A machine learning-based framework for prognostic prediction and tumor microenvironment characterization of locally advanced cervical cancer with concurrent chemoradiotherapy Authors: Yue Feng, Zijian Sun, Yuqiang Li, Fang Wang, Qiyang Li, Jiahui Ma, Xiaoyong Zhang, Hui Ye, Xiaojuan Lv, Zhao Wang, Lei Shi, Zhen Zhang, Jiayu Song, Tao Feng, Haowen Li, Chunwei Xu, Yihan Wang, Jingkui Tian, Yunyan Zhang, Liting Shi, Hanmei Lou, Wei Zhu Published: 2025 ABSTRACT Accurate prognosis prediction for locally advanced cervical cancer (LACC) after concurrent chemoradiotherapy (CCRT) is essential for individualized treatment decision making. We aimed to develop a multitask prognostic model and reveal radiomic-phenotypic associations for LACC patients after CCRT. The framework consists of (1) A deep learning–based fully automated model (DeepMR-LACC) which used T2-weighted magnetic resonance images obtained before CCRT to predict patient outcomes; (2) Proteomics profiling from paired cervical biopsy samples for tumor microenvironment characterization and radioproteomics-based risk stratification. The DeepMR-LACC predicted progression-free survival (PFS) and overall survival (OS) in training [C-indices, 0.80 (95% confidence interval, 0.75–0.84) and 0.83 (0.80–0.87)], internal test [0.67 (0.59–0.75) and 0.70 (0.61–0.78)], external test [0.69 (0.59–0.78) and 0.65 (0.55–0.76)] cohorts. The DeepMR-LACC effectively stratified patients into high- or low-risk groups, outperforming current clinical risk factors. Furthermore, proteomic profiling revealed an immunosuppressive microenvironment in the high-risk group. Finally, radioproteomics-based risk stratification showed superior prognostic performance compared to the DeepMR-LACC for PFS and OS in the radioproteomics cohort [C-indices 0.85 (0.74–0.96) and 0.85 (0.73–0.96)]. The DeepMR-LACC enabled accurate prognostic prediction and in-depth tumor microenvironment characterization for LACC, aiding personalized long-term management. Introduction Cervical cancer is a substantial global health challenge; although there are scale-up of prevention programs, approximately 604,127 new cases and nearly 341,831 related deaths of the disease were reported in 20201. Particularly in countries with low or medium human development indices, most cases are initially diagnosed as locally advanced cervical cancer (LACC) and have a worse prognosis2. Cisplatin-based concurrent chemoradiotherapy (CCRT) is the definitive treatment for such cases3. However, nearly 30–50% of LACC patients experienced progression or recurrence after CCRT, with a 5-year survival rate of less than 50%4. Thus, a prognostic model is needed to identify LACC cases with high risks of recurrence and mortality after CCRT. Tumor microenvironment heterogeneity is a key factor contributing to the failure of CCRT5. In contemporary clinical practice, the International Federation of Gynecology and Obstetrics (FIGO) staging system is employed for prognostic assessment. Still, this system is limited due to the lack of consideration of prognostic heterogeneity among cases of the same stage6. The tumor microenvironment (TME), a major determinant of tumor heterogeneity, consists of extracellular matrix (ECM), subpopulations of immune cells, and biochemical factors7. TME components play crucial roles in treatment responses and clinical outcomes across human cancers. Recent studies have demonstrated that the immune cell composition in cervical tumors affects the efficacy of chemoradiotherapy and that immunotherapy combined with chemoradiotherapy can improve cervical cancer outcomes for part of LACC patients8,9. Thus, the assessment of different prognostic phenotypes in patients with LACC and its association with TME is imperative. Radiomics facilitates the extraction of medical imaging features that correlate with imaging phenotypes and disease prognoses10. Deep learning (DL) techniques further enhance the ability to predict patient outcomes via the identification of imaging patterns that extend beyond traditional radiomic features11. This approach has demonstrated significant utility in the assessment of tumor heterogeneity, as it provides a mesoscopic view of the TME12. Concurrently, proteomics serves as a powerful tool to reflect the intricacies of bodily biology and pathophysiology, thereby enabling the identification of tumor microenvironments (TMEs) that are intricately linked to clinical outcomes13,14. Several groups including ours have developed biology signatures based on proteomics that were correlated with disease diagnosis and outcomes15–17. An integrated radioproteomics approach was recently used to explore different treatment response phenotypes18. However, the radioproteomics illustration of imaging-based tumor heterogeneity and its biological foundation remains unclear. To address these limitations, we developed and validated a DL-based fully automated model—termed DeepMR-LACC—for LACC prognosis prediction based on pre-treatment T2-weighted magnetic resonance (MR) images. Moreover, we established a radioproteomics-based risk stratification strategy by integrating proteomics data from paired tumor tissues with the DeepMR-LACC model. This integrated approach may contribute to the improvement of predictive accuracy and identification of distinct molecular pathways and immune cell infiltration patterns associated with different prognoses, offering a biological framework for model-based prediction and informing targeted clinical interventions. Results Clinical characteristics This study included 693 patients from two independent cancer centers: training cohort, internal test cohort, external test cohort, and radioproteomics cohort (Supplementary Fig. S1). The clinical characteristics of patients in the four cohorts are summarized in Table 1. Among these patients, a total of 173 (25.0%) patients experienced recurrence, and 186 (26.8%) patients died within the follow-up time.Table 1Patient characteristics in the training, internal test, external test and radioproteomics cohortsTraining cohort (n = 386)Internal test cohort (n = 168)External test cohort (n = 93)Radioproteomics cohort (n = 46)Age (%) <59177 (45.9)79 (47.0)51 (54.8)18 (39.1) ≥59209 (54.1)89 (53.0)42 (45.2)28 (60.9)Stage (%) Ⅰ/Ⅱ200 (51.8)85 (50.6)43 (46.2)23 (50.0) Ⅲ/Ⅳ186 (48.2)83 (49.4)50 (53.8)23 (50.0)Pre-SCC (ng/mL) <7.6192 (49.7)80 (47.6)66 (71.0)26 (56.5) ≥7.6193 (50.0)86 (51.2)27 (29.0)20 (43.5) NA1 (0.3)2 (1.2)00Tumor Size (cm) <4.0129 (33.4)63 (37.5)41 (44.1)16 (34.8) ≥4.0257 (66.6)105 (62.5)52 (55.9)30 (65.2) Duration of CCRT (day)57.89 (8.14)57.09 (8.72)66.11 (17.35)54.22 (9.38) EBRT (Gy)46.22 (1.12)46.07 (1.02)43.54 (4.01)46.05 (0.97) ICBT (Gy)28.06 (3.51)28.39 (3.33)34.46 (7.07)25.58 (4.82)Synchronous chemotherapy (%) No129 (33.4)59 (35.1)2 (2.2)11 (23.9) Yes257 (66.6)109 (64.9)91 (97.8)35 (76.1)Recurrence (%) No290 (75.1)126 (75.0)73 (78.5)31 (67.4) Yes96 (24.9)42 (25.0)20 (21.5)15 (32.6)Outcome (%) Survival286 (74.1)124 (73.8)68 (73.1)29 (63.0) Death100 (25.9)44 (26.2)25 (26.9)17 (37.0) Table 1Patient characteristics in the training, internal test, external test and radioproteomics cohortsTraining cohort (n = 386)Internal test cohort (n = 168)External test cohort (n = 93)Radioproteomics cohort (n = 46)Age (%) <59177 (45.9)79 (47.0)51 (54.8)18 (39.1) ≥59209 (54.1)89 (53.0)42 (45.2)28 (60.9)Stage (%) Ⅰ/Ⅱ200 (51.8)85 (50.6)43 (46.2)23 (50.0) Ⅲ/Ⅳ186 (48.2)83 (49.4)50 (53.8)23 (50.0)Pre-SCC (ng/mL) <7.6192 (49.7)80 (47.6)66 (71.0)26 (56.5) ≥7.6193 (50.0)86 (51.2)27 (29.0)20 (43.5) NA1 (0.3)2 (1.2)00Tumor Size (cm) <4.0129 (33.4)63 (37.5)41 (44.1)16 (34.8) ≥4.0257 (66.6)105 (62.5)52 (55.9)30 (65.2) Duration of CCRT (day)57.89 (8.14)57.09 (8.72)66.11 (17.35)54.22 (9.38) EBRT (Gy)46.22 (1.12)46.07 (1.02)43.54 (4.01)46.05 (0.97) ICBT (Gy)28.06 (3.51)28.39 (3.33)34.46 (7.07)25.58 (4.82)Synchronous chemotherapy (%) No129 (33.4)59 (35.1)2 (2.2)11 (23.9) Yes257 (66.6)109 (64.9)91 (97.8)35 (76.1)Recurrence (%) No290 (75.1)126 (75.0)73 (78.5)31 (67.4) Yes96 (24.9)42 (25.0)20 (21.5)15 (32.6)Outcome (%) Survival286 (74.1)124 (73.8)68 (73.1)29 (63.0) Death100 (25.9)44 (26.2)25 (26.9)17 (37.0) Patient characteristics in the training, internal test, external test and radioproteomics cohorts Prognostic value of DeepMR-LACC Our prognostic model achieved high accuracy in predicting outcomes of CCRT for LACC. The C-indices for PFS and OS were 0.80 [95% confidence interval (CI), 0.75–0.84] and 0.83 (95% CI, 0.80–0.87), respectively, for the training cohort and 0.67 (95% CI, 0.59–0.75) and 0.70 (95% CI, 0.61–0.78), respectively, for the internal testing cohort. In the training cohort, the AUCs (area under the ROC curve) for 1-, 3-, and 5-year PFS were 0.81 (95% CI, 0.74–0.87), 0.82 (95% CI, 0.77–0.87), and 0.85 (95% CI, 0.80–0.89), respectively (Supplementary Fig. S2A). The AUCs for 1-, 3-, and 5-year OS were 0.79 (95% CI, 0.68–0.87), 0.86 (95% CI, 0.81–0.91), and 0.89 (95% CI, 0.85–0.93), respectively. In the internal test cohort, the AUCs for 1-, 3-, and 5-year PFS were 0.67 (95% CI, 0.54–0.78), 0.73 (95% CI, 0.63–0.82), and 0.68 (95% CI, 0.58–0.77), respectively (Supplementary Fig. S2B). The AUCs for 1-, 3-, and 5-year OS were 0.56 (95% CI, 0.23–0.90), 0.71 (95% CI, 0.61–0.82), and 0.73 (95% CI, 0.63–0.82), respectively. Based on the training cohort results, the threshold DeepMR-LACC score of 1.921 was used to classify the patients into high- or low-risk groups in all the cohorts. The clinical characteristics of these groups are presented in Supplementary Table S5. In the training cohort, compared with those in the low-risk group, patients in the high-risk group had significantly lower PFS [HR, 5.70 (95% CI, 3.41–9.53); p < 0.0001] and OS [HR, 13.31 (95% CI, 6.70–26.42); p < 0.0001; Fig. 1A]. In the internal test cohort, patients in the high-risk group also had lower PFS [HR, 3.41 (95% CI, 1.58–7.37); p = 0.0018] and OS [HR, 3.71 (95% CI, 1.72–8.01); p = 0.0008] than those in the low-risk group (Fig. 1B).Fig. 1Kaplan-Meier survival curves in the training and test cohorts.A training cohort. B internal test cohort. C external test cohort. Patients were categorized into high-risk and low-risk groups. The p value was calculated using the log-rank test, and HR was derived from univariate Cox regression analysis. Kaplan-Meier survival curves in the training and test cohorts. A training cohort. B internal test cohort. C external test cohort. Patients were categorized into high-risk and low-risk groups. The p value was calculated using the log-rank test, and HR was derived from univariate Cox regression analysis. We assessed the generalization capability of DeepMR-LACC using data from the independent external test cohort. In this cohort, the C-indices for PFS and OS were 0.69 (95% CI, 0.59–0.78) and 0.65 (95% CI, 0.55–0.76), respectively. AUCs for 1-, 3-, and 5-year PFS and OS were 0.64 (95% CI, 0.43–0.85), 0.66 (95% CI, 0.51–0.80), 0.68 (95% CI, 0.55–0.81), 0.34 (95% CI, 0.13–0.57), 0.63 (95% CI, 0.46–0.78), and 0.65 (95% CI, 0.51–0.79), respectively, indicating robust predictive accuracy (Supplementary Fig. S2C). Patients in the high-risk group had significantly lower PFS [HR, 3.50 (95% CI, 1.27–9.64); p = 0.0154] and OS [HR, 2.44 (95% CI, 1.05–5.67); p = 0.0373], confirming the model’s effective risk stratification capability (Fig. 1C). DeepMR-LACC score is independent of clinical features The DeepMR-LACC model demonstrated superior clinical utility compared to clinical predictors. Decision curve analysis showed superior net benefit among clinical predictors, supporting its potential for clinical application (Supplementary Fig. S3). The model maintained excellent calibration for both OS (Brier Score=0.17) and PFS (Brier Score=0.18) in the training cohort, with predicted probabilities closely matching observed outcomes (Supplementary Fig. S4). Subgroup analyses indicated that the DeepMR-LACC scores were strong predictors in the training and internal testing cohorts (Supplementary Fig. S5-8). In most clinical subgroups, the scores effectively distinguished between high- and low-risk groups, with the high-risk group showing lower PFS and OS relative to the low-risk group. In the OS subcohorts of the internal test cohort stratified by pre-SCC ≥ 7.6 ng/mL and disease stage of III/IV, the high- and low-risk groups were significantly distinguished (log-rank p = 0.0027 and 0.0016, respectively), underscoring the substantially better prognoses in the latter group. The DeepMR-LACC score emerged as an independent predictor in the multivariate Cox regression analysis (Fig. 2 and Supplementary Table S6). The HRs for PFS in the training, internal test, and external test cohorts were 5.63 (95% CI, 3.34–9.49; p < 0.0001), 3.10 (95% CI, 1.41–6.80; p = 0.0048), and 3.47 (95% CI, 1.23–9.79; p = 0.0188), respectively. For OS, the HRs were 12.94 (95% CI, 6.48–25.87; p < 0.0001), 3.19 (95% CI, 1.44–7.06; p = 0.0042), and 2.32 (95% CI, 0.96–5.59; p = 0.0604), respectively. Across all three cohorts, the HRs for PFS and OS associated with the DeepMR-LACC score were higher than those associated with other variables. These findings were reflected in the C-indices for these variables (Supplementary Table S7), underscoring the robust predictive ability of the DeepMR-LACC score, independent of other clinical variables.Fig. 2Forest plots of multivariate Cox regression analysis.A training cohort. B internal test cohort. C external test cohort. PFS progression-free survival, OS overall survival, Pre-SCC pre-treatment squamous cell carcinoma antigen level. Forest plots of multivariate Cox regression analysis. A training cohort. B internal test cohort. C external test cohort. PFS progression-free survival, OS overall survival, Pre-SCC pre-treatment squamous cell carcinoma antigen level. The distribution of DeepMR-LACC scores with the attention distribution overlaid on the original images in the internal test cohort are shown in Fig. 3. This visualization clearly demonstrated that the model’s attention was not limited to the tumor ROIs, but also encompassed the peritumoral tissues. Additionally, according to the statistical results (Supplementary Table S8), 86.9% attention heatmap were classified into peritumoral attention in the internal test cohort, with 84/100 patients in the high-risk group and 62/68 patients in the low-risk group. The results validated our findings that the DeepMR-LACC paid more attention on the peritumoral regions for prognostic prediction, and highlighted that peritumoral microenvironment played crucial role in cancer progression and therapy response19,20. Furthermore, an independent intermediate radiology expert (Q.L.) was invited to assess the attention maps, which revealed that the peritumoral areas prioritized by DeepMR-LACC encompassed regions with fibrous hyperplasia.Fig. 3The DeepMR-LACC score distribution and attention weight visualization of patients in the internal test cohort.The left side presents a bar chart of the DeepMR-LACC score distribution among patients in the internal test cohort. On the right side, a heat map visualizing the attention weights of patient images obtained through attention rollout is displayed. The DeepMR-LACC score distribution and attention weight visualization of patients in the internal test cohort. The left side presents a bar chart of the DeepMR-LACC score distribution among patients in the internal test cohort. On the right side, a heat map visualizing the attention weights of patient images obtained through attention rollout is displayed. Dysregulated molecular pathways and immune infiltration associated with prognosis To further explore the biological mechanisms associated with DeepMR-LACC high- and low-risk groups, radioproteomics analysis was performed based on paired tumor biopsy samples from the radioproteomics cohort. In total, 642 DEPs among 7722 proteins with fold changes (FC) > 1.2 and Wilcoxon test p values < 0.05 were identified (Fig. 4A, Supplementary Table S9). The heatmap of the 642 DEPs showed that 484 DEPs were remarkably downregulated and 158 DPEs were upregulated (Fig. 4B). The glycine, serine, and threonine metabolism were significantly upregulated in the DeepMR-LACC high-risk groups (Fig. 4C). Sun et al. have also found that the growth of cervical cancer was crucially dependent on serine, which was regulated by serine incorporators 221. Our GSEA results also revealed dysregulated pathways including downregulated complement and coagulation cascades, upregulated ECM-receptor interaction, focal adhesion and vascular smooth muscle contraction, which were also identified in a poorer prognostic cohort from a proteogenomic study of cervical cancer22. Meanwhile, the biological processes associated with the DEPs were involved mainly in actin filament organization, phagocytosis, and ribosome biogenesis (Fig. 4D). The upregulated ribosome biogenesis identified in high DeepMR-LACC risk group was aligned with the results in the previous proteogenomic study22, which validated the reliability of our enrichment analysis.Fig. 4Proteomic profiling of dysregulated molecular pathways and immune cell infiltration.A Volcano plot of differentially expressed proteins (DEPs) between DeepMR-LACC high- and low-risk groups. B Heatmap of DEPs between DeepMR-LACC high- and low-risk groups. C Enriched KEGG pathways based on GSEA analysis. D Enrichment of biological process based on DEPs. E Boxplot of immune score between high- and low-risk groups. F Heatmap of relative immune cell infiltration. G Correlation between DeepMR-LACC risk score and immune score. H Correlation between the relative DeepMR-LACC risk score and immune cell infiltration. Proteomic profiling of dysregulated molecular pathways and immune cell infiltration. A Volcano plot of differentially expressed proteins (DEPs) between DeepMR-LACC high- and low-risk groups. B Heatmap of DEPs between DeepMR-LACC high- and low-risk groups. C Enriched KEGG pathways based on GSEA analysis. D Enrichment of biological process based on DEPs. E Boxplot of immune score between high- and low-risk groups. F Heatmap of relative immune cell infiltration. G Correlation between DeepMR-LACC risk score and immune score. H Correlation between the relative DeepMR-LACC risk score and immune cell infiltration. As immune-related pathways were enriched, we assumed that immune cell infiltration would differ between the DeepMR-LACC high- and low-risk groups. Immune scores were increased significantly in the low-risk group compared with the high-risk group (Fig. 4E). The low-risk group showed significant enrichment in monocytes, resting mast cells, plasma B cells, memory B cells, gamma delta T cells, neutrophils, and common lymphoid progenitors (all p < 0.05), representing higher immune cell infiltration (Fig. 4F). Lower resting mast cells has shown poorer prognosis and higher risk of recurrence in cervical cancer23,24. Kim at al. have revealed that the plasma B cells can improve OS in cervical cancer which was stimulated by the radiotherapy25. The infiltration of memory B cells in the tumor microenvironment contributes to improved overall and disease-free survival rates in cervical cancer26. Furthermore, Li et al. has validated that the presence of gamma delta T cells in cervical tumor tissues is positively linked to improved patient outcomes27. These research results have confirmed the higher immune cells infiltration was correlated with lower risk of outcomes in our study. The DeepMR-LACC score correlated negatively with the immune score (Fig. 4G), which was mainly caused by neutrophils, memory B cells, and resting mast cells (Fig. 4H). Development of radioproteomics-based prognostic stratification Based on the univariate Cox regression and the random survival forest algorithm, seven key prognostic proteins were selected (Fig. 5A). Of these seven proteins, DKC128 and NOP229 have been reported in cervical cancer, and CEBPZ30 and WDR7431 have been reported in squamous cell carcinoma; MRPS15, PWP2, and RBM28 have not been reported in cervical cancer yet (Supplementary Table S10). NOP2, DKC1, WDR74, and PWP2 were found to be related to ribosome biogenesis and rRNA processing, which play key roles in cancer development (Fig. 5B)32.Fig. 5Protein biomarker selection and radioproteomics-based prognostic stratification.A Workflow of key prognostic protein biomarker selection. B Protein-protein interaction between seven prognostic protein biomarkers. C Heatmap and clustering of seven prognostic protein biomarkers. D Forest plot based on multivariate Cox analysis for overall survival. E Correlation between seven prognostic protein biomarkers and DeepMR-LACC risk score. F Sankey plot based on DeepMR-LACC, proteomic-based risk stratification, and radioproteomics-based stratification. G Kaplan-Meier curve analysis based on radioproteomics-based stratification. Protein biomarker selection and radioproteomics-based prognostic stratification. A Workflow of key prognostic protein biomarker selection. B Protein-protein interaction between seven prognostic protein biomarkers. C Heatmap and clustering of seven prognostic protein biomarkers. D Forest plot based on multivariate Cox analysis for overall survival. E Correlation between seven prognostic protein biomarkers and DeepMR-LACC risk score. F Sankey plot based on DeepMR-LACC, proteomic-based risk stratification, and radioproteomics-based stratification. G Kaplan-Meier curve analysis based on radioproteomics-based stratification. Furthermore, the profiles of the seven proteins divided the radioproteomics cohort into two clusters by Canberra distance (Fig. 5C). Cluster 1 contained 30 patients whose prognoses were more favorable than those of patients in cluster 2 (Supplementary Fig. S9). In addition, the proteomics-based cluster was associated significantly with prognosis, outperforming the DeepMR-LACC score and clinical variables (Fig. 5D). The Spearman correlation analysis showed that DeepMR-LACC scores negatively correlated with levels of the seven key prognostic proteins (Fig. 5E). Finally, we combined the proteomics-based clustering with the DeepMR-LACC risk levels for radioproteomics-based risk stratification into high-, medium-, and low-risk groups (Fig. 5F). The radioproteomics-based low-risk group had a more favorable prognosis than the high-risk group (Fig. 5G and Supplementary Fig. S10). For the DeepMR-LACC score, the C-indices for PFS and OS in the radioproteomics cohort were 0.66 (95% CI, 0.47–0.85) and 0.70 (95% CI, 0.57–0.83), respectively (Supplementary Table S7 and Supplementary Fig. S11). Radioproteomics-based risk stratification increased the C-indices for PFS and OS to 0.85 (95% CI, 0.74–0.96) and 0.85 (95% CI, 0.73–0.96), respectively, reflecting improved prognostic accuracy for LACC. Discussion In this study, we developed an automatic prognostic prediction system for LACC following CCRT by integrating MR imaging and proteomics data. Specifically, the DeepMR-LACC model showed promising predictive performance and was consistently validated in an independent cohort. The DeepMR-LACC score was shown to be a predictor independent of clinical variables in a multivariate Cox analysis, outperforming current clinical variables. Notably, we showed that the radioproteomics-based framework could not only identify the distinct immunosuppressive microenvironment in the DeepMR-LACC high-risk group but also significantly improve prognostic accuracy. To our knowledge, DeepMR-LACC is the first DL-based model for LACC prognosis prediction using sagittal MR imaging data. Of note, our study demonstrates several key advantages over previously published works (Supplementary Table S11)33–38. First, unlike most existing single-center studies, our model was developed and validated on a multicenter cohort of 693 patients from two independent cancer centers, which significantly enhances its generalizability and clinical applicability. Second, our end-to-end automated pipeline eliminates the need for manual region-of-interest delineation, thereby reducing inter-observer variability—a common limitation in manual annotation-based approaches. Third, by utilizing sagittal T2-weighted MRI, our model captures the craniocaudal tumor extent essential for accurate FIGO staging and radiotherapy planning, whereas relied on axial plane imaging39. As a result, our model achieved superior prognostic performance, with C-indices of 0.85 (95% CI: 0.74–0.96) for PFS and 0.85 (95% CI: 0.73–0.96) for OS—the highest among all compared studies—highlighting its strong potential for clinical adoption. The DeepMR-LACC model facilitates risk-stratified treatment personalization. For low-risk patients, de-escalation strategies may include: (1) six-week carboplatin–paclitaxel induction chemotherapy before chemoradiotherapy40; (2) substitution of radical hysterectomy for chemoradiation if induction yields complete or near-complete response with residual tumor ≤4 cm; (3) confinement of radiation to the pelvis for para-aortic node-negative cases; and (4) use of single-agent platinum for radiosensitization. These approaches can reduce toxicity, improve quality of life, and decrease costs, consistent with guidelines minimizing overtreatment. For high-risk patients, the 2025-V4 NCCN (National Comprehensive Cancer Network) Guidelines recommend: (1) chemoradiotherapy plus pembrolizumab with maintenance41; (2) extended-field radiotherapy for para-aortic node involvement42; and (3) MRI-guided adaptive brachytherapy for bulky or refractory tumors43. The DeepMR-LACC model shows potential to aid in early identification of high-risk patients, which may help clinicians tailor treatment intensity with the goal of improving survival outcomes, while also potentially reducing unnecessary toxicity for those classified as lower risk. The intrinsic black-box nature of current DL models poses a significant challenge for interpretability and transparency, which constrains the generalizability of the model developed in this study44. To address this issue, we utilized attention rollout to visualize the distribution of attention weights in the tumor ROIs in the model’s input. This result suggests that peripheral lesions and adjacent tissues outside the tumor region may also be correlated with patient prognosis, offering a more comprehensive understanding of the model’s decision-making process in real-world clinical practice45. DeepMR-LACC highlighted fibrous hyperplasia within the peritumoral area, offering a biologically grounded interpretation of its predictions. By remodeling the collagen network and activating cancer-associated fibroblasts, this fibrosis fosters an immunosuppressive microenvironment, thereby driving cervical cancer progression and modulating therapeutic response46. Given the noninvasiveness and availability of MRI examination, the deep learning–based fully automated model is conducive to clinical deployment, especially in economically underdeveloped areas. Our comprehensive framework further assessed the complex biological mechanisms that lead to differential clinical outcomes of imaging-based risk stratification. Notably, KEGG pathways that promoted tumor progression were found to be enriched in the high-risk group based on the exploration of the proteomic data. Specifically, the excessive activation of amino acid metabolism is required for tumor progression47. Furthermore, our results showed that ECM receptor interaction, focal adhesion, and P53 signaling pathways pose significant obstacles to the achievement of favorable radiotherapy outcomes, aligning with previous research emphasizing their contribution to radioresistance48–50. Additionally, the cancer-altered ECM modulates immune cells, creating an immunosuppressive microenvironment that results in cancer development and metastasis51. The molecular pathway enrichment findings were consistent with the results of the immune infiltration analysis, revealing a notably activated TME in the DeepMR-LACC low-risk group. This enhanced anti-tumor immunity has been demonstrated to impede tumor progression and improve prognoses52. Our study exhibits a novel perspective, enriching insights into tumor heterogeneity. Integrating tumor proteomic data into the prognostic models serves decision-making explanation for DeepMR-LACC. CCRT facilitates the liberation of tumor antigens through the induction of immunogenic cell death, augments the infiltration of lymphocytes into tumors by triggering the synthesis of injury-related molecular patterns, and elevates the expression levels of immune-checkpoint molecules on the surface of cancer cells53,54. Therefore, the researchers have postulated that incorporating immunotherapy into the first-line treatment of LACC could lead to more patients being cured55. Our finding shows significant clinical implications since the identification of low-risk LACC cases within the same clinical stage can help prevent unnecessary incorporation of therapy. On the other hand, an immunosuppressive microenvironment in the DeepMR-LACC high-risk group may enhance the immune evasion of tumor cells, which compromises the radiation-induced immune regulation. However, the lower immune cell infiltration may invent more effective targetable molecules for LACC in high-risk groups to overcome treatment resistance and improve survival. The seven key prognostic proteins may be worth further investigation as new therapeutic targets to enhance CCRT response in high-risk groups. Proteomic profiling provides high-throughput protein quantification capabilities while requiring minimal tissue samples, making it feasible for clinical applications using pretreatment cervical biopsies. This approach enables pre-therapeutic risk stratification, post-treatment prognosis evaluation, and precision target selection56. On the basis of DeepMR-LACC, further integrating proteomic stratification enhances the reliability of predicting outcomes. Three patients with follow-up durations exceeding 60 months who did not experience recurrence or death were incorrectly placed in the high-risk group by the DeepMR-LACC but were accurately placed in the low-risk group (cluster 1) in the proteomics-based stratification. Conversely, one patient who died within 10 months was placed in the low-risk group in the proteomics-based stratification, but was more reasonably placed in the high-risk group by the DeepMR-LACC. Although the two stratification systems displayed some discrepancies in prognostic assessment, they demonstrated a certain level of complementarity. Notably, integrating multimodal data (e.g., MRI, genomics, and serum biomarkers) provides a more comprehensive and accurate tumor assessment through multidimensional characterization. Radiomic features mainly highlight tumor texture, whereas proteomic features focus on molecular function. Both contribute to a richer depiction of tumor heterogeneity and assist in risk stratification. By harnessing the strengths of proteomics and radiomics, clinicians can gain a more holistic assessment of patient outcomes. This study has several limitations. First, our dataset was derived exclusively from Chinese patients, the clinical characteristics and pathological subtype distribution may differ for patients from other regions. Second, the radioproteomics cohort was relatively small, which may limit the broad applicability of our radioproteomics-based risk stratification. Multimodal integration of image features and molecular signatures has shown to be superior for cancer detection and prognosis57,58. However, most currently accessible datasets are incomplete and do not include all require modalities. This limitation is primarily attributable to the limited availability of matched clinical samples, and high cost or labor intensity of multi-omics experiments. We for the first time proposed a radioproteomics based model for cancer prognosis, and we emphasized that the robustness of the model needs external validation in larger multimodal datasets before clinical translation. Third, the retrospective design of this study carries an inherent risk of bias that might have influenced the interpretation of the results. To ensure the effective application of the integrated framework in the future, a larger prospective cohort with paired proteomics data is still warranted. In summary, we developed and externally validated a deep learning-based, fully automated model (DeepMR-LACC) for long-term prognostic prediction based on MR imaging. The DeepMR-LACC score emerged as an independent prognostic factor among clinical features, highlighting its potential in clinical practice. Additionally, paired proteomics analysis revealed a distinct tumor microenvironment associated with the DeepMR-LACC score. Radioproteomics-based risk stratification further enhanced model performance. Integrating routine MR imaging and paired proteomic profiling can contribute to individualized clinical decision making. Methods Patient cohorts This retrospective study was approved by the Research Ethics Committees of the Zhejiang Cancer Hospital [IRB-2024-434(IIT)]. In addition, we recruited patients from Harbin Medical University Cancer Hospital following approval by the ethics board (2024-331-R). The requirement for obtaining informed consent from the patients was waived. The study was conducted according to the principles of the Declaration of Helsinki. In total, 693 patients with locally advanced cervical squamous cell carcinoma (CSCC) forming four cohorts who attended the two participating hospitals were enrolled. The primary cohort (n = 600) was enrolled in Zhejiang Cancer Hospital between December 2012 and July 2018. It was divided into a training cohort (n = 386), an internal test cohort (n = 168), and a radioproteomics cohort (n = 46) for which tissue proteomics data were available. The external test cohort (n = 93) was from Harbin Medical University Cancer Hospital between January 2015 and December 2017. The inclusion criteria were: (A) diagnosis of cervical squamous cell carcinoma, confirmed histologically within 2 weeks before treatment initiation; (B) availability of pre-treatment magnetic resonance imaging (MRI) examinations performed within 2 weeks before treatment initiation and showing measurable cervical tumor lesions; (C) 2009 International Federation of Gynecology and Obstetrics (FIGO) stage indicating locally advanced disease (IB2–IVA); (D) treatment consisting of concurrent chemoradiotherapy; and (E) availability of complete follow-up data for the first, third, and fifth years after treatment. The exclusion criteria were: (A) concurrent malignant tumor or history of tumor, (B) receipt of neoadjuvant therapy. Image acquisition and automatic segmentation All enrolled patients underwent MRI examinations within 2 weeks before treatment. MRI examinations were performed using Siemens and Philips 1.5 T and 3.0 T MRI scanners, with pixel spacing of 0.391–0.875 and slice thicknesses of 3.0–6.5 mm; the specific parameter settings are provided in Supplementary Table S1. Slices from T2-weighted sequences were selected for analysis. In this study, all patient MRI data underwent standardized preprocessing to address image quality issues arising from MRI signal intensity inhomogeneity. This included N4 bias field correction followed by intensity normalization using min-max scaling. To optimize preprocessing, we systematically evaluated the impact of image resampling by comparing model performance between resampled and non-resampled datasets. Results showed that model performance did not improve after resampling; therefore, resampling was not applied in the final preprocessing pipeline (Supplementary Table S2 and S3). To mitigate the labor-intensive task of manually delineating tumor regions of interest (ROIs), we employed the UNet59 segmentation model for automatic segmentation (Fig. 6A). To improve the accuracy of the tumor masks generated by the segmentation model, we used a publicly available medical image segmentation dataset (Synapse Multi-Organ Segmentation Dataset, https://www.synapse.org/#!Synapse:syn3193805/wiki/217789) to pretrain the UNet.Fig. 6Overall study design.A Image segmentation process. The image data undergoes preprocessing using the trained segmentation model UNet for automatic delineation of the ROI, which serves as the input for the subsequent DeepMR-LACC model. B The architecture of DeepMR-LACC. The DeepMR-LACC consists of Inception V1, Channel Attention, Spatial Attention, and Vision Transformer. C Clinical cohorts and model evaluation. Model evaluation was based on training cohort, internal test cohort, external test cohort and, radioproteomics cohort. The evaluation includes C-Index, time-dependent ROC curve, KM survival curve, and Cox regression analysis. D Overview of proteomic analysis and construction of radioproteomics-based prognostic stratification. ROI region of interest, C-Index Concordance Index, ROC receiver operating characteristic, KM Kaplan-Meier, DIA data-independent acquisition. Overall study design. A Image segmentation process. The image data undergoes preprocessing using the trained segmentation model UNet for automatic delineation of the ROI, which serves as the input for the subsequent DeepMR-LACC model. B The architecture of DeepMR-LACC. The DeepMR-LACC consists of Inception V1, Channel Attention, Spatial Attention, and Vision Transformer. C Clinical cohorts and model evaluation. Model evaluation was based on training cohort, internal test cohort, external test cohort and, radioproteomics cohort. The evaluation includes C-Index, time-dependent ROC curve, KM survival curve, and Cox regression analysis. D Overview of proteomic analysis and construction of radioproteomics-based prognostic stratification. ROI region of interest, C-Index Concordance Index, ROC receiver operating characteristic, KM Kaplan-Meier, DIA data-independent acquisition. To achieve automatic segmentation, images from 440 patients were manually annotated. A radiologist (F.W.) and a radiation oncologist (Y.F.) with 13 and 8 years of experience in cervical cancer imaging diagnosis and target delineation, respectively, independently delineated ROIs on all T2-weighted slices. During manual tumor segmentation, in cases where the two experts disagreed, the final delineation was determined by a senior radiology expert (L.S.) with 17 years of specialization in cervical cancer imaging. Slices without tumors or with unclear tumors were excluded, resulting in a total of 1779 MRI slices with tumor ROI. The inter-observer variability was assessed with a mean Dice coefficient of 0.77, indicating the robustness of our segmentation ground truth for model training. All patients were randomly divided into the training, validation, and testing cohorts in a 6:2:2 ratio, with each MRI slice considered as an independent sample. The pre-trained model was fine-tuned on the training cohort to achieve automatic segmentation of tumors in patients. To mitigate overfitting during the training process, we applied data augmentation to the input images, including random rotation (fixed 90° rotation and free rotation within the range of −30° to 30°), random horizontal, and vertical flipping. The network employed the sum of the Dice60 and binary cross-entropy losses as its loss function. Dice loss evaluates regional overlap between predicted and ground-truth masks, penalizing misalignment of the entire foreground area. Binary cross-entropy loss evaluates pixel-wise prediction accuracy, penalizing individual false positives and false negatives. The input slices were resized to a resolution of 224 × 224, and the stochastic gradient descent (SGD) was used as the optimizer. The model was trained for 100 epochs with a batch size of 24. The initial learning rate was set to 2 × 10−3 and was reduced by 0.1 each time it reached a plateau. We used PyTorch’s ReduceLROnPlateau scheduler (mode = ‘max’, patience=5) to dynamically reduce the learning rate with the goal of maximizing the Dice coefficient in the validation cohort. The code was implemented in the PyTorch 1.13.1 platform, and the model was trained on an NVIDIA GeForce RTX 4070 Super GPU and an AMD Ryzen 9 5950X CPU. In the end, the segmentation model achieved a Dice coefficient of 0.7792 in the testing cohort. Prognostic model (DeepMR-LACC) In this study, we proposed a prognostic prediction network (DeepMR-LACC) which was based on the Vision Transformer (ViT) architecture61, incorporating modules from the InceptionV1 convolutional network62, channel attention63, and spatial attention64 (Fig. 6B). The pre-trained ViT can be obtained from the timm library and has a modified ViT-B/16 architecture, with the number of layers reduced from 12 to 9 to accommodate the dataset size. The slice from each patient with the largest tumor area was selected as the model input. A minimal bounding box was precisely positioned to encapsulate the tumor region. In order to recognizing the potential prognostic significance of satellite lesions and adjacent tissues beyond the tumor margins, the bounding box was then expanded by 10 pixels following normalization and cropping. The cropped tumor regions were reshaped to 224 × 224. We augmented the input data to minimize the risk of overfitting during training, using random rotation (fixed 90° rotation and free rotation within the range of –30° to 30°), random horizontal and vertical flipping, and random shifting (within 0–10% of the total width and height). We compared two input strategies for model training: (1) triplicate replication of the slice with the largest tumor area and (2) stacking the tumor slice with its adjacent superior and inferior slices, both generating 224×224×3 input images. Comparative analysis of two input strategies revealed that triplicating the slice with the largest tumor area yielded superior prognostic performance in the DeepMR-LACC model, and was consequently selected for subsequent analyses (Supplementary Table S4). We used the Adam optimizer with a batch size of 24. The initial learning rate was set to 1 × 10−5 and was reduced by 0.1 every five epochs. The Cox loss function, which relies on Cox partial likelihood to differentiate survival risks among patients was used for model training45,65. Our model jointly optimizes PFS and OS through a shared Cox loss, representing a multi-objective learning framework without task-specific heads. The code was implemented in the PyTorch platform, and the model was trained on an NVIDIA GeForce RTX 4070 Super GPU and an AMD Ryzen 9 5950X CPU. The total loss function formulation:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{total}}={L\left(\theta \right)}_{{PFS}}+{L\left(\theta \right)}_{{OS}}$$\end{document}Ltotal=LθPFS+LθOSwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{total}}$$\end{document}Ltotal denotes the total loss, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L\left(\theta \right)}_{{PFS}}$$\end{document}LθPFS and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L\left(\theta \right)}_{{OS}}$$\end{document}LθOS represent the Cox loss for PFS and OS, respectively. The Cox loss is defined as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L\left(\theta \right)=-\mathop{\sum }\limits_{i:{E}_{i}=1}\left({\hat{h}}_{\theta }\left({x}_{i}\right)-\log \mathop{\sum }\limits_{j{\mathfrak{\in }}{\mathfrak{R}}\left({T}_{i}\right)}{e}^{{\hat{h}}_{\theta }\left({x}_{j}\right)}\right)$$\end{document}Lθ=−∑i:Ei=1hˆθxi−log∑j∈RTiehˆθxjwhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${E}_{i}$$\end{document}Ei represents the patient’s outcome status, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{h}}_{\theta }$$\end{document}hˆθ is the DeepMR-LACC score, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}$$\end{document}xi represents the ith sample, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathfrak{R}}\left({T}_{i}\right)$$\end{document}RTi is the list of patients with shorter survival times than patient i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{j}$$\end{document}xj represents the jth patient on the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathfrak{R}}\left({T}_{i}\right)$$\end{document}RTi. Considering the black-box nature of DL models, we used attention rollout to visualize the model’s attention and statistically analyzed the attention distribution between intratumoral and peritumoral regions. High attention regions were displayed in yellow to red, while low-attention regions appeared blue. Therefore, high- and low-attention regions can be distinguished by their RGB values. We identified high-attention regions using the R-channel values of the RGB attention heatmap, with the channel threshold of 50. When the R-channel value exceeds this threshold, the point is considered a high-attention area. By combining the high-attention area with the tumor masks generated by the segmentation model, we established the following classification criteria: (1) Intratumoral attention: When >70% of the high-attention area falls within the tumor mask region. (2) Peritumoral attention: When >70% of the high-attention area lies outside the tumor mask. (3) Dual attention: All other cases where attention is distributed both inside and outside the tumor region. Additionally, an independent intermediate radiology expert (Q.L.) was invited to evaluate the attention maps. Clinical outcomes The clinical outcomes examined in this study were progression-free survival (PFS) and overall survival (OS). PFS refers to the time from treatment to any tumor progression (local or distant recurrence, other malignancy, or last known date without progression) or death from any cause. OS refers to the time from treatment to death. Model evaluation and statistical analysis The Harrell concordance index (C-index) was used to evaluate the prognostic performance of the model for all cohorts. The 95% confidence intervals were estimated through bootstrap resampling with 1000 iterations. We further assessed the model by plotting Kaplan-Meier survival and time-dependent receiver operating characteristic (ROC) curves and calculated the areas under the curves (AUCs) for 1-, 3-, and 5-year PFS and OS. In the Kaplan-Meier survival analysis, we used the median DeepMR-LACC score from the training cohort as the threshold for stratification into high (score > threshold) and low (score ≤ threshold) risk groups. We then examined prognostic differences between the groups. In the ROC analysis, we normalized the DeepMR-LACC scores, treating them as probabilities of 1-, 3-, and 5-year PFS and OS. We first calculated the mean and standard deviation of the DeepMR-LACC scores in the training cohort and applied z-score normalization across all cohorts. Scores were then scaled to [0,1] using a sigmoid function. Model clinical utility was evaluated through decision curve analysis comparing DeepMR-LACC against established clinical predictors, and calibration analysis to assess the agreement between predicted probabilities and observed outcomes. Additionally, we conducted univariate and multivariate Cox regression analyses to estimate the hazard ratio (HRs) for DeepMR-LACC score and predictive factors, including age, stage, pre-treatment squamous cell carcinoma antigen level (pre-SCC), and tumor size. Finally, subgroup analyses were performed according to the clinical factors of age, stage, pre-SCC level, and tumor size (Fig. 6C). The criteria for selecting thresholds of age and pre-SCC level were based on the median values in the training cohort (59 years and 7.6 ng/mL respectively). The tumor size threshold was set as 4.0 cm, which was based on the clinical practice that 4 cm represents the optimal cutoff for predicting long-term survival in cervical cancer66. The statistical analysis of the clinical characteristics was performed using IBM SPSS 26 (IBM Corp., Armonk, NY, USA). The unpaired t test was used for continuous variables and the chi-squared test was employed for categorical variables. All statistical tests were two tailed, with p < 0.05 indicating significance. Proteomics data acquisition and preprocessing Proteomics data were obtained from the patients in the radioproteomics cohort, who underwent cervical biopsy, using liquid chromatography-mass spectrometry (Fig. 6D). The 46 cervical biopsy samples were homogenized in lysis buffer, followed by sonication and centrifugation to obtain tissue proteins. Proteins from each sample were then reduced with dithiothreitol, alkylated with iodoacetamide, and digested with trypsin. The tryptic peptides were analyzed on an EASY-nLC 1200 nano-liquid chromatography system (Thermo Fisher Scientific, Dreieich, Germany) coupled with an Orbitrap Exploris 480 mass spectrometer (Thermo Fisher Scientific, Dreieich, Germany) operating in data-independent acquisition (DIA) mode. The raw data were processed using Spectronaut v.18 (Biognosys, Zurich, Switzerland). Subsequently, the resulting 12,427 proteins were converted into a protein abundance matrix for downstream analysis in R 4.2.1 software (https://www.r-project.org/). Data pre-processing consisted of the following steps. Protein abundance values below 10 were regarded as missing values. Median normalization was applied, followed by imputation using the k-nearest neighbor approach. Samples with more than 20% missing values were excluded, and the remaining data were log2 transformed. In total, 7722 high-quality proteins were retained for further analysis. Biological pathway and immune microenvironment analyses For the biological interpretation of the DeepMR-LACC scores, pathological mechanisms were explored in the high- and low-risk groups. The differentially expressed proteins (DEPs) were identified with fold change (FC) ≥ 1.2 and Wilcoxon test p < 0.05 (uncorrected) to retain biologically relevant proteins that might show subtle but coordinated expression changes in disease-related pathways67. The KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway enrichment analysis was conducted using gene set enrichment analysis (GSEA), and the gene ontology enrichment analysis was conducted using enrichGO, both implemented with the “clusterProfiler” R package68. Relative quantification of immune cell infiltration in the radioproteomics cohort was performed based on the xCell69, CIBERSORT70, and ESTIMATE71 methods. The ESTIMATE was employed for calculating the overall immune score, which represented the infiltration of immune cells. The xCell and CIBERSORT were used for relative quantification of specific immune cell types. The comparison of specific immune cell infiltration between DeepMR-LACC high- and low-risk groups was conducted by Kruskal–Wallis test, with p value < 0.05 considered significant. Selection of prognostic proteins To explore the potential of proteomics data in the prediction of LACC prognosis, machine learning methods were applied to select key prognostic proteins. Features were selected as follows. Based on proteomics expression profiles, univariate Cox regression analysis was performed to identify 321 prognostic proteins for OS and 135 prognostic proteins for PFS (all p < 0.01, uncorrected). This approach follows established practice in biomarker discovery studies72–74. The threshold (p < 0.01) balances statistical stringency with biological comprehensiveness, allowing downstream modeling to autonomously determine the most relevant features. The remaining proteins were then used for feature selection based on a random survival forest model. The models were trained 100 times with different initial states. In each training run, the 50 proteins of greatest importance were recorded. Proteins selected more than 50 times were defined as key prognostic features for OS or PFS. Proteins identified as prognostic for both OS and PFS were selected to construct a prognostic protein panel. A protein-protein interaction network was constructed using GeneMania75. Heatmaps were generated using the “pheatmap” R package, with hierarchical clustering according to Canberra distance. The clustering divided patients into two groups with different risks based on protein panel abundance. The leave-one-out cross-validation (LOOCV) was employed to validate the robustness of the clustering method based on the seven protein biomarkers. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting summary Supplementary Information Reporting summary