spatial context. Digital pathology provides a means to decode these patterns [13–15]. Previous studies have begun to leverage this resource; for instance, we reported a tumor immune dysfunction and exclusion (TIDE)-based pathomics model that accurately predicted CIT-related survival outcomes in lung adenocarcinoma [16]. Similarly, Chen et al. developed a morphological signature based on pathological features for gastric cancer stratification [17]. Collectively, these studies suggest that routine histology contains signals associated with clinically relevant outcomes, motivating topology-aware modeling in whole-slide images (WSIs).

However, to fully exploit the spatial information contained in WSIs, more sophisticated modeling strategies may be required. Conventional ensemble methods and multiple-instance learning (MIL) often treat WSIs as unordered collections (“bags”) of patches, potentially neglecting the spatial topology underlying immune exclusion or infiltration [18–20]. In this context, graph convolutional networks (GCNs) offer a promising solution. By constructing graphs based on spatial adjacency—where patches are represented as nodes and spatial relationships as edges—GCNs explicitly model inter-patch relationships [21]. This approach may help capture higher-order histopathologic organization that is potentially relevant to immune–tumor interactions.

Beyond GCN-derived pathomics, patient clinical features remain critical determinants of treatment response and prognosis, including age, gender, smoking status, body mass index (BMI), Eastern Cooperative Oncology Group performance status (ECOG PS), and metastatic sites [22, 23]. In this study, we aimed to develop and validate a GCN-based pathomics model combined with clinical features to predict the objective response to CIT in advanced LUSC. Model performance was evaluated in terms of discrimination for objective response, calibration, and clinical utility, along with risk stratification to support clinically meaningful patient grouping. The model was developed and internally validated retrospectively using data from a prospective, multicenter trial cohort and externally validated in two independent cohorts. Furthermore, visualization tools were incorporated to enhance qualitative interpretability by illustrating histologic regions that most influenced the model outputs. These qualitative visualizations support clinician-facing review and improve the transparency of model behavior.

Methods

Study design and participants

Figure 1 summarizes the overall workflow. This multicenter retrospective study was conducted in accordance with the Declaration of Helsinki and received approval from the institutional review boards of all participating centers: Shanghai Chest Hospital (approval no. LS1907), Chinese PLA General Hospital (approval no. S2024-792-01), and Chongqing University Three Gorges Hospital (approval no. 2024200). The requirement for informed consent was waived due to the retrospective observational study design.

Fig. 1.

Workflow of the study design. Pretreatment H&E WSIs were retrospectively collected and divided into 512 × 512-pixel patches, followed by color normalization and feature extraction using ImageNet-pretrained CNNs (ResNet50, DenseNet121, and Inception_v3). Based on comparative analysis in the CIT_validation cohort, Inception_v3 was selected as the patch-level backbone. Patch-level predictions were further visualized using Grad-CAM and converted into prediction and probability maps to highlight response-related regions. For patient-level integration, patch embeddings were organized into spatial graphs using k-NN adjacency and processed by a GCN to generate a pathomics signature, which was combined with clinical variables to construct the final model. The final model was subsequently evaluated for predictive performance, risk stratification (high-risk vs. low-risk), and survival analyses. H&E, hematoxylin and eosin; WSI, whole-slide image; CNN, convolutional neural network; CIT, chemoimmunotherapy; Grad-CAM, gradient-weighted class activation mapping; k-NN, k-nearest neighbor; GCN, graph convolutional network

A total of 436 patients with pathologically confirmed LUSC and available formalin-fixed, paraffin-embedded (FFPE) histopathological WSIs from three independent centers were ultimately included. Of these, 343 were enrolled in the phase III AK105-302 randomized controlled trial (NCT03866993)—a prospective, multicenter study conducted across 74 hospitals in China that compared penpulimab (AK105; an anti–programmed cell death protein 1 [PD-1] antibody) plus chemotherapy with chemotherapy alone. Within AK105-302, specimens were categorized by tissue origin: 263 patients had lung-origin tumor tissue, and 80 had non-lung-origin biopsies. The lung-origin CIT subset (n = 128) was randomly partitioned (7:3) into a training set (CIT_training, n = 89) and an internal validation set (CIT_validation, n = 39) for model development and tuning. The remaining lung-origin chemotherapy cohort (Chemo, n = 135) was reserved to evaluate treatment-specific generalizability. The non-lung group within AK105-302 comprised 80 patients (CIT_non-lung, n = 42; Chemo_non-lung, n = 38) and was used to assess generalization across treatment regimens and tissue origins. For external validation, two independent CIT cohorts were assembled: the CIT_external 1 cohort (n = 63) from the Chinese PLA General Hospital and the CIT_external 2 cohort (n = 30) from Chongqing University Three Gorges Hospital. The patient selection process is detailed in Supplementary Information I and Fig. 2.

Fig. 2.

Flowchart of the patient selection process. LUSC, lung squamous cell carcinoma; CIT, chemoimmunotherapy; Chemo, chemotherapy

The prespecified primary endpoint was ORR, defined as the proportion of participants achieving a confirmed best overall response of complete response (CR) or partial response (PR) prior to the first imaging-confirmed progressive disease (PD). Secondary endpoints included progression-free survival (PFS) and overall survival (OS). PFS was defined as the duration from treatment initiation to disease progression or death from any cause, whichever occurred first. OS was defined as the time from treatment initiation to death from any cause. Tumor assessments and response categorization adhered to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1.

WSI acquisition and pre-processing

WSIs of H&E-stained tissue sections were acquired using high-resolution digital slide scanners with a 20× objective, yielding a resolution of 0.5 μm/pixel. Each WSI was divided into non-overlapping 512 × 512-pixel tiles. To ensure adequate tissue content for analysis, the OKT-patch2predict (OnekeyAI) algorithm was employed to identify and exclude non-informative regions (e.g., blank background and non-tissue areas). The remaining tiles were channel-wise normalized using the mean and standard deviation of the ImageNet dataset and subsequently center-cropped to 224 × 224 pixels (or to 299 × 299 pixels for Inception_v3) before being fed into the models.

Patch level prediction

This study employed a weakly supervised learning framework, utilizing ORR as the supervisory signal to train a patch-level classifier. Specifically, all patches derived from a single WSI were assigned the same label as the corresponding patient, thereby enabling patch-level training under weak supervision [24, 25].

For feature extraction, three canonical convolutional neural network (CNN) models pre-trained on the ImageNet dataset—ResNet50, DenseNet121, and Inception_v3—were utilized due to their robust representational capacity and computational efficiency [26]. The primary objective was to estimate the likelihood of each patch being classified into the response category corresponding to its WSI. For each patch, the model generated a prediction probability () and a derived binary label (). During patch-level model training, online data augmentation techniques, including random horizontal and vertical flips, were applied to enhance model robustness and mitigate the risk of overfitting. Detailed training specifications are provided in Supplementary Information II.

Visualization and qualitative interpretability

To facilitate the qualitative interpretation of patch-level predictions, gradient-weighted class activation mapping (Grad-CAM) was employed. Grad-CAM generated heatmaps that highlighted image regions most associated with the model’s predictions, thereby providing visual evidence of model attention [27]. For each WSI, two visualization maps were generated: (i) a prediction map that assigns each patch a categorical label (response vs. non-response); and (ii) a probability map that visualizes continuous patch-level prediction probabilities using a color gradient to reflect model confidence.

Patient level integration

To aggregate patch-level features into patient-level predictions, three integration strategies were systematically compared: ensemble, MIL, and GCN. Both ensemble and MIL implemented late-fusion schemes. Ensemble aggregated patch-level predicted probabilities using summary statistics, whereas MIL constructed a pathomics signature from histogram- and bag-of-words (BoW)-based features for patient-level classification. In contrast, GCN employed early-fusion: for each patch, features were extracted from the penultimate layer of the CNN models. These patch-level features were then integrated using the GCN, which applied multiple graph convolution layers to propagate and aggregate information across spatially adjacent patches, thereby producing the final patient-level prediction [28, 29]. Detailed descriptions of all fusion algorithms are provided in Supplementary Information III-V.

Model development and evaluation

Using the three patient-level integration strategies—ensemble, MIL, and GCN—three distinct pathomics signatures were derived. In parallel, a clinical signature was developed by applying machine-learning algorithms to baseline clinical variables, prioritizing those collected in as many cohorts as possible to improve cross-cohort comparability, including age, gender, smoking status, pathologic stage, ECOG PS, and PD-L1 expression. These parameters were pre-specified because they are routinely available before treatment and reflect patient condition, tumor burden, and established clinicopathologic determinants relevant to prognosis and CIT response in advanced NSCLC. To contextualize the clinical model, the relative contribution of individual clinical variables was additionally summarized using a random forest–based permutation importance approach (mean decrease in accuracy) [30, 31]. Subsequently, a combined model was constructed by integrating the best-performing pathomics signature with the same clinical variables.

Predictive performance for ORR was assessed using receiver operating characteristic (ROC) analysis, with the area under the curve (AUC) serving as the primary threshold-independent metric of discrimination. Calibration was evaluated using calibration curves to assess the agreement between predicted probabilities and observed outcomes, while decision-curve analysis (DCA) was employed to quantify clinical utility. Risk stratification was performed using cohort-specific Youden-optimal cutoffs derived from ROC analysis, which were used solely to dichotomize the continuous model score for within-cohort stratification analyses. Survival analyses were subsequently conducted to assess the model’s prognostic and predictive value. In the training cohort, five-fold cross-validation was used for model development and hyperparameter tuning to mitigate overfitting and assess stability. The trained model was then applied to the internal and external validation cohorts without further retraining or parameter tuning to evaluate generalizability.

Statistical analysis

Continuous data were summarized as mean (standard deviation [SD]) or median (interquartile range [IQR]), as appropriate. Between-group comparisons of continuous variables were conducted using Student’s t-test or the Wilcoxon rank-sum test, while categorical variables were compared using the χ² test or Fisher’s exact test. Model performance metrics for predicting ORR included the AUC with 95% confidence intervals (CIs), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Survival outcomes were estimated using the Kaplan-Meier method and compared using log-rank tests. Treatment-by-risk interaction effects were assessed using p-values from the interaction terms in the Cox analysis. Univariate and multivariable Cox regression analyses were performed to estimate hazard ratios (HRs) with 95% CIs.

All statistical tests were two-sided, with p < 0.05 considered statistically significant. Statistical analyses were performed using Python (version 3.7.12), R (version 3.6.1), IBM SPSS Statistics (version 23.0), and GraphPad Prism (version 9.5). For GCN, PyTorch Geometric (torch_geometric; version 2.1.0) was utilized. Machine learning algorithms, including the Support Vector Machine (SVM), were implemented using Scikit-learn (version 1.0.2). Deep-learning models were developed in PyTorch (version 1.11.0) and accelerated using CUDA (version 11.3.1) and cuDNN (version 8.2.1).

Results

Patient characteristics

The baseline characteristics of the 436 enrolled patients are shown in Table 1. The median age was 62.0 years (IQR 56.0–67.0 years). The majority of patients were male (87.4%), and 91.7% were former or current smokers. PD-L1 expression was predominantly intermediate (53.2% with 1 to < 50% expression), and high expression (≥ 50%) was observed in 18.1%. The most common baseline metastatic sites were bone (9.9%) and liver (6.2%), whereas brain metastases were uncommon (1.1%). The median baseline diameter of tumor lesions was 149.0 mm (IQR 99.0–204.0 mm). Baseline clinical characteristics varied across cohorts (Table S1), reflecting expected case-mix heterogeneity.

Table 1.

Baseline characteristics of patients

Characteristics	Total (n = 436)	CIT_training (n = 89)	CIT_validation (n = 39)	CIT_external 1 (n = 63)	CIT_external 2 (n = 30)	Chemo (n = 135)	CIT_non-lung (n = 42)	Chemo_non-lung (n = 38)
Age, years	62.0 [56.0, 67.0]	61.0 [55.0, 66.0]	59.0 [53.0, 65.0]	65.0 [60.0, 70.0]	59.0 [54.8, 67.0]	63.0 [57.0, 66.0]	62.0 [57.0, 68.3]	62.0 [56.0, 65.0]
Gender
Male	381 (87.4)	83 (93.3)	36 (92.3)	61 (96.8)	28 (93.3)	126 (93.3)	39 (92.9)	34 (89.5)
Female	55 (12.6)	6 (6.7)	3 (7.7)	2 (3.2)	2 (6.7)	9 (6.7)	3 (7.1)	4 (10.5)
Smoking status
Never	36 (8.3)	11 (12.4)	3 (7.7)	0 (0.0)	10 (33.3)	8 (5.9)	4 (9.5)	0 (0.0)
Former/Current	400 (91.7)	78 (87.6)	36 (92.3)	63 (100.0)	20 (66.7)	127 (94.1)	38 (90.5)	38 (100.0)
BMI, kg/m2	23.2 [20.6, 25.2]	23.2 [21.0, 24.7]	24.0 [21.3, 26.0]	24.2 [22.5, 26.3]	NA	22.2 [19.9, 24.7]	23.8 [19.9, 25.6]	22.5 [20.1, 25.4]
Pathologic stage
IIIB/C	94 (21.6)	11 (12.4)	7 (17.9)	29 (46.0)	17 (56.7)	21 (15.6)	5 (11.9)	4 (10.5)
IV	342 (78.4)	78 (87.6)	32 (82.1)	34 (54.0)	13 (43.3)	114 (84.4)	37 (88.1)	34 (89.5)
ECOG PS
0	90 (20.6)	17 (19.1)	10 (25.6)	0 (0.0)	6 (20.0)	34 (25.2)	15 (35.7)	8 (21.1)
1	344 (78.9)	72 (80.9)	29 (74.4)	63 (100.0)	22 (73.3)	101 (74.8)	27 (64.3)	30 (78.9)
2	2 (0.5)	0 (0.0)	0 (0.0)	0 (0.0)	2 (6.7)	0 (0.0)	0 (0.0)	0 (0.0)
PD-L1 expression
<1%	125 (28.7)	33 (37.1)	10 (25.6)	4 (6.3)	6 (20.0)	47 (34.8)	15 (35.7)	10 (26.3)
1%≤ PD-L1 < 50%	232 (53.2)	41 (46.1)	18 (46.2)	49 (77.8)	19 (63.3)	64 (47.4)	21 (50.0)	20 (52.6)
≥50%	79 (18.1)	15 (16.8)	11 (28.2)	10 (15.9)	5 (16.7)	24 (17.8)	6 (14.3)	8 (21.1)
Metastatic site
Brain	5 (1.1)	2 (2.2)	0 (0.0)	NA	NA	3 (2.2)	0 (0.0)	0 (0.0)
Liver	27 (6.2)	5 (5.6)	2 (5.1)	NA	NA	13 (9.6)	2 (4.8)	5 (13.2)
Bone	43 (9.9)	8 (9.0)	6 (15.4)	NA	NA	20 (14.8)	4 (9.5)	5 (13.2)
Diameter of lesions, mm	149.0 [99.0, 204.0]	157.0 [105.0, 206.0]	155.0 [114.0, 210.0]	NA	NA	157.0 [99.0, 204.0]	142.0 [93.5, 194.0]	143.0 [78.0, 199.5]

Data are presented as median [IQR] and n (%)

CIT, chemoimmunotherapy; Chemo, chemotherapy; BMI, body mass index; ECOG PS, Eastern Cooperative Oncology Group Performance Status; PD-L1, programmed death-ligand 1; IQR, interquartile range; NA, not available

Development and validation of the model

At the patch level, three CNNs—ResNet50, DenseNet121, and Inception_v3—were trained in the CIT_training cohort and evaluated in the CIT_validation cohort. In the CIT_training cohort (Fig. 3A), Inception_v3 demonstrated superior performance in predicting ORR (AUC 0.983, 95% CI 0.982–0.984), compared with ResNet50 and DenseNet121. This advantage persisted in the CIT_validation cohort (AUC 0.653, 95% CI 0.644–0.662; Fig. 3B) and was further validated externally (CIT_external 1: AUC 0.582, 95% CI 0.578–0.586; Fig. 3C; CIT_external 2: AUC 0.713, 95% CI 0.704–0.722; Fig. 3D). Table S2 presents additional performance metrics for Inception_v3, including accuracy, sensitivity, specificity, PPV, and NPV. Based on these findings, Inception_v3 was selected for subsequent analyses.

Fig. 3.

Performance comparison of patch-level classifiers for predicting ORR in the internal and external CIT cohorts. (A) ROC curve of three patch-level CNN classifiers—ResNet50, DenseNet121, and Inception_v3—in the CIT_training cohort. (B) ROC curve of the same three patch-level CNN classifiers in the CIT_validation cohort. (C) ROC curve of the Inception_v3 patch-level classifier in the CIT_external 1 cohort. (D) ROC curve of the Inception_v3 patch-level classifier in the CIT_external 2 cohort. ORR, objective response rate; CIT, chemoimmunotherapy; ROC, receiver operating characteristic; CNN, convolutional neural network; AUC, area under the curve

To further illustrate qualitative interpretability, we presented two representative cases with Grad-CAM alongside prediction and probability maps (Fig. 4). The highlighted regions corresponded to histologically plausible tumor areas, providing qualitative support for the model’s attention in examples of both non-response and response.

Fig. 4.

Patch-level visualization of model prediction. Pretreatment H&E WSIs were divided into 512 × 512-pixel patches and processed using CNNs (ResNet50, DenseNet121, and Inception_v3) to extract patch-level features. Following comparative evaluation, Inception_v3 was selected as the patch-level backbone. To enhance interpretability, Grad-CAM was applied to highlight regions most influential for predictions. For each case, a prediction map (response vs. non-response) and a probability map (0–1) were generated. Representative examples of a non-response WSI (left) and a response WSI (right) are shown. H&E, hematoxylin and eosin; WSI, whole-slide image; CNN, convolutional neural network; Grad-CAM, gradient-weighted class activation mapping

At the patient level, patch features derived from Inception_v3 were aggregated to construct three pathomics models (ensemble, MIL, and GCN). Details of the ensemble and MIL strategies are provided in Supplementary Information III-IV and Fig. S1. Concurrently, a clinical-only comparator was built using baseline clinical variables, with the relative importance of these variabled visualized in Fig. S2. Permutation importance analysis indicated that ECOG PS contributed most strongly, followed by age and BMI, while the remaining variables demonstrated relatively limited contributions. Following development in the CIT_training cohort (Fig. 5A), the GCN showed superior predictive performance for ORR compared with the clinical, ensemble, and MIL strategies, achieving AUCs of 0.869 (95% CI 0.738-1.000) in the CIT_validation cohort, 0.743 (95% CI 0.596–0.890) in the CIT_external 1 cohort, and 0.837 (95% CI 0.620-1.000) in the CIT_external 2 cohort (Fig. 5B-D). The sensitivity of the GCN was notably high in the CIT_validation cohort (0.900) and both CIT_external cohorts (0.818 and 1.000, respectively), whereas specificity was moderate (0.577–0.828). The NPV for the GCN exceeded 0.900 across all four cohorts, while the PPV ranged from 0.267 to 0.815 (Fig. 5E-H and Table S3).

Fig. 5.

Performance comparison of patient-level models for predicting ORR in the internal and external CIT cohorts. (A-D) ROC curves of the clinical, ensemble, MIL, GCN, and combined models in the CIT_training, CIT_validation, CIT_external 1, and CIT_external 2 cohorts. (E-H) Radar plots summarizing AUC, ACC, SEN, SPE, PPV, NPV, and the Youden index for the same five models in the four cohorts. (I-L) Calibration curves showing the relationship between predicted and observed probabilities for the same five models in the four cohorts. (M-P) DCA comparing the net clinical benefit of the five models across a range of risk thresholds in the four cohorts. ORR, objective response rate; CIT, chemoimmunotherapy; ROC, receiver operating characteristic; MIL, multiple-instance learning; GCN, graph convolutional network; AUC, area under the curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; DCA, decision curve analysis

Building on these comparisons, a combined model was constructed by integrating the best-performing pathomics signature—the GCN-derived pathomics signature—with clinical features. This design aimed to determine whether augmenting the GCN-derived pathomics signature with clinical factors provided complementary predictive value for ORR and supported meaningful patient stratification. Across the CIT_training, CIT_validation, and both CIT_external cohorts, the combined model consistently achieved the highest or near-highest discriminative performance and exhibited greater stability than the other four models, as reflected by higher AUCs (0.738–0.964; Fig. 5A-D). Additional performance metrics are summarized in radar plots (Fig. 5E-H and Table S3). For the combined model, calibration curves demonstrated good agreement between predicted and observed probabilities (Fig. 5I-L), and DCA indicated greater net benefit across clinically relevant thresholds (Fig. 5M-P).

Generalization across therapeutic regimens and tissue origins

An exploratory analysis was conducted to assess the model’s generalizability across different therapeutic regimens and tissue origins. In the Chemo, CIT_non-lung, and Chemo_non-lung cohorts, the patch-level Inception_v3 classifier showed limited transferability, with AUCs of 0.449, 0.546, and 0.347, respectively (Fig. S3).

At the patient level, neither the GCN pathomics model nor the combined model showed a distinct or consistent advantage over the clinical, ensemble, or MIL models. The ROC curves largely overlapped, yielding uniformly low AUCs (0.576–0.678; Fig. S4A-C). Radar plots showed no consistent improvements in accuracy, sensitivity, specificity, PPV, NPV, or the Youden index (Fig. S4D-F and Table S3). Calibration curves deviated from the ideal line (Fig. S4G-I), and DCA indicated minimal net benefit across clinically relevant thresholds (Fig. S4J-L).

Risk stratification and survival analysis based on the combined model

Patients were stratified into high- and low-risk groups using cohort-specific Youden-optimal cutoffs derived from the combined model. The CIT_training and CIT_validation cohorts were pooled to form an internal CIT cohort. Baseline characteristics of patients in the internal CIT and chemotherapy cohorts are presented in Table S4, with risk group–stratified characteristics provided in Table S5. Among low-risk patients, CIT was associated with a significantly higher response rate than chemotherapy (95.6% vs. 53.7%; p < 0.001; Fig. 6A). Conversely, among high-risk patients, the ORR was significantly lower with CIT than with chemotherapy (18.4% vs. 54.3%; p = 0.004; Fig. 6A). Compared with the high-risk group, the waterfall and spider plots demonstrated greater tumor regression and more favorable longitudinal tumor-burden dynamics in the low-risk CIT group. Subsequently, the association between risk stratification and long-term outcomes was evaluated. Among low-risk patients, CIT was associated with significantly longer PFS than chemotherapy (12 months: 35.6% vs. 7.3%; 24 months: 8.9% vs. 1.0%; p < 0.001; Fig. 6D-E). Moreover, Kaplan-Meier curves demonstrated a significant treatment-by-risk interaction for PFS, with greater benefit observed in the low-risk CIT group (interaction p = 0.019; Fig. 6F). Similarly, OS favored low-risk patients receiving CIT compared with chemotherapy (12 months: 92.2% vs. 64.5%; 24 months: 37.8% vs. 15.5%; p < 0.001; Fig. 6G-H). Kaplan-Meier curves further indicated that the OS benefit of CIT was concentrated in low-risk patients (interaction p < 0.001; Fig. 6I). However, the 12- and 24-month PFS and OS rates did not differ significantly between CIT and chemotherapy in the high-risk group. In univariate and multivariable Cox models within the internal CIT cohort, risk stratification derived from the combined model remained a key independent prognostic factor for both PFS (Table S6) and OS (Table S7).

Fig. 6.

Risk stratification and survival analysis based on the combined model in the internal CIT and Chemo cohorts. (A) Donut plots of ORR by treatment and risk stratification. (B) Waterfall plot of the best percentage change from baseline by risk stratification in the internal CIT cohort. (C) Longitudinal tumor-burden curves by risk stratification in the internal CIT cohort. (D, E) Distribution of PFS time (< 12 months, 12 to < 24 months, ≥ 24 months) by treatment and risk stratification. (F) Kaplan-Meier curves for PFS by treatment and risk stratification. (G, H) Distribution of OS time (< 12 months, 12 to < 24 months, ≥ 24 months) by treatment and risk stratification. (I) Kaplan-Meier curves for OS by treatment and risk stratification. CIT, chemoimmunotherapy; Chemo, chemotherapy; ORR, objective response rate; PFS, progression-free survival; OS, overall survival

These findings were reproduced in two external CIT cohorts. Baseline characteristics for high- versus low-risk groups in the two external CIT cohorts are provided in Table S8 and Table S9, respectively. In both cohorts, the low-risk group consistently exhibited a higher response rate than the high-risk group (CIT_external 1: 96.4% vs. 71.4%; p = 0.024; CIT_external 2: 100.0% vs. 69.2%; p = 0.026; Fig. 7A and D) and longer PFS (CIT_external 1: HR 0.48, 95% CI 0.23–0.99; p = 0.043; CIT_external 2: HR 0.31, 95% CI 0.11–0.85; p = 0.015; Fig. 7B and E), with a trend toward OS benefits (CIT_external 1: HR 0.42, 95% CI 0.19–0.93; p = 0.045; CIT_external 2: HR 0.26, 95% CI 0.05–1.36; p = 0.088; Fig. 7C and F).

Fig. 7.

Risk stratification and survival analysis based on the combined model in both CIT_external cohorts. (A) Bar plot of ORR by risk stratification in the CIT_external 1 cohort. (B, C) Kaplan-Meier curves for PFS and OS by risk stratification in the CIT_external 1 cohort. (D) Bar plot of ORR by risk stratification in the CIT_external 2 cohort. (E, F) Kaplan-Meier curves for PFS and OS by risk stratification in the CIT_external 2 cohort. CIT, chemoimmunotherapy; ORR, objective response rate; PFS, progression-free survival; OS, overall survival; HR, hazard ratio; mPFS, median progression-free survival; mOS, median overall survival; NR, not reached

However, risk stratification did not demonstrate a clear advantage in terms of response rate (Figs. S5A and S5D), PFS (Figs. S5B and S5E), or OS (Figs. S5C and S5F) in exploratory non-lung-origin cohorts.

Discussion

In this multicenter retrospective study, we developed and validated a deep learning-based model that integrated a GCN-derived pathomics signature with clinical features to predict objective response to CIT in patients with advanced LUSC. The GCN-based pathomics model outperformed conventional ensemble and MIL approaches, as well as the clinical-only model. Baseline clinical variables were subsequently incorporated to construct the final combined model used for risk stratification and survival analyses. Using this framework, low-risk patients had a higher probability of response and more favorable survival outcomes with CIT, whereas model performance and risk stratification were less consistent in the chemotherapy-only cohort and in the non-lung-origin tissue cohorts.

Conventional pathomics approaches often emphasize cellular morphology and the local composition of tumor microenvironmental, providing detailed microstructural insights. However, these approaches may underutilize information related to the global spatial organization of tumor tissue. With the widespread adoption of high-resolution whole-slide imaging and deep learning, digital pathology has increasingly shifted toward the systematic exploration of spatial tissue architecture [32, 33]. Ahmedt-Aristizabal et al. provided a comprehensive overview of the application of graph-based structures in computational pathology, highlighting their utility for diagnostic support and outcome prediction across multiple cancer types, including lung, colorectal, and breast cancers [34]. Additionally, Taki et al. quantified tumor–stroma spatial adjacency to develop the Spatial Tumor-Stroma Distribution Index (STSDI), which independently stratified prognosis in LUSC beyond conventional clinicopathologic variables [35].

Methodologically, these observations support the central premise of our study. The GCN approach may offer an advantage over conventional pathomics models by integrating patch-level morphologic signals with spatial context encoded by graphs, enabling multiscale modeling of tumor architecture. At the micro-level, the model may leverage H&E-visible cues reflecting tumor–stroma–immune composition and architecture. For example, patches characterized by dense stromal bands around tumor nests, a paucity of peritumoral lymphocytes, or prominent lymphocyte aggregates at the tumor–stroma interface were frequently highlighted in qualitative visualizations, suggesting that such patterns may carry discriminative information for response-related phenotypes [36, 37]. At the macro-level, the GCN may further capture how these morphologic patterns are organized across the WSI, such as neighborhood context, spatial clustering, and regional heterogeneity, which are not captured when patches are treated as independent instances [38]. Complementary to the histologic architecture captured by the GCN-based pathomics signature, baseline clinical factors provide additional context on patient fitness and disease burden. Accordingly, we incorporated clinical features—such as ECOG performance status and BMI—to complement the pathomics signature [39–41]. This integrative strategy reflects the multifactorial nature of treatment response and illustrates that supplementing GCN-based pathomics with clinical factors can yield more clinically relevant predictions than pathomics alone. Based on this combined model, we illustrate a potential stratification framework to optimize patient selection. In our retrospective cohorts, patients classified as low-risk (higher predicted response) exhibited greater benefit from CIT, supported by treatment-by-risk interaction analyses.

A common concern with GCN-based deep learning models is their inherently “black-box” nature, which can hinder clinical adoption [42]. To improve transparency, we used Grad-CAM together with prediction and probability maps to visualize image regions that most influenced the model outputs. These heatmaps provided a qualitative indication of where the model attended within the WSI and were intended to support face validity and clinician-facing review, rather than to provide a mechanistic explanation.

Beyond predictive accuracy, our model also achieved clinically meaningful risk stratification. Importantly, a significant treatment-by-risk interaction was observed, indicating that the model not only identifies responders but also distinguishes patients who derive greater benefit from CIT compared with chemotherapy alone. Patients classified as low-risk by the combined model exhibited significantly higher ORR, prolonged PFS, and improved OS compared with high-risk patients receiving CIT. These associations were reproduced in two independent external cohorts, further supporting the robustness and generalizability of the model in lung-origin, CIT-treated populations.

The reduced performance in the chemotherapy-only and non-lung-origin cohorts suggests that the predictive signals captured by spatial pathomics may be context-dependent. CIT response is more closely associated with the tumor microenvironment and immune–tumor–stroma interactions, some of which may be reflected in H&E-visible spatial patterns [43]. In contrast, chemotherapy response may depend more on tumor-intrinsic determinants (e.g., proliferation, necrosis, and cytotoxic sensitivity) that are not consistently represented by immune-associated spatial organization in routine histology [44]. Moreover, non-lung-origin biopsies introduce microenvironmental and sampling-site heterogeneity, which can alter histologic appearance and spatial architecture, further challenging transportability. Collectively, these findings clarify the primary intended application of our framework—CIT-treated, lung-origin specimens—and motivate future work to improve robustness across tissue origins and treatment contexts.

Despite the promising findings, several limitations should be acknowledged. First, although multicenter, prospective data were incorporated and external validation was performed, model development and evaluation were retrospective, which may introduce selection bias and limit causal inference. Second, biological interpretability remains constrained: immune-infiltrated/excluded architectures, stromal exclusion patterns, and tumor–immune proximity were not directly quantified, and our visualizations provide qualitative support rather than mechanistic evidence. Future studies integrating pathologist-guided annotations, explicit spatial/statistical metrics, and orthogonal assays (e.g., immunophenotyping or spatial profiling) are warranted. Third, this study primarily supports use in CIT-treated patients with lung-origin biopsies. Generalizability to non–lung-origin biopsies, chemotherapy-only cohorts, and populations outside China requires further confirmation. Fourth, as risk stratification relied on cohort-specific Youden-optimal thresholds, threshold-dependent metrics and dichotomized subgroup analyses should be interpreted with caution. Importantly, this framework should be considered hypothesis-generating at this stage, and prospective validation under standardized workflows is essential before clinical implementation and independent clinical decision-making.

Conclusion

In conclusion, we developed a novel deep learning-based model that integrates GCN-derived pathomics signatures with clinical features to predict objective response to CIT in advanced LUSC. The robust performance of this combined model across internal and external cohorts suggests its potential utility for response enrichment and risk stratification in CIT-treated patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CIT

Chemoimmunotherapy

LUSC

Lung squamous cell carcinoma

GCN

Graph convolutional network

WSI

Whole-slide image

ORR

Objective response rate

CNN

Convolutional neural network

AUC

Area under the curve

MIL

Multiple-instance learning

PFS

Progression-free survival

OS

Overall survival

PD-L1

Programmed death-ligand 1

TMB

Tumor mutational burden

TPS

Tumor proportion score

H&E

Hematoxylin and eosin

TIDE

Tumor immune dysfunction and exclusion

BMI

Body mass index

ECOG PS

Eastern Cooperative Oncology Group performance status

FFPE

Formalin-fixed: paraffin-embedded

Chemo

Chemotherapy

CR

Complete response

PR

Partial response

PD

Progressive disease

RECIST v1.1

Response Evaluation Criteria in Solid Tumors version 1.1

Grad-CAM

Gradient-weighted class activation mapping

BoW

Bag-of-words

ROC

Receiver operating characteristic

DCA

Decision-curve analysis

SD

Standard deviation

IQR

Interquartile range

CI

Confidence interval

PPV

Positive predictive value

NPV

Negative predictive value

HR

Hazard ratio

SVM

Support vector machine

STSDI

Spatial Tumor-Stroma Distribution Index

k-NN

k-nearest neighbor

ACC

Accuracy

SEN

Sensitivity: SPE: specificity

mPFS

median progression-free survival

mOS

median overall survival

NR

Not reached

NA

Not available

Author contributions

DYW, SM, and MHZ were involved in the literature search, figures, study design, data collection, data analysis, and writing. GYT, SYW, YCS and GXX were involved in literature search and data collection. XYZ, LC, BHH, HZ, and WN were involved in study design and article review. All authors read and approved the final manuscript.

Funding

This study was supported by National Natural Science Foundation of China (No. 82373425 and No. 82372722), the Medical Innovation Research Special Project of the Science and Technology Commission of Shanghai Municipality (No. 23Y11904200), Shanghai Innovative Medical Device Application Demonstration Project 2023 (No. 23SHS02600), Shanghai “Rising Stars of Medical Talents” Youth Development Program (Youth Medical Talents-Specialist Program, No. SHWSRS(2024)_070), and the Shanghai Key Laboratory Open Project (STCSM No. 22DZ2229005).

Data availability

Any reasonable requests for access to available data underlying the results reported in this article will be considered. Such proposals should be submitted to the corresponding author.

Declarations

Ethics approval and consent to participate

This multicenter retrospective study was conducted in accordance with the Declaration of Helsinki and received approval from the institutional review boards of all participating centers: Shanghai Chest Hospital (approval no. LS1907), Chinese PLA General Hospital (approval no. S2024-792-01), and Chongqing University Three Gorges Hospital (approval no. 2024200). The requirement for informed consent was waived because of the observational design.

Consent for publication

Not applicable.

Role of funder/sponsor

The funding sources had no role in the design or conduct of the study, including the collection, management, analysis, or interpretation of the data; the preparation, review, or approval of the manuscript; or the decision to submit the manuscript for publication.

Competing interests

All authors have declared no conflicts of interest.