Evaluating Deconvolution Methods Using Real Bulk RNA-expression Data for Robust Prognostic Insights in Pan-Cancer

Evaluating Deconvolution Methods Using Real Bulk RNA-expression Data for Robust Prognostic Insights in Pan-Cancer | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Evaluating Deconvolution Methods Using Real Bulk RNA-expression Data for Robust Prognostic Insights in Pan-Cancer Minghan Li, Yuqing Su, Weidong Tian This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6389993/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 21 Jan, 2026 Read the published version in Genome Biology → Version 1 posted 9 You are reading this latest preprint version Abstract Deconvolution of bulk RNA-expression data unlocks the cellular complexity of cancer, yet traditional pseudobulk benchmarks falter in real-world settings where absolute cell proportions are unknown. Here, we introduce a novel real-data framework, leveraging 16 real bulk RNA-expression cohorts (4,576 samples) across eight cancer types to evaluate six deconvolution methods based on relative changes in differentially proportioned (DP) cell types—an impartial and reliable metric. Across three innovative benchmark scenarios—consistency with scRNA-seq, reproducibility across cohorts, and prognostic relevance—ReCIDE, Bisque, and BayesPrism have been demonstrated to be the three most robust deconvolution methods. analysis of ten cancer types revealed matrix cancer-associated fibroblasts (mCAF) as a poor prognosis marker (p = 0.0081) and CLEC9A + dendritic cells (cDC_CLEC9A) as a favorable one (p = 0.016). Furthermore, a prognostic indicator (ASC% - mCAF%) developed using ReCIDE was validated across five TCGA and three GEO cohorts. This study broadens deconvolution benchmarking, offering actionable tools for precision oncology and guiding method selection for translational research. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction The cellular heterogeneity of cancer critically impacts patient outcomes and treatment strategies, yet unraveling this complexity in clinical samples remains a difficult task 1 , 2 , 3 , 4 . Single-cell RNA sequencing (scRNA-seq) has begun to pierce this veil, revealing gene expression at unprecedent resolution 5 , 6 , 7 , but its cost restricts large-scale studies linking cell types to patient prognosis 8 , 9 . Conversly, bulk RNA-seq and microarray data from repositories like TCGA and GEO 10 , 11 , 12 , 13 , rich with clinical information, offer a scalable alternative. Deconvolution methods bridge these domains, estimating cell type proportions from bulk data using scRNA-seq references to probe disease associations 14 , 15 . Yet, a notable challenge persists. Despite advances in methods like CIBERSORT 16 , DWLS 17 , MuSiC 18 , Bisque 19 , BayesPrism 20 , and our recently developed ReCIDE 21 , current benchmarking studies for deconvolution methods invariably lean on pseudobulk data or flow cytometry 22 , 23 , 24 , 25 , assuming known absolute cell type proportions. In real bulk RNA-expression deconvolution, such precision is a mirage, confounded by biological and technical noise, underminging deconvolution’s clinical utility 26 , 27 . To address this, we propose a novel benchmark strategy: evaluating methods based on relative changes in differentially proportioned (DP) cell types—those shifting significantly between conditions—a metric robust across datasets and aligned with clinical research priorities. Here, we assessed six deconvolution methods across 16 real bulk RNA-expression cohorts from eight cancer types, totalling 4,576 samples using three scenarios: (1) DP cell type alignment with scRNA-seq, (2) reproducibility of DP cell types between cohorts, and (3) reproducibility of prognositic related (PR) cell types between cohorts. Our analysis revealed that ReCIDE, Bisque, and BayesPrism exhibit superior performance across these benchmarks. Furthermore, pan-cancer analyses using these three methods identified matrix cancer-associated fibroblasts (mCAF) and CLEC9A + dendritic cells (cDC_CLEC9A) as key survival predictors, unlocking actionable insights. Leveraging these findings, we constructed three prognostic prediction indicators utilizing TCGA cohorts. Among them, the (ASC% - mCAF%) indicator, derived from ReCIDE deconvolution results, demonstrated robust clinical relevance through rigorous validation across three independent GEO cohorts. This real-data approach not only refines method selection but also advances precision oncology by linking cellular insights to patient outcomes. Results Benchmark design Traditional benchmarks of cell type deconvolution methods rely on pseudo-bulk expression data, generated by artificially mixing scRNA-seq profiles with predefined proportions and added noise. Yet, evidence mounts that these fail to mirror real bulk RNA-seq or microarray profiles 26 , 27 , where absolute cell proportions remain elusive. As an alternative, we designed a novel benchmark that emphasizes the identification of differentially proportioned (DP) cell types that exhibit significant relative proportional changes across disease conditions, rather than striving for precise absolute quantification ( Fig. 1 A-E ) . This approach aligns with the priorities of clinical research, as investigators are predominantly focused on cell types that have significant implications for disease progression and clinical outcomes. Initially, we assembled datasets from 10 cancer types, including Breast Cancer (BRCA) 28 , Colon Adenocarcinoma (COAD) 29 , Esophageal Carcinoma (ESCA) 30 , Head and Neck Squamous Cell Carcinoma (HNSC) 31 , Kidney Renal Clear Cell Carcinoma (KIRC) 32 , Liver Hepatocellular Carcinoma (LIHC) 33 , Non-Small Cell Lung Cancer (NSCLC) 34 , Pancreatic Adenocarcinoma (PAAD) 35 , Prostate Adenocarcinoma (PRAD) 36 , and Stomach Adenocarcinoma (STAD) 37 , each contrasting two conditions (cancer vs. normal in seven, subtypes in three), paired with two matching bulk cohorts from GEO and TCGA. When compare the cellular composition in subdivided cancer subtypes, we analyze each cell type individually, treating tumor cells and epithelial cells as two distinct cell types. When evaluating the cellular composition in tumor and normal tissues, we adopted the cell type proportion comparison strategy proposed by Lee et al. to ensure objective benchmarking 38 . This strategy involved combining tumor cells and epithelial cells in tumor tissues and comparing them with epithelial cells in normal tissues. Due to the lack of explicit annotation of epithelial cells in KIRC and STAD, we prudently excluded these two cancer types from the benchmark evaluation. Therefore, in the benchmark, we included a total of eight cancer types (ten RNA-seq cohorts, six microarray cohorts; 4,576 samples; Table 2 ). In selecting deconvolution methods for our benchmark, we referred to the benchmark tests conducted by Cobos et al. 23 and Chu et al. 20 Based on their benchmark results, we selected five high-performenced deconvolution methods, including CIBERSORT, DWLS, MuSiC, Bisque, BayesPrism, and incorporated our newly developed method ReCIDE for comprehensive benchmarking across diverse scenarios ( Fig. 1 A, B ) . Table 2 RNA-seq/Microarray datasets used in this research. Disease TCGA_Bulk GEO_Bulk GEO_number GEO_type BRCA 726ER + BC and 175TNBC 11ER + BC and 9TNBC GSE176078 62 RNA-seq COAD 41N and 465T 19N and 566T *GSE39582 63 Microarray ESCA 13N and 173T 10N and 10T GSE149609 64 RNA-seq HNSC 65 HPV- HNSC and 21 HPV + HNSC 121 HPV- HNSC and 60 HPV + HNSC *GSE65858 65 Microarray KIRC 72N and 532T 72N and 72T GSE53757 66 Microarray LIHC 50N and 369T 30N and 64T GSE87630 67 Microarray NSCLC 58LUAD and 51LUSC 133LUAD and 43LUSC *GSE42127 68, 69 Microarray PAAD 4N and 142T 46N and 145T *GSE71729 70 Microarray PRAD 51N and 483T 160N and 264T GSE62872 71, 72 Microarray STAD 13N and 97T 100N and 300T GSE66229 73 Microarray * represents GEO cohorts with survival information. All TCGA cohorts have survival information. We defined DP cell types from scRNA-seq as the reference, then evaluated each method across three innovative​ scenarios: (1) DP cell types alignment with scRNA-seq ( Fig. 1 C ) , (2) reproducibility of DP cell types between cohorts ( Fig. 1 D ) , and (3) reproducibility of prognosis-related (PR) cell types between cohorts ( Fig. 1 D ) , leveraging survival data where available. This comprehensive workflow enhances the scope of deconvolution benchmarking by emphasizing clinical applicability and methodological robustness. Details on data download and DP/PR cell type identification are provided in the Methods an Data available section. Consistency of DP cell types between deconvolved bulk and scRNA-seq data To evaluate how well deconvolution methods identify differentially proportioned (DP) cell types in real bulk expression data, we compared their results to DP cell types defined by scRNA-seq across six methods. Specifically, we employed the F1-score (Weighted harmonic mean of precision and recall) as our primary metric, using a dual-evaluation strategy—conservative and permissive criteria—to account for sample size disparities between scRNA-seq and bulk cohorts ( Fig. 2 A ) . Under conservative criteria, precision reflects the fraction of DP cell types from deconvolution that match statistically significant with scRNA-seq findings, while permissive criteria include those cell types that, although not statistically significant in scRNA-seq data, exhibit trends consistent with the deconvolution results. Recall, under both, measures the proportion of scRNA-seq DP cell types recovered by deconvolution. The calculations for recall, precision, and F1-score are detailed in the Methods section. We first analyzed Colon Adenocarcinoma (COAD), leveraging its large scRNA-seq dataset (35 normal, 53 tumor samples; 19 cell types). Here, 15 DP cell types were identified from scRNA-seq dataset via Wilcoxon test with FDR correction, with eight enriched and seven reduced in tumors ( Fig. 2 B ) . While absolute proportions from deconvolution diverged from scRNA-seq (e.g., dendritic cells absolute proportion in TCGA-COAD: scRNA-seq median 0.014 tumor vs. 0.005 normal; ReCIDE 0.009 vs. 0.005; Bisque 0.008 vs. 0; DWLS 0.002 vs. 0, Supplementary Fig. 1 ), the relative trends generated by ReCIDE, Bisque, and DWLS were consistent with those from scRNA-seq. These three methods consistently detected significant upregulation of dendritic cells in tumors, mirroring scRNA-seq findings. This reinforces our DP-focused benchmark: despite absolute discrepancies, well-performing methods capture proportional shifts accurately. In COAD, under conservative criteria, ReCIDE led with an F1-score of 0.67 (9/12 DP cell types validated, recall 0.6, precision 0.75), followed by Bisque at 0.62 (8/11 validated, recall 0.53, precision 0.73). MuSiC lagged at 0.24 (3/10 validated, recall 0.2, precision 0.3). Under permissive criteria, the F1-scores of ReCIDE and MuSiC remained unchanged, as none of the DP cell types identified by these two methods were newly validated. In contrast, among the 11 DP cell types identified by Bisque, two cell types were newly validated that aligned with scRNA-seq trends, leading to an improvement in precision to 0.91, and an increase in F1-score to 0.67. Notably, MuSiC, CIBERSORT, and Bisque estimated a proportion of 0 for certain cell types in some samples. For example, in the TCGA-COAD dataset, MuSiC assigned a dendritic cell proportion as 0 for 91% of the samples, which hindered the effective identification of the distribution trends of dendritic cells between tumor colon tissues and normal colon tissues in the deconvolution results. Extending to a pan-cancer analysis across eight cancer types, ReCIDE and Bisque maintained superior consistency with scRNA-seq DP cell types. ReCIDE achieved the highest average recall (0.6) under both conservative and permissive criteria, while Bisque excelled in average precision (0.35 conservative, 0.74 permissive) versus ReCIDE (0.31, 0.69) ( Supplementary Fig. 2 ). Average F1-scores underscored their robust performance: ReCIDE at 0.34 (conservative) and 0.60 (permissive), Bisque at 0.32 and 0.53 (Fig. 2 C, D). In contrast, DWLS and MuSiC scored below 0.3, reflecting limitations in handling heterogeneous data. It is worth noting that the average F1-scores of most deconvolution methods under permissive conditions are higher than that under conservative conditions, primarily attributable to the fact that the DP cell types identified by the deconvolution methods are more easily validated under permissive conditions, thus leading to a general increase in precision and F1-score. In summary, ReCIDE and Bisque are the top performers in this benchmark scenario. ReCIDE’s high recall suits applications like cancer screening, ensuring detection of all relevant DP cell types, while Bisque’s precision advantages favor precise subtyping. Reproducibility of DP cell types across deconvolved bulk cohorts Ensuring deconvolution methods reproducibly identify differentially proportioned (DP) cell types across bulk cohorts with matching disease conditions is vital for real-world applicability, as shown by Fonseca et al.’s validation of ciliated endometrial epithelial cell enrichment in multiple ovarian cancer cohorts using MuSiC 39 . Here, we assessed the reproducibility of DP cell types across paired bulk cohorts for six deconvolution methods. We measured reproducibility with the Jaccard index (Jaccard), calculating the overlap of DP cell types between paired cohorts, and evaluated accuracy using the F1-score, which compares shared DP cell types (those consistently identified across both cohorts) to scRNA-seq-defined DP cell types ( Fig. 3 A ) . We applied conservative and permissive criteria: conservative criteria required statistical significance in both cohorts for a shared DP cell type, while permissive criteria accepted a significant association in one cohort with a consistent trend in the other. For F1-scores, precision is the fraction of shared DP cell types matching scRNA-seq among all discovered shared DP cell types, and recall is the proportion of scRNA-seq DP cell types recovered among shared ones, balancing both metrics (Methods) . Similar to the previous section, we first investigated the reproducibility of different deconvolution methods in Colon Adenocarcinoma (COAD), where 15 scRNA-seq DP cell types were identified. Under conservative criteria, ReCIDE identified 16 DP cell types across TCGA and GEO cohorts, with 12 shared (Jaccard = 0.75) ( Fig. 3 B ) . Of these, nine aligned trends with scRNA-seq, and eight also showed significant changes in scRNA-seq, yielding an F1-score of 0.62 (recall = 0.53, precision = 0.75). Bisque detected 11 DP cell types across TCGA and GEO cohorts, with seven shared (Jaccard = 0.64), among which all aligned trends with scRNA-seq, and six also showed significant changes in scRNA-seq (F1-score = 0.57, recall = 0.40, precision = 1.0). MuSiC identified 14 DP cell types, but only three were shared (Jaccard = 0.21), none aligned trends with scRNA-seq (F1-score, recall, and precision = 0). Under permissive criteria, ReCIDE added one shared cell type (Jaccard = 0.81), but it mismatched scRNA-seq’s trend, dropping F1-score to 0.60 (recall = 0.53, precision 0.69). Bisque added three shared cell types (Jaccard = 0.91), among which all aligned trends with scRNA-seq, and two also showed significant changes in scRNA-seq, raising F1-score to 0.70 (recall = 0.53, precision = 1). MuSiC’s Jaccard rose to 0.29 with one added shared cell type, but F1-score remained 0 as none showed consistent and significant changes in scRNA-seq (recall = 0, precision = 0.25). Across eight cancer types, ReCIDE led with average Jaccard of 0.56 (conservative) and 0.88 (permissive), followed by Bisque (0.40 and 0.78) and BayesPrism (0.43 and 0.71) ( Fig. 3 C, E ) . Average F1-scores for shared DP cell types vs. scRNA-seq ranked ReCIDE (0.67 conservative, 0.70 permissive), Bisque (0.46 and 0.68), and BayesPrism (0.38 and 0.48) highest ( Fig. 3 D, F ) . Compared to the conservative criteria, the gap in average F1-scores between ReCIDE and Bisque narrowed under permissive criteria. This shift occurred because, under conservative criteria, Bisque identified only 3.88 shared DP cell types across cohorts compared to ReCIDE’s 9.88 (Supplementary Fig. 3A) , which resulted in Bisque achieving a lower average recall of 0.34 compared to ReCIDE’s 0.62 (Supplementary Fig. 4A) , impacting its average F1-score. Under permissive criteria, the number of shared DP cell types identified by Bisque increased to 8.5 (ReCIDE: 11.88, Supplementary Fig. 3B ), enhancing Bisque’s average recall to 0.54 (ReCIDE: 0.66, Supplementary Fig. 4B ) and reducing the gap with ReCIDE from 82–22%, leading to a narrowing of the difference in their average F1-scores. Another crucial observation is that shared DP cell types identified by most deconvolution methods under permissive criteria exhibit higher average F1-scores compared to those under conservative criteria ( Fig. 3 F ) . This outcome primarily stems from the fact that, under permissive criteria, the number of shared DP cell types increased compared to conservative criteria, leading to higher recall (Supplementary Fig. 4C) . At the same time, precision remained stable (Supplementary Fig. 4D) . As a result, F1-scores showed improvement. In addition, permissive Jaccard mirrored F1-score ranking ( Fig. 3 E,F ) , offers a robust alternative metric when scRNA-seq validation is unavailable. Based on the comprehensive evaluation of both Jaccard and F1-scores, ReCIDE and Bisque demonstrated superior performance in terms of reproducibility and accuracy. Moreover, under permissive criteria, DP cell types shared between two cohorts demonstrated clear advantages in precision and F1-scores compared to DP cell types identified within individual cohorts ( Fig. 2 D, 3 E ) . This result supports the high reliability of consistent DP cell types identified using deconvolution methods across multiple cohorts with the matching disease conditions in clinical practice. Reproducibility of prognosis-related cell types across deconvolved bulk cohorts Identifying prognosis-related (PR) cell types that correlate with cancer patient survival is crucial for stratifying patients and designing targeted therapies. Here, we assessed the ability of six deconvolution methods to reproducibly detect PR cell types across paired bulk RNA-expression cohorts with matching disease conditions, leveraging survival data from five cancer entities: COAD, HPV-HNSC, LUAD, LUSC, and PAAD, sourced from TCGA and GEO. We evaluated reproducibility using the Jaccard index (Jaccard), which measures the overlap of PR cell types between cohorts, under conservative and permissive criteria. Conservative criteria required statistical significance in both cohorts for a shared PR cell type, while permissive criteria accepted significance in one cohort with a consistent trend in the other ( Methods ). In COAD (Fig. 4 A), ReCIDE identified six PR cell types across TCGA and GEO cohorts, with two shared under conservative criteria (Jaccard = 0.33) and five under permissive (Jaccard = 0.83). BayesPrism detected nine PR cell types, with two shared conservatively (Jaccard = 0.22) and seven permissively (Jaccard = 0.78). In contrast, MuSiC identified no PR cell types in either criteria (Jaccard = 0). Extending to all five cancers (Fig. 4 B, C), ReCIDE achieved the highest average Jaccard: 0.218 (conservative) and 0.740 (permissive), reflecting robust detection of reproducible PR cell types. BayesPrism followed with 0.064 and 0.634, while Bisque scored 0 (conservative) and 0.642 (permissive), excelling only when criteria relaxed. In contrast, under permissive criteria, MuSiC demonstrated an average Jaccard below 0.5, suggesting potential limitations in identifying prognosis-associated cell types in real-world data. In this section, ReCIDE exhibited superior performance in consistently identifying PR cell types across cohorts compared to other methods, followed by BayesPrism and Bisque. These three deconvolution methods proved to be the most valuable for conducting deconvolution-based prognostic analysis. Comprehensive performance across three evaluation scenarios We evaluated six deconvolution methods across three scenarios—consistency of differentially proportioned (DP) cell types with scRNA-seq (Scenario 1), reproducibility across bulk cohorts (Scenario 2), and identification of prognosis-related (PR) cell types (Scenario 3)—using F1-score and Jaccard under conservative and permissive criteria. Detailed results for each scenario are presented in prior subsections. Here, we synthesize overall performance across all eight metrics ( Fig. 4 D, E ) . ReCIDE emerged as the top performer, achieving an average ranking of 1.79 across all test datasets and metrics, reflecting its robust ability to detect and reproduce DP and PR cell types. Bisque followed with an average ranking of 2.78, demonstrating strong performance particularly in DP-focused tasks, while BayesPrism ranked third at 3.04, notable for PR cell type identification (Fig. 4 E). These rankings aggregate method consistency, reproducibility, and prognostic relevance, aligning with our benchmark’s emphasis on biologically meaningful cell types. Therefore, for users in need of cancer data deconvolution analysis, we highly recommend using ReCIDE, followed by Bisque ad BayesPrism. Pan-cancer prognostic analysis based on deconvolution results In this section, we constructed a pan-cancer atlas by leveraging all the cancer scRNA-seq data collected in this study, which includes two cancer types (KIRC and STAD) whose cell annotations from original research did not meet the requirements for benchmark evaluation. The constructed pan-cancer atlas encompasses 1.4 million cells including 10 major cancer types: BRCA, COAD, ESCA, HNSC, KIRC, LIHC, NSCLC, PAAD, PRAD, and STAD. Among these, three cancer types contain distinct subtypes: BRCA includes TNBC and ER + BC; HNSC comprises HPV- HNSC and HPV + HNSC; and NSCLC consists of LUSC and LUAD. Based on cell type marker genes obtained from published studies 40 , 41 , 42 , 43 , 44 , this comprehensive atlas has been systematically annotated into eight major cell types and 36 distinct cell subtypes (Fig. 5 A, B; Methods ), establishing a robust foundation for pan-cancer analysis. Using ReCIDE, Bisque, and BayesPrism—top performers from our benchmark—we deconvolved 14 cohorts (10 TCGA, 4 GEO, Table 2 ) with survival data across 10 cancer types. For cancer types with distinct subtypes, we conducted independent analyses for each subtype (e.g., BRCA, HNSC, NSCLC), excluding HPV + HNSC from TCGA due to the absence of mortality events, resulting in a total of 12 distinct cancer entities analyzed. Univariate Cox Regression analysis was performed to assess the correlation between the proportions of cell subtypes and patient survival outcomes within each cancer entity, where hazard ratios (HR) were calculated, with -ln(HR) > 0 indicating favorable prognosis and -ln(HR) < 0 denoting poor prognosis (Fig. 6 A). Futhermore, we applied a one-sided Wilcoxon Signed-Rank Test with FDR correction to the median -ln(HR) values across all 12 entities to identify cell subtypes that are consistently associated with positive or negative cancer prognosis at the pan-cancer level, defining their pan-cancer significance ( Fig. 6 B ) . ReCIDE revealed matrix cancer-associated fibroblasts (mCAF) as a poor prognosis marker (p = 0.0081), showing a negative correlation with survival in ER + BC, HPV- HNSC, and KIRC ( Supplementary Fig. 2A ). BayesPrism identified cDC_CLEC9A as a favorable prognosis marker (p = 0.016), showing a positive correlation in TNBC, LIHC, LUAD, and PAAD ( Supplementary Fig. 2B ). Bisque detected no significant prognosis markers. GO enrichment analysis of genes specifically expressed in mCAF and cDC_CLEC9A populations revealed distinct functional characteristics. Beyond the general features shared across the fibroblast class, mCAF showed enrichment in pathways including 'cellular response to transforming growth factor beta stimulus' and 'cellular response to Vascular Endothelial Growth Factor (VEGF) stimulus' ( Fig. 7 A ) . The enrichment of VEGF response pathways suggests that mCAFs participate in angiogenesis, which not only supports tumor growth but also facilitates tumor spread and metastasis 43 . The GO enrichment analysis of cDC_CLEC9A-specific genes showed significant enrichment in biological processes related to 'peptide antigen assembly with MHC class II protein complex' and 'antigen processing and presentation of exogenous peptide antigen via MHC class II' ( Fig. 7 B ) , indicating its primary function involves assembling antigen-MHC complexes to execute antigen presentation 45 , 46 . Building on these findings, we explored whether combining mCAF or cDC_CLEC9A cell proportions with other cell types could generate prognostic indicators applicable across at least five TCGA cancer entities, with subsequent validation in GEO cohorts. The results demonstrated that for the deconvolution results of ReCIDE, the prognostic evaluation metric calculated by subtracting mCAF from antibody-secreting cell (ASC) proportions was identified as the optimal indicator, which correlated with a positive prognosis in five TCGA cancer entities ( Fig. 7 C ) . For BayesPrism deconvolution results, the addition of cDC_CLEC9A to either Mast_KIT or hypoxic cell proportions showed significant correlation with positive prognosis across five TCGA cancer entities ( Fig. 7 D, E ) . To further validate these findings, we tested the constructed prognostic indicators across four GEO cohorts (five cancer entities, LUAD and LUSC in one cohort). ReCIDE’s (ASC% - mCAF%) indicator exhibited significant positive correlation with prognosis across three GEO cancer entities ( Fig. 7 F ) , while BayesPrism’s (Hypoxia% + cDC_CLEC9A%) indicator showed significant prognostic relevance in one GEO cancer entity ( Fig. 7 G ) . In contrast, BayesPrism’s (Mast_KIT% + cDC_CLEC9A%) indicator failed to demonstrate any significant correlation with cancer prognosis in the GEO cohorts (Supplementary Fig. 3A) . Overall, ReCIDE’s (ASC% - mCAF%) indicator demonstrated the highest reliability for patient stratification across multiple cancer types, emphasizing its superior clinical utility in pan-cancer prognostic studies. Discussion Deconvolution methods are pivotal bioinformatics tools for dissecting the cellular composition of tumor microenvironments, widely applied in cancer research 47 , 48 . Unlike prior pseudobulk-based benchmarks, our study introduces a novel evaluation framework using real bulk RNA-seq and microarray data from 16 clinical cohorts, emphasizing biologically and clinically relevant metrics over artificial simulations. Central to our approach is the use of relative changes in differentially proportioned (DP) cell types between conditions as the evaluation standard, rather than absolute proportions prone to biological and technical biases. For example, in TCGA-COAD dataset, ReCIDE, Bisque, and DWLS consistently detected dendritic cell upregulation in tumors versus normal tissue, aligning with trends observed in scRNA-seq data despite discrepancies in absolute proportions. This directional consistency underscores the reliability of relative shifts, forming the basis for three evaluation scenarios: (1) consistency with scRNA-seq, (2) reproducibility across cohorts (internal robustness), and (3) identification of prognosis-related (PR) cell types using survival data (external utility). These scenarios, detailed in Results , progress from biological fidelity to clinical applicability, offering a comprehensive assessment of deconvolution performance. Across these scenarios, ReCIDE demonstrated consistent advantages over other methods, particularly in recall and Jaccard, making it a potentially valuable deconvolution method for minimizing false negatives in cancer studies. Bisque showed strength in precision, effectively reducing false positives, while BayesPrism provided a balanced performance across metrics. Notably, DWLS and MuSiC, which have shown strong results in previous pseudobulk benchmarks 23 , 49 , exhibited limited performance with real-world data, presumably due to their sensitivity to inherent noise. This discrepancy underscores the importance of evaluating methods using real-world data rather than relying solely on simulated benchmarks. Building on this framework, we conducted a pan-cancer prognostic analysis across 12 cancer entities using ReCIDE, Bisque, and BayesPrism. ReCIDE identified matrix cancer-associated fibroblasts (mCAF) as a consistent poor prognosis marker (p = 0.0081), while BayesPrism linked cDC_CLEC9A to favorable outcomes (p = 0.016). The most robust prognostic indicator, ReCIDE’s subtraction of mCAF from antibody-secreting cell (ASC) proportions, correlated with survival in five TCGA cohorts and validated in three GEO cohorts, demonstrating superior utility for patient stratification compared to prognostic indicators based on BayesPrism. Despite these advances, limitations persist. Cohort heterogeneity—despite matching tumor subtypes and stages—may introduce variability in DP and PR cell types, uniformly affecting all methods but potentially undermining reproducibility. Additionally, our reliance on disease grouping and survival data, due to limited clinical annotations, excludes other relevant metrics like therapy response (e.g., antibody treatments 50 , 51 , 52 , an emerging field). Future studies should address cohort variability and expand clinical dimensions to enhance generalizability. In summary, our benchmark provides a real-data-driven framework that surpasses pseudobulk limitations, rigorously evaluating six deconvolution methods. This work complements existing research and sets a foundation for refining deconvolution tools in translational cancer studies. Methods Data collection scRNA-seq references We sourced scRNA-seq datasets for 10 cancer types from TISH2 53 , CellXGene 54 , and Curated Cancer Cell Atlas databases 55 , applying these criteria: (1) datasets included > 5 samples, (2) samples spanned distinct clinical conditions (e.g., tumor vs. adjacent normal, cancer subtypes), and (3) cell type annotations were provided. If suitable datasets were unavailable, we searched PubMed using the pubmed.mine 56 R package with keywords "single cell" and the disease’s full name. Details are in Table 1 . Table 1 scRNA-seq datasets used in this research. Disease Number of samples (N: normal, T: tumor) scRNA_source_article BRCA 15ER + BC and 11TNBC Bassez A et al. 2021 28 COAD 35N and 53T Pelka et al. 2021 29 ESCA 4N and 60T Zhang X et al. 2021 30 HNSC 12HPV- HNSC and 6HPV + HNSC Kürten et al. 2021 31 KIRC 10N and 12T Mei S et al.2022 32 LIHC 9N and 14T Sharma et al. 2020 33 NSCLC 18LUAD and 16LUSC Salcher S et al. 2022 34 PAAD 11N and 24T Peng et al. 2019 35 PRAD 13N and 18T Hirz T et al.2023 36 STAD 9N and 23T Kumar V et al.2022 37 Bulk RNA-seq and microarray cohorts Bulk RNA-expression cohorts from TCGA and GEO were selected based on: (1) histopathological alignment with scRNA-seq; (2) primary tumor origin; and (3) preference for cohorts with the largest sample size. TCGA expression matrices were retrieved via TCGAbiolinks 57 v2.30.4, with survival data from UCSC Xena ( https://xenabrowser.net/ ). GEO matrices were obtained using GEOquery 58 v2.70.0. See Table 2 for details. Reference selection for deconvolution To ensure reliable deconvolution, we matched scRNA-seq references to bulk data by condition. For cancerous tissues, scRNA-seq from tumor samples served as the reference; for non-cancerous tissues, scRNA-seq from normal samples served as the reference. Similarly, for cancer subtypes (e.g., HPV + vs. HPV- HNSC), subtype-specific scRNA-seq references deconvolved corresponding bulk cohorts. Evaluation metrics We assessed deconvolution performance across three scenarios using metrics detailed below, with formulas in Supplementary Note 1 . Scenario 1: Consistency with scRNA-seq Conservative and permissive F1-scores evaluated consistency between deconvolution results and scRNA-seq-defined differentially proportioned (DP) cell types. Recall and precision were calculated, with F1-scores as their weighted harmonic mean. Recall and conservative precision required statistical significance in scRNA-seq for true positives (TP), while permissive precision additionally incorporated consistent trends. Here, "consistent trends" refer to cell types that, although not reaching statistical significance in scRNA-seq analysis, exhibited proportional changes in the same direction as the deconvolution results, with directionality determined by median cell proportions. Scenario 2: Reproducibility across cohorts Conservative and permissive Jaccard measured DP cell type overlap between paired bulk cohorts, with conservative Jaccard requiring significance in both and permissive Jaccard additionally incorporated consistent trends. Here, "consistent trends" refer to cell types that, although not reaching statistical significance in one cohort, exhibited proportional changes in the same direction as the other cohort, with directionality determined by median cell proportions. F1-scores further assessed shared DP cell types against scRNA-seq, mirroring scenario 1’s permissive approach. Scenario 3: Prognostic relevance For cohorts with survival data, conservative and permissive Jaccard evaluated consistency of prognosis-related (PR) cell types across bulk datasets, using the same significance and trend criteria as scenario 2. Here, "consistent trends" were defined as cell types that, although not reaching statistical significance in one cohort, exhibited the same directionality of prognostic correlation as observed in another cohort. Directionality was determined by whether -ln(HR) was greater than 0 (indicating favorable prognosis) or less than 0 (denoting poor prognosis). Pan-cancer atlas construction and visualization The pan-cancer atlas was built separately for five major cell types (T & NK_cells, myeloid_cells, fibroblasts, endothelial_cells, B_cells). For each major cell types, we normalized scRNA-seq data using Seurat 59 v4.4.0’s NormalizeData(), selected 2,000 variable genes with FindVariableFeatures(), and removed batch effects via IntegrateData(). Then, we performed PCA with RunPCA() (top 30 components), followed by unsupervised clustering (FindNeighbors(), FindClusters()) and UMAP visualization (RunUMAP()). Clusters were annotated using published marker genes 40 , 41 , 42 , 43 , 44 . Survival analysis Kaplan-Meier survival analysis, implemented with survival v3.5 and survminer v0.4.9 R packages, used median cell proportions or prognostic indicators as high/low-risk cutoffs. The pairwise_survdiff() and ggsurvplot() functions from the survminer v0.4.9 R package were used for plotting Kaplan-Meier plots and log-rank tests. Cox regression was conducted using coxph(). GO enrichment analysis Marker genes for mCAF and cDC_CLEC9A were identified with Seurat’s FindMarkers() (|log2FC| > 1, adj.p.val < 0.05), followed by enrichment analysis using clusterProfiler 60 v4.10.0’s enrichGO(). Statistical tests Wilcoxon tests used ggpubr v0.6.0’s stat_compare_means(), Fisher’s tests used stats v4.3.3’s fisher.test(), and hierarchical clustering used pheatmap v1.0.12’s pheatmap(). Deconvolution methods CIBERSORT CIBERSORT was obtained from https://github.com/Moonerss/CIBERSORT and utilized following the instructions provided at https://github.com/Moonerss/CIBERSORT/blob/main/README.md . The inputs included the bulk data and MGM (Marker Gene Matrix). MGM construction began by initially selecting 100 marker genes for each cell type using the cosg() function from the COSG R package 61 . Marker genes with a SecondFC (Second Fold Change) greater than 1.5 were retained 23 , and the mean expression values of marker genes for each cell type were utilized to construct the MGM. All other parameters were left at their default settings. DWLS The DWLS R package v0.1.0 was obtained from https://github.com/dtsoucas/DWLS and utilized by following the tutorials provided at https://github.com/dtsoucas/DWLS/blob/master/Manual.docx . Inputs for DWLS included the bulk data, raw counts gene expression matrix of reference scRNA-seq data, and cell type labels, with the latter being annotations from the reference scRNA-seq dataset. To enhance computational efficiency, the findmarker() function from the Seurat package, invoked by the buildSignatureMatrixUsingSeurat() function in DWLS, was substituted with the cosg() function from the COSG R package 61 . All other parameters were maintained at their default settings. BayesPrism The BayesPrism R package v2.0 was obtained from https://github.com/Danko-Lab/BayesPrism and utilized by following the tutorial available at https://github.com/Danko-Lab/BayesPrism/blob/main/tutorial_deconvolution.pdf . Inputs for BayesPrism included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and cell state labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset. In accordance with the benchmark paper by Tran et al. 24 , we opted not to utilize the cell state labels option to ensure comparability with other methods, specifying only cell types for BayesPrism. When tumor cells were present, the key parameters in the new.prism() function were configured as tumor cells, while all other parameters were set to the default options of the algorithm. BisqueRNA (Bisque) The Bisque R package v1.0.5 was obtained from https://github.com/cozygene/bisque and utilized by following the guidelines provided in the tutorials available at https://github.com/cozygene/bisque/blob/master/vignettes/bisque.Rmd . Inputs for Bisque included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were obtained from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were configured to the default settings of the algorithm. MuSiC The MuSiC R package v1.0.0 was obtained from https://github.com/xuranw/MuSiC and implemented by following the instructions provided in the tutorials at https://xuranw.github.io/MuSiC/articles/MuSiC.html . Inputs to MuSiC included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were left at their default settings. ReCIDE The ReCIDE R package v1.0.0 was obtained from https://github.com/limingham/ReCIDE and implemented by following the instructions provided in the tutorials at https://github.com/limingham/ReCIDE/blob/main/README.md . Inputs to ReCIDE included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were left at their default settings. Declarations Data available This study did not generate any new data. All scRNA-seq and bulk data used in this study are listed in Table 1 and Table 2 , and the download links for all datasets can be accessed at https://github.com/TianLab-Bioinfo/Benchmark-realdata. Code availability The R scripts used for analyses in this study is available at https://github.com/TianLab-Bioinfo/Benchmark-realdata. Funding This work was supported by the National Natural Science Foundation of China [32170667, 32370719]. Competing interests The authors declare that they have no competing interests. Author Contribution L. and S. contributed equally to this work. L. and T. designed the study. L. and S. performed the analyses and wrote the manuscript. T. supervised the project, provided critical revisions, and approved the final version. All authors reviewed and approved the manuscript. References Du W, Elemento O. Cancer systems biology: embracing complexity to develop better anticancer therapeutic strategies. Oncogene. 2015;34:3215–25. Cha YJ, Park EJ, Baik SH, Lee KY, Kang J. Clinical significance of tumor-infiltrating lymphocytes and neutrophil-to-lymphocyte ratio in patients with stage III colon cancer who underwent surgery followed by FOLFOX chemotherapy. Sci Rep. 2019;9:11617. Idos GE, Kwok J, Bonthala N, Kysh L, Gruber SB, Qu C. The Prognostic Implications of Tumor Infiltrating Lymphocytes in Colorectal Cancer: A Systematic Review and Meta-Analysis. Sci Rep. 2020;10:3360. Maibach F, Sadozai H, Seyed Jafari SM, Hunger RE, Schenk M. Tumor-Infiltrating Lymphocytes and Their Prognostic Value in Cutaneous Melanoma. Front Immunol. 2020;11:2105. Huang D, et al. Advances in single-cell RNA sequencing and its applications in cancer research. J Hematol Oncol. 2023;16:98. Aissa AF, et al. Single-cell transcriptional changes associated with drug tolerance and response to combination therapies in cancer. Nat Commun. 2021;12:1628. Tang F, et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nat Methods. 2009;6:377–82. Jovic D, Liang X, Zeng H, Lin L, Xu F, Luo Y. Single-cell RNA sequencing technologies and applications: A brief overview. Clin Transl Med. 2022;12:e694. Wang S et al. The Evolution of Single-Cell RNA Sequencing Technology and Application: Progress and Perspectives. Int J Mol Sci 24, (2023). Clough E, Barrett T. The Gene Expression Omnibus Database. Methods Mol Biol. 2016;1418:93–110. Hutter C, Zenklusen JC. The Cancer Genome Atlas: Creating Lasting Value beyond Its Data. Cell. 2018;173:283–5. Edgar R, Domrachev M, Lash AE. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 2002;30:207–10. Barrett T, et al. NCBI GEO: archive for functional genomics data sets–update. Nucleic Acids Res. 2013;41:D991–995. Im Y, Kim Y. A Comprehensive Overview of RNA Deconvolution Methods and Their Application. Mol Cells. 2023;46:99–105. Avila Cobos F, Vandesompele J, Mestdagh P, De Preter K. Computational deconvolution of transcriptomics data from mixed cell populations. Bioinformatics. 2018;34:1969–79. Newman AM, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12:453–7. Tsoucas D, Dong R, Chen H, Zhu Q, Guo G, Yuan G-C. Accurate estimation of cell-type composition from gene expression data. Nat Commun. 2019;10:2975. Wang X, Park J, Susztak K, Zhang NR, Li M. Bulk tissue cell type deconvolution with multi-subject single-cell expression reference. Nat Commun. 2019;10:380. Jew B, et al. Accurate estimation of cell composition in bulk expression through robust integration of single-cell information. Nat Commun. 2020;11:1971. Chu T, Wang Z, Pe’er D, Danko CG. Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology. Nat Cancer. 2022;3:505–17. Li M, Su Y, Gao Y, Tian W. ReCIDE: robust estimation of cell type proportions by integrating single-reference-based deconvolutions. Brief Bioinform 25, (2024). Newman AM, Gentles AJ, Liu CL, Diehn M, Alizadeh AA. Data normalization considerations for digital tumor dissection. Genome Biol. 2017;18:128. Avila Cobos F, Alquicira-Hernandez J, Powell JE, Mestdagh P, De Preter K. Benchmarking of cell type deconvolution pipelines for transcriptomics data. Nat Commun. 2020;11:5650. Tran KA, et al. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. Nat Commun. 2023;14:5758. Hu M, Chikina M. Heterogeneous pseudobulk simulation enables realistic benchmarking of cell-type deconvolution methods. Genome Biol. 2024;25:169. Garmire LX, et al. Challenges and perspectives in computational deconvolution of genomics data. Nat Methods. 2024;21:391–400. Hippen AA, et al. Performance of computational algorithms to deconvolve heterogeneous bulk ovarian tumor tissue depends on experimental factors. Genome Biol. 2023;24:239. Bassez A, et al. A single-cell map of intratumoral changes during anti-PD1 treatment of patients with breast cancer. Nat Med. 2021;27:820–32. Pelka K, et al. Spatially organized multicellular immune hubs in human colorectal cancer. Cell. 2021;184:4734–e47524720. Zhang X, et al. Dissecting esophageal squamous-cell carcinoma ecosystem by single-cell transcriptomic analysis. Nat Commun. 2021;12:5291. Kürten CHL, et al. Investigating immune and non-immune cell interactions in head and neck tumors by single-cell RNA sequencing. Nat Commun. 2021;12:7338. Mei S, et al. Single-cell analysis of immune and stroma cell remodeling in clear cell renal cell carcinoma primary tumors and bone metastatic lesions. Genome Med. 2024;16:1. Sharma A, et al. Onco-fetal Reprogramming of Endothelial Cells Drives Immunosuppressive Macrophages in Hepatocellular Carcinoma. Cell. 2020;183:377–e394321. Salcher S, et al. High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell. 2022;40:1503–e15201508. Peng J, et al. Single-cell RNA-seq highlights intra-tumoral heterogeneity and malignant progression in pancreatic ductal adenocarcinoma. Cell Res. 2019;29:725–38. Hirz T, et al. Dissecting the immune suppressive human prostate tumor microenvironment via integrated single-cell and spatial transcriptomic analyses. Nat Commun. 2023;14:663. Kumar V, et al. Single-Cell Atlas of Lineage States, Tumor Microenvironment, and Subtype-Specific Expression Programs in Gastric Cancer. Cancer Discov. 2022;12:670–91. Lee H-O, et al. Lineage-dependent gene expression programs influence the immune landscape of colorectal cancer. Nat Genet. 2020;52:594–603. Fonseca MAS, et al. Single-cell transcriptomic analysis of endometriosis. Nat Genet. 2023;55:255–67. Yang Y, et al. Pan-cancer single-cell dissection reveals phenotypically distinct B cell subtypes. Cell. 2024;187:4790–e48114722. Qian J, et al. A pan-cancer blueprint of the heterogeneous tumor microenvironment revealed by single-cell profiling. Cell Res. 2020;30:745–62. Li W, et al. A single-cell pan-cancer analysis to show the variability of tumor-infiltrating myeloid cells in immune checkpoint blockade. Nat Commun. 2024;15:6142. Ma C, et al. Pan-cancer spatially resolved single-cell analysis reveals the crosstalk between cancer-associated fibroblasts and tumor microenvironment. Mol Cancer. 2023;22:170. Li J, Wang D, Tang F, Ling X, Zhang W, Zhang Z. Pan-cancer integrative analyses dissect the remodeling of endothelial cells in human cancers. Natl Sci Rev. 2024;11:nwae231. Murphy TL, Murphy KM. Dendritic cells in cancer immunology. Cell Mol Immunol. 2022;19:3–13. Liu P, Zhao L, Kroemer G, Kepp O. Conventional type 1 dendritic cells (cDC1) in cancer immunity. Biol Direct. 2023;18:71. Schelker M, et al. Estimation of immune cell content in tumour tissue using single-cell RNA-seq data. Nat Commun. 2017;8:2032. Tirosh I, et al. Dissecting the multicellular ecosystem of metastatic melanoma by single-cell RNA-seq. Science. 2016;352:189–96. Nguyen H, Nguyen H, Tran D, Draghici S, Nguyen T. Fourteen years of cellular deconvolution: methodology, applications, technical evaluation and outstanding challenges. Nucleic Acids Res. 2024;52:4761–83. Zhang Y, et al. Single-cell analyses reveal key immune cell subsets associated with response to PD-L1 blockade in triple-negative breast cancer. Cancer Cell. 2021;39:1578–e15931578. Hargadon KM, Johnson CE, Williams CJ. Immune checkpoint blockade therapy for cancer: An overview of FDA-approved immune checkpoint inhibitors. Int Immunopharmacol. 2018;62:29–39. Yin S, et al. Patient-derived tumor-like cell clusters for personalized chemo- and immunotherapies in non-small cell lung cancer. Cell Stem Cell. 2024;31:717–e733718. Han Y, et al. TISCH2: expanded datasets and new tools for single-cell transcriptome analyses of the tumor microenvironment. Nucleic Acids Res. 2022;51:D1425–31. Abdulla S, et al. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Res. 2025;53:D886–900. Tyler M et al. The Curated Cancer Cell Atlas: comprehensive characterisation of tumours at single-cell resolution. bioRxiv , 2024.2010.2011.617836 (2024). Rani J, Shah AR, Ramachandran S. pubmed.mineR: An R package with text-mining algorithms to analyse PubMed abstracts. J Biosci. 2015;40:671–82. Colaprico A, et al. TCGAbiolinks: an R/Bioconductor package for integrative analysis of TCGA data. Nucleic Acids Res. 2015;44:e71–71. Davis S, Meltzer PS. GEOquery: a bridge between the Gene Expression Omnibus (GEO) and BioConductor. Bioinformatics. 2007;23:1846–7. Hao Y, et al. Integrated analysis of multimodal single-cell data. Cell. 2021;184:3573–e35873529. Yu G, Wang L-G, Han Y, He Q-Y. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. OMICS. 2012;16:284–7. Dai M, Pei X, Wang XJ. Accurate and fast cell marker gene identification with COSG. Brief Bioinform 23, (2022). Wu SZ, et al. A single-cell and spatially resolved atlas of human breast cancers. Nat Genet. 2021;53:1334–47. Marisa L, et al. Gene expression classification of colon cancer into molecular subtypes: characterization, validation, and prognostic value. PLoS Med. 2013;10:e1001453. Cao W, et al. Multi-faceted epigenetic dysregulation of gene expression promotes esophageal squamous cell carcinoma. Nat Commun. 2020;11:3675. Wichmann G, et al. The role of HPV RNA transcription, immune response-related gene expression and disruptive TP53 mutations in diagnostic and prognostic profiling of head and neck cancer. Int J Cancer. 2015;137:2846–57. von Roemeling CA, et al. Neuronal pentraxin 2 supports clear cell renal cell carcinoma by activating the AMPA-selective glutamate receptor-4. Cancer Res. 2014;74:4796–810. Woo HG, et al. Integrative analysis of genomic and epigenomic regulation of the transcriptome in liver cancer. Nat Commun. 2017;8:839. Hight SK, et al. An in vivo functional genomics screen of nuclear receptors and their co-regulators identifies FOXA1 as an essential gene in lung tumorigenesis. Neoplasia. 2020;22:294–310. Tang H, et al. A 12-gene set predicts survival benefits from adjuvant chemotherapy in non-small cell lung cancer patients. Clin Cancer Res. 2013;19:1577–86. Moffitt RA, et al. Virtual microdissection identifies distinct tumor- and stroma-specific subtypes of pancreatic ductal adenocarcinoma. Nat Genet. 2015;47:1168–78. Penney KL, et al. Association of prostate cancer risk variants with gene expression in normal and tumor tissue. Cancer Epidemiol Biomarkers Prev. 2015;24:255–60. Labbé DP, et al. High-fat diet fuels prostate cancer progression by rewiring the metabolome and amplifying the MYC program. Nat Commun. 2019;10:4358. Oh SC, et al. Clinical and genomic landscape of gastric cancer with a mesenchymal phenotype. Nat Commun. 2018;9:1777. Additional Declarations No competing interests reported. Supplementary Files Supplementaryinformation.docx Cite Share Download PDF Status: Published Journal Publication published 21 Jan, 2026 Read the published version in Genome Biology → Version 1 posted Editorial decision: Revision requested 18 Jun, 2025 Reviews received at journal 13 Jun, 2025 Reviews received at journal 20 May, 2025 Reviewers agreed at journal 08 May, 2025 Reviewers agreed at journal 06 May, 2025 Reviewers invited by journal 06 May, 2025 Editor assigned by journal 10 Apr, 2025 Submission checks completed at journal 07 Apr, 2025 First submitted to journal 07 Apr, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6389993","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":454539940,"identity":"edbbe5dd-af4a-40e9-92db-a5313eee4b37","order_by":0,"name":"Minghan Li","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Minghan","middleName":"","lastName":"Li","suffix":""},{"id":454539941,"identity":"7983a699-b3ef-4c3b-916e-144c4491c2ab","order_by":1,"name":"Yuqing Su","email":"","orcid":"","institution":"Fudan University","correspondingAuthor":false,"prefix":"","firstName":"Yuqing","middleName":"","lastName":"Su","suffix":""},{"id":454539942,"identity":"7a62a8da-be90-4aa4-b64b-44727cab5d0b","order_by":2,"name":"Weidong Tian","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCUlEQVRIiWNgGAWjYFACxgYJMM0MxB9sgFwQh4cYLTxALYwz0ojSwsAgAVPEzEOMFoPjzY03Pu6oZbBnZ3722CbhsGz/7AbGB2/bGOTNcWk5c7DZcuaZ40CHsZkb5yQcNp5x5wCz4dw2BsOdDdi1mN1IbJPmbTsG8ouZdO6Pw4kNNxLYgCIMCQYHcGi5/xCmhf2btEXC4cT5NxLYf+PVcoMRpKUGqIXHTJoBqGUD0BZmfFrszyQC/dJ2gIfnME+ZZE9CuvHGG4nNknPOSRhuwKFFsv34wxsf2+rk2PuPb5P4kWAtO+9G8sEPb8ps5HHZAgWHkSMCHDUSeNUDQR0hBaNgFIyCUTCSAQB+V1pd1d7WDgAAAABJRU5ErkJggg==","orcid":"","institution":"Fudan University","correspondingAuthor":true,"prefix":"","firstName":"Weidong","middleName":"","lastName":"Tian","suffix":""}],"badges":[],"createdAt":"2025-04-07 04:23:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6389993/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6389993/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s13059-026-03942-1","type":"published","date":"2026-01-21T15:58:30+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":82506849,"identity":"95b9d5fe-9aee-45fb-9e34-eb558eeeb1ed","added_by":"auto","created_at":"2025-05-12 09:47:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":10991273,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic representation of the benchmarking study and the experimental design of our research. (A) Summary of datasets, deconvolution methods, and evaluation metrics involved in this study. (B-E) Illustration of the three evaluation scenarios. (C) Scenario 1 assesses consistency of differentially proportioned (DP) cell types between deconvolution results and scRNA-seq data. (D) Scenario 2 evaluates reproducibility of DP cell types across TCGA and GEO cohorts. (E) Scenario 3 examines reproducibility of prognosis-related (PR) cell types across TCGA and GEO cohorts.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/cad0f4170a8d820a9bf3aa7d.png"},{"id":82506850,"identity":"1ba648b2-c444-464a-9b0b-1ee092496a9a","added_by":"auto","created_at":"2025-05-12 09:47:17","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":8179392,"visible":true,"origin":"","legend":"\u003cp\u003eThe consistency between the DP cell types identified by the deconvolution methods and the scRNA-seq standard. (A) Detailed description of the calculation for conservative and permissive F1-score. (B) Detailed representation of conservative and permissive F1-score within the TCGA-COAD dataset. (C-D) Results of the consistency between deconvolution results and scRNA-seq standard across eight cancer types for six methods. For PRAD, both recall and F1-scores are 0 due to the absence of scRNA-seq DP cell type. (C) Conservative F1-scores. (D) Permissive F1-scores.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/85cb5b7a19e082ccc2100b7f.png"},{"id":82506851,"identity":"1590d307-47f2-4e22-8202-d370f3e3d449","added_by":"auto","created_at":"2025-05-12 09:47:17","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":9454001,"visible":true,"origin":"","legend":"\u003cp\u003eReproducibility of DP cell types across cohorts. (A) Detailed description of the calculation for conservative and permissive Jaccard. (B) Detailed representation of conservative and permissive Jaccard for six methods in COAD (TCGA and GEO cohorts). (C-F) Results of the reproducibility and consistency across eight cancer types for six methods. (C) Conservative Jaccard. (D) Conservative F1-scores of shared DP cell types vs. scRNA-seq. (E) Permissive Jaccard. (F) Permissive F1-scores of shared DP cell types vs. scRNA-seq.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/ee91c20f694aca1f9cb48df8.png"},{"id":82506854,"identity":"c24ce1d0-d88f-4861-a087-b687766ef7e3","added_by":"auto","created_at":"2025-05-12 09:47:17","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":7951836,"visible":true,"origin":"","legend":"\u003cp\u003eReproducibility of PR cell types and overall method performance. (A) Detailed representation of conservative and permissive Jaccard for six methods in COAD (TCGA and GEO cohorts). (B-C) Results of the reproducibility across eight cancer types for six methods. (B) Conservative Jaccard. (C) Permissive Jaccard. (D) Detailed rankings of the six deconvolution methods across all scenarios and metrics. (E) Average rankings of the 6 deconvolution methods across all scenarios.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/8731b0e13a7b1bac8dbc3e53.png"},{"id":82507641,"identity":"209fed2b-8e59-4604-a893-0fe911bd6026","added_by":"auto","created_at":"2025-05-12 09:55:17","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":11851725,"visible":true,"origin":"","legend":"\u003cp\u003ePan-cancer cell atlas composition. (A) UMAP visualization of the atlas, featuring eight major cell types and 36 subtypes (nine T \u0026amp; NK_cells, eight myeloid_cells, six fibroblasts, five endothelial_cells, five B_cells, three others). (B) The composition of cell subtypes within each major cell type, with marker genes listed for annotation.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/90d8cec53d775919e297e7e3.png"},{"id":82506866,"identity":"b1565c36-691b-4491-bc70-01640e7e4e7a","added_by":"auto","created_at":"2025-05-12 09:47:18","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":9325007,"visible":true,"origin":"","legend":"\u003cp\u003ePan-cancer prognostic analysis results. (A) Heatmap of prognostic correlations between 33 cell subtypes and survival across 12 cancer entities. (B) Boxplot of -ln(HR) values across 12 cancer entities, with significant pan-cancer prognostic cell types (Wilcoxon test, FDR-corrected) highlighted.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/e02bb68c5f3751fd1b1dcfcd.png"},{"id":82507640,"identity":"997b07fb-ffef-4636-9e34-57e5e5943788","added_by":"auto","created_at":"2025-05-12 09:55:17","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":11123971,"visible":true,"origin":"","legend":"\u003cp\u003eGO enrichment and prognostic indicators performance. (A) GO enrichment results of mCAFs marker genes. (B) GO enrichment results of cDC_CLEC9A marker genes. (C-E) Kaplan-Meier survival curves in TCGA cohorts for prognostic indicators. (C) ReCIDE’s (ASC% - mCAF%) across five TCGA entities. (D) BayesPrism’s (Mast_KIT% + cDC_CLEC9A%) across five TCGA entities. (E) BayesPrism’s (Hypoxia% + cDC_CLEC9A%) across five TCGA entities. (F) Kaplan-Meier curves for ReCIDE’s (ASC% - mCAF%) in three GEO entities (G) Kaplan-Meier curves for BayesPrism’s (Hypoxia% + cDC_CLEC9A%) in GEO-HPV- HNSC.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/9a270ed3dda400b0b4632e04.png"},{"id":101152138,"identity":"58566b15-ae3e-4d7a-939f-4f6fb84a66dc","added_by":"auto","created_at":"2026-01-26 16:10:32","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":70013200,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/e3c1949c-d5a6-4b45-aac1-a05f01330367.pdf"},{"id":82506856,"identity":"1f6fe290-1a90-449b-84e7-249ed78866a1","added_by":"auto","created_at":"2025-05-12 09:47:17","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":3237307,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryinformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-6389993/v1/96da70d3cebedf58e0d40678.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Evaluating Deconvolution Methods Using Real Bulk RNA-expression Data for Robust Prognostic Insights in Pan-Cancer","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe cellular heterogeneity of cancer critically impacts patient outcomes and treatment strategies, yet unraveling this complexity in clinical samples remains a difficult task\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Single-cell RNA sequencing (scRNA-seq) has begun to pierce this veil, revealing gene expression at unprecedent resolution\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e, but its cost restricts large-scale studies linking cell types to patient prognosis\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Conversly, bulk RNA-seq and microarray data from repositories like TCGA and GEO\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, rich with clinical information, offer a scalable alternative. Deconvolution methods bridge these domains, estimating cell type proportions from bulk data using scRNA-seq references to probe disease associations\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eYet, a notable challenge persists. Despite advances in methods like CIBERSORT\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, DWLS\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, MuSiC\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, Bisque\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, BayesPrism\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e, and our recently developed ReCIDE\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, current benchmarking studies for deconvolution methods invariably lean on pseudobulk data or flow cytometry\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, assuming known absolute cell type proportions. In real bulk RNA-expression deconvolution, such precision is a mirage, confounded by biological and technical noise, underminging deconvolution\u0026rsquo;s clinical utility\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. To address this, we propose a novel benchmark strategy: evaluating methods based on relative changes in differentially proportioned (DP) cell types\u0026mdash;those shifting significantly between conditions\u0026mdash;a metric robust across datasets and aligned with clinical research priorities.\u003c/p\u003e \u003cp\u003eHere, we assessed six deconvolution methods across 16 real bulk RNA-expression cohorts from eight cancer types, totalling 4,576 samples using three scenarios: (1) DP cell type alignment with scRNA-seq, (2) reproducibility of DP cell types between cohorts, and (3) reproducibility of prognositic related (PR) cell types between cohorts. Our analysis revealed that ReCIDE, Bisque, and BayesPrism exhibit superior performance across these benchmarks. Furthermore, pan-cancer analyses using these three methods identified matrix cancer-associated fibroblasts (mCAF) and CLEC9A\u0026thinsp;+\u0026thinsp;dendritic cells (cDC_CLEC9A) as key survival predictors, unlocking actionable insights. Leveraging these findings, we constructed three prognostic prediction indicators utilizing TCGA cohorts. Among them, the (ASC% - mCAF%) indicator, derived from ReCIDE deconvolution results, demonstrated robust clinical relevance through rigorous validation across three independent GEO cohorts. This real-data approach not only refines method selection but also advances precision oncology by linking cellular insights to patient outcomes.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eBenchmark design\u003c/h2\u003e \u003cp\u003eTraditional benchmarks of cell type deconvolution methods rely on pseudo-bulk expression data, generated by artificially mixing scRNA-seq profiles with predefined proportions and added noise. Yet, evidence mounts that these fail to mirror real bulk RNA-seq or microarray profiles\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e, where absolute cell proportions remain elusive. As an alternative, we designed a novel benchmark that emphasizes the identification of differentially proportioned (DP) cell types that exhibit significant relative proportional changes across disease conditions, rather than striving for precise absolute quantification \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eA-E\u003cb\u003e)\u003c/b\u003e. This approach aligns with the priorities of clinical research, as investigators are predominantly focused on cell types that have significant implications for disease progression and clinical outcomes.\u003c/p\u003e \u003cp\u003eInitially, we assembled datasets from 10 cancer types, including Breast Cancer (BRCA)\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e, Colon Adenocarcinoma (COAD)\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e, Esophageal Carcinoma (ESCA)\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, Head and Neck Squamous Cell Carcinoma (HNSC)\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, Kidney Renal Clear Cell Carcinoma (KIRC)\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, Liver Hepatocellular Carcinoma (LIHC)\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e, Non-Small Cell Lung Cancer (NSCLC)\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e, Pancreatic Adenocarcinoma (PAAD)\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e, Prostate Adenocarcinoma (PRAD)\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e, and Stomach Adenocarcinoma (STAD)\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e, each contrasting two conditions (cancer vs. normal in seven, subtypes in three), paired with two matching bulk cohorts from GEO and TCGA. When compare the cellular composition in subdivided cancer subtypes, we analyze each cell type individually, treating tumor cells and epithelial cells as two distinct cell types. When evaluating the cellular composition in tumor and normal tissues, we adopted the cell type proportion comparison strategy proposed by Lee et al. to ensure objective benchmarking\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. This strategy involved combining tumor cells and epithelial cells in tumor tissues and comparing them with epithelial cells in normal tissues. Due to the lack of explicit annotation of epithelial cells in KIRC and STAD, we prudently excluded these two cancer types from the benchmark evaluation. Therefore, in the benchmark, we included a total of eight cancer types (ten RNA-seq cohorts, six microarray cohorts; 4,576 samples; Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e2\u003c/span\u003e). In selecting deconvolution methods for our benchmark, we referred to the benchmark tests conducted by Cobos et al.\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e and Chu et al.\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e Based on their benchmark results, we selected five high-performenced deconvolution methods, including CIBERSORT, DWLS, MuSiC, Bisque, BayesPrism, and incorporated our newly developed method ReCIDE for comprehensive benchmarking across diverse scenarios \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eA, B\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eRNA-seq/Microarray datasets used in this research.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDisease\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTCGA_Bulk\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGEO_Bulk\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGEO_number\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGEO_type\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBRCA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e726ER\u0026thinsp;+\u0026thinsp;BC and 175TNBC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11ER\u0026thinsp;+\u0026thinsp;BC and 9TNBC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE176078\u003csup\u003e62\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRNA-seq\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCOAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e41N and 465T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e19N and 566T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e*GSE39582\u003csup\u003e63\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eESCA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13N and 173T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10N and 10T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE149609\u003csup\u003e64\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eRNA-seq\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHNSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e65 HPV- HNSC and 21 HPV\u0026thinsp;+\u0026thinsp;HNSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e121 HPV- HNSC and 60 HPV\u0026thinsp;+\u0026thinsp;HNSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e*GSE65858\u003csup\u003e65\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKIRC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e72N and 532T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e72N and 72T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE53757\u003csup\u003e66\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLIHC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e50N and 369T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30N and 64T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE87630\u003csup\u003e67\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNSCLC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e58LUAD and 51LUSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e133LUAD and 43LUSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e*GSE42127\u003csup\u003e68, 69\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePAAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4N and 142T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e46N and 145T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e*GSE71729\u003csup\u003e70\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePRAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e51N and 483T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e160N and 264T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE62872\u003csup\u003e71, 72\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSTAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13N and 97T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100N and 300T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eGSE66229\u003csup\u003e73\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eMicroarray\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e* represents GEO cohorts with survival information. All TCGA cohorts have survival information.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eWe defined DP cell types from scRNA-seq as the reference, then evaluated each method across three innovative​ scenarios: (1) DP cell types alignment with scRNA-seq \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eC\u003cb\u003e)\u003c/b\u003e, (2) reproducibility of DP cell types between cohorts \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eD\u003cb\u003e)\u003c/b\u003e, and (3) reproducibility of prognosis-related (PR) cell types between cohorts \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e1\u003c/span\u003eD\u003cb\u003e)\u003c/b\u003e, leveraging survival data where available. This comprehensive workflow enhances the scope of deconvolution benchmarking by emphasizing clinical applicability and methodological robustness. Details on data download and DP/PR cell type identification are provided in the \u003cb\u003eMethods\u003c/b\u003e an \u003cb\u003eData available\u003c/b\u003e section.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eConsistency of DP cell types between deconvolved bulk and scRNA-seq data\u003c/h3\u003e\n\u003cp\u003eTo evaluate how well deconvolution methods identify differentially proportioned (DP) cell types in real bulk expression data, we compared their results to DP cell types defined by scRNA-seq across six methods. Specifically, we employed the F1-score (Weighted harmonic mean of precision and recall) as our primary metric, using a dual-evaluation strategy\u0026mdash;conservative and permissive criteria\u0026mdash;to account for sample size disparities between scRNA-seq and bulk cohorts \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eA\u003cb\u003e)\u003c/b\u003e. Under conservative criteria, precision reflects the fraction of DP cell types from deconvolution that match statistically significant with scRNA-seq findings, while permissive criteria include those cell types that, although not statistically significant in scRNA-seq data, exhibit trends consistent with the deconvolution results. Recall, under both, measures the proportion of scRNA-seq DP cell types recovered by deconvolution. The calculations for recall, precision, and F1-score are detailed in the \u003cb\u003eMethods\u003c/b\u003e section.\u003c/p\u003e \u003cp\u003eWe first analyzed Colon Adenocarcinoma (COAD), leveraging its large scRNA-seq dataset (35 normal, 53 tumor samples; 19 cell types). Here, 15 DP cell types were identified from scRNA-seq dataset via Wilcoxon test with FDR correction, with eight enriched and seven reduced in tumors \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eB\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eWhile absolute proportions from deconvolution diverged from scRNA-seq (e.g., dendritic cells absolute proportion in TCGA-COAD: scRNA-seq median 0.014 tumor vs. 0.005 normal; ReCIDE 0.009 vs. 0.005; Bisque 0.008 vs. 0; DWLS 0.002 vs. 0, \u003cb\u003eSupplementary Fig.\u0026nbsp;1\u003c/b\u003e), the relative trends generated by ReCIDE, Bisque, and DWLS were consistent with those from scRNA-seq.\u0026nbsp;These three methods consistently detected significant upregulation of dendritic cells in tumors, mirroring scRNA-seq findings. This reinforces our DP-focused benchmark: despite absolute discrepancies, well-performing methods capture proportional shifts accurately. In COAD, under conservative criteria, ReCIDE led with an F1-score of 0.67 (9/12 DP cell types validated, recall 0.6, precision 0.75), followed by Bisque at 0.62 (8/11 validated, recall 0.53, precision 0.73). MuSiC lagged at 0.24 (3/10 validated, recall 0.2, precision 0.3). Under permissive criteria, the F1-scores of ReCIDE and MuSiC remained unchanged, as none of the DP cell types identified by these two methods were newly validated. In contrast, among the 11 DP cell types identified by Bisque, two cell types were newly validated that aligned with scRNA-seq trends, leading to an improvement in precision to 0.91, and an increase in F1-score to 0.67. Notably, MuSiC, CIBERSORT, and Bisque estimated a proportion of 0 for certain cell types in some samples. For example, in the TCGA-COAD dataset, MuSiC assigned a dendritic cell proportion as 0 for 91% of the samples, which hindered the effective identification of the distribution trends of dendritic cells between tumor colon tissues and normal colon tissues in the deconvolution results.\u003c/p\u003e \u003cp\u003eExtending to a pan-cancer analysis across eight cancer types, ReCIDE and Bisque maintained superior consistency with scRNA-seq DP cell types. ReCIDE achieved the highest average recall (0.6) under both conservative and permissive criteria, while Bisque excelled in average precision (0.35 conservative, 0.74 permissive) versus ReCIDE (0.31, 0.69) (\u003cb\u003eSupplementary Fig.\u0026nbsp;2\u003c/b\u003e). Average F1-scores underscored their robust performance: ReCIDE at 0.34 (conservative) and 0.60 (permissive), Bisque at 0.32 and 0.53 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eC, D). In contrast, DWLS and MuSiC scored below 0.3, reflecting limitations in handling heterogeneous data.\u003c/p\u003e \u003cp\u003eIt is worth noting that the average F1-scores of most deconvolution methods under permissive conditions are higher than that under conservative conditions, primarily attributable to the fact that the DP cell types identified by the deconvolution methods are more easily validated under permissive conditions, thus leading to a general increase in precision and F1-score.\u003c/p\u003e \u003cp\u003eIn summary, ReCIDE and Bisque are the top performers in this benchmark scenario. ReCIDE\u0026rsquo;s high recall suits applications like cancer screening, ensuring detection of all relevant DP cell types, while Bisque\u0026rsquo;s precision advantages favor precise subtyping.\u003c/p\u003e\n\u003ch3\u003eReproducibility of DP cell types across deconvolved bulk cohorts\u003c/h3\u003e\n\u003cp\u003eEnsuring deconvolution methods reproducibly identify differentially proportioned (DP) cell types across bulk cohorts with matching disease conditions is vital for real-world applicability, as shown by Fonseca et al.\u0026rsquo;s validation of ciliated endometrial epithelial cell enrichment in multiple ovarian cancer cohorts using MuSiC\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. Here, we assessed the reproducibility of DP cell types across paired bulk cohorts for six deconvolution methods.\u003c/p\u003e \u003cp\u003eWe measured reproducibility with the Jaccard index (Jaccard), calculating the overlap of DP cell types between paired cohorts, and evaluated accuracy using the F1-score, which compares shared DP cell types (those consistently identified across both cohorts) to scRNA-seq-defined DP cell types \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eA\u003cb\u003e)\u003c/b\u003e. We applied conservative and permissive criteria: conservative criteria required statistical significance in both cohorts for a shared DP cell type, while permissive criteria accepted a significant association in one cohort with a consistent trend in the other. For F1-scores, precision is the fraction of shared DP cell types matching scRNA-seq among all discovered shared DP cell types, and recall is the proportion of scRNA-seq DP cell types recovered among shared ones, balancing both metrics \u003cb\u003e(Methods)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eSimilar to the previous section, we first investigated the reproducibility of different deconvolution methods in Colon Adenocarcinoma (COAD), where 15 scRNA-seq DP cell types were identified. Under conservative criteria, ReCIDE identified 16 DP cell types across TCGA and GEO cohorts, with 12 shared (Jaccard\u0026thinsp;=\u0026thinsp;0.75) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eB\u003cb\u003e)\u003c/b\u003e. Of these, nine aligned trends with scRNA-seq, and eight also showed significant changes in scRNA-seq, yielding an F1-score of 0.62 (recall\u0026thinsp;=\u0026thinsp;0.53, precision\u0026thinsp;=\u0026thinsp;0.75). Bisque detected 11 DP cell types across TCGA and GEO cohorts, with seven shared (Jaccard\u0026thinsp;=\u0026thinsp;0.64), among which all aligned trends with scRNA-seq, and six also showed significant changes in scRNA-seq (F1-score\u0026thinsp;=\u0026thinsp;0.57, recall\u0026thinsp;=\u0026thinsp;0.40, precision\u0026thinsp;=\u0026thinsp;1.0). MuSiC identified 14 DP cell types, but only three were shared (Jaccard\u0026thinsp;=\u0026thinsp;0.21), none aligned trends with scRNA-seq (F1-score, recall, and precision\u0026thinsp;=\u0026thinsp;0). Under permissive criteria, ReCIDE added one shared cell type (Jaccard\u0026thinsp;=\u0026thinsp;0.81), but it mismatched scRNA-seq\u0026rsquo;s trend, dropping F1-score to 0.60 (recall\u0026thinsp;=\u0026thinsp;0.53, precision 0.69). Bisque added three shared cell types (Jaccard\u0026thinsp;=\u0026thinsp;0.91), among which all aligned trends with scRNA-seq, and two also showed significant changes in scRNA-seq, raising F1-score to 0.70 (recall\u0026thinsp;=\u0026thinsp;0.53, precision\u0026thinsp;=\u0026thinsp;1). MuSiC\u0026rsquo;s Jaccard rose to 0.29 with one added shared cell type, but F1-score remained 0 as none showed consistent and significant changes in scRNA-seq (recall\u0026thinsp;=\u0026thinsp;0, precision\u0026thinsp;=\u0026thinsp;0.25).\u003c/p\u003e \u003cp\u003eAcross eight cancer types, ReCIDE led with average Jaccard of 0.56 (conservative) and 0.88 (permissive), followed by Bisque (0.40 and 0.78) and BayesPrism (0.43 and 0.71) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eC, E\u003cb\u003e)\u003c/b\u003e. Average F1-scores for shared DP cell types vs. scRNA-seq ranked ReCIDE (0.67 conservative, 0.70 permissive), Bisque (0.46 and 0.68), and BayesPrism (0.38 and 0.48) highest \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eD, F\u003cb\u003e)\u003c/b\u003e. Compared to the conservative criteria, the gap in average F1-scores between ReCIDE and Bisque narrowed under permissive criteria. This shift occurred because, under conservative criteria, Bisque identified only 3.88 shared DP cell types across cohorts compared to ReCIDE\u0026rsquo;s 9.88 \u003cb\u003e(Supplementary Fig.\u0026nbsp;3A)\u003c/b\u003e, which resulted in Bisque achieving a lower average recall of 0.34 compared to ReCIDE\u0026rsquo;s 0.62 \u003cb\u003e(Supplementary Fig.\u0026nbsp;4A)\u003c/b\u003e, impacting its average F1-score. Under permissive criteria, the number of shared DP cell types identified by Bisque increased to 8.5 (ReCIDE: 11.88, \u003cb\u003eSupplementary Fig.\u0026nbsp;3B\u003c/b\u003e), enhancing Bisque\u0026rsquo;s average recall to 0.54 (ReCIDE: 0.66, \u003cb\u003eSupplementary Fig.\u0026nbsp;4B\u003c/b\u003e) and reducing the gap with ReCIDE from 82\u0026ndash;22%, leading to a narrowing of the difference in their average F1-scores.\u003c/p\u003e \u003cp\u003eAnother crucial observation is that shared DP cell types identified by most deconvolution methods under permissive criteria exhibit higher average F1-scores compared to those under conservative criteria \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eF\u003cb\u003e)\u003c/b\u003e. This outcome primarily stems from the fact that, under permissive criteria, the number of shared DP cell types increased compared to conservative criteria, leading to higher recall \u003cb\u003e(Supplementary Fig.\u0026nbsp;4C)\u003c/b\u003e. At the same time, precision remained stable \u003cb\u003e(Supplementary Fig.\u0026nbsp;4D)\u003c/b\u003e. As a result, F1-scores showed improvement. In addition, permissive Jaccard mirrored F1-score ranking \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eE,F\u003cb\u003e)\u003c/b\u003e, offers a robust alternative metric when scRNA-seq validation is unavailable.\u003c/p\u003e \u003cp\u003eBased on the comprehensive evaluation of both Jaccard and F1-scores, ReCIDE and Bisque demonstrated superior performance in terms of reproducibility and accuracy. Moreover, under permissive criteria, DP cell types shared between two cohorts demonstrated clear advantages in precision and F1-scores compared to DP cell types identified within individual cohorts \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003eD, \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003eE\u003cb\u003e)\u003c/b\u003e. This result supports the high reliability of consistent DP cell types identified using deconvolution methods across multiple cohorts with the matching disease conditions in clinical practice.\u003c/p\u003e\n\u003ch3\u003eReproducibility of prognosis-related cell types across deconvolved bulk cohorts\u003c/h3\u003e\n\u003cp\u003eIdentifying prognosis-related (PR) cell types that correlate with cancer patient survival is crucial for stratifying patients and designing targeted therapies. Here, we assessed the ability of six deconvolution methods to reproducibly detect PR cell types across paired bulk RNA-expression cohorts with matching disease conditions, leveraging survival data from five cancer entities: COAD, HPV-HNSC, LUAD, LUSC, and PAAD, sourced from TCGA and GEO.\u003c/p\u003e \u003cp\u003eWe evaluated reproducibility using the Jaccard index (Jaccard), which measures the overlap of PR cell types between cohorts, under conservative and permissive criteria. Conservative criteria required statistical significance in both cohorts for a shared PR cell type, while permissive criteria accepted significance in one cohort with a consistent trend in the other (\u003cb\u003eMethods\u003c/b\u003e). In COAD (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003eA), ReCIDE identified six PR cell types across TCGA and GEO cohorts, with two shared under conservative criteria (Jaccard\u0026thinsp;=\u0026thinsp;0.33) and five under permissive (Jaccard\u0026thinsp;=\u0026thinsp;0.83). BayesPrism detected nine PR cell types, with two shared conservatively (Jaccard\u0026thinsp;=\u0026thinsp;0.22) and seven permissively (Jaccard\u0026thinsp;=\u0026thinsp;0.78). In contrast, MuSiC identified no PR cell types in either criteria (Jaccard\u0026thinsp;=\u0026thinsp;0).\u003c/p\u003e \u003cp\u003eExtending to all five cancers (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003eB, C), ReCIDE achieved the highest average Jaccard: 0.218 (conservative) and 0.740 (permissive), reflecting robust detection of reproducible PR cell types. BayesPrism followed with 0.064 and 0.634, while Bisque scored 0 (conservative) and 0.642 (permissive), excelling only when criteria relaxed. In contrast, under permissive criteria, MuSiC demonstrated an average Jaccard below 0.5, suggesting potential limitations in identifying prognosis-associated cell types in real-world data.\u003c/p\u003e \u003cp\u003eIn this section, ReCIDE exhibited superior performance in consistently identifying PR cell types across cohorts compared to other methods, followed by BayesPrism and Bisque. These three deconvolution methods proved to be the most valuable for conducting deconvolution-based prognostic analysis.\u003c/p\u003e\n\u003ch3\u003eComprehensive performance across three evaluation scenarios\u003c/h3\u003e\n\u003cp\u003eWe evaluated six deconvolution methods across three scenarios\u0026mdash;consistency of differentially proportioned (DP) cell types with scRNA-seq (Scenario 1), reproducibility across bulk cohorts (Scenario 2), and identification of prognosis-related (PR) cell types (Scenario 3)\u0026mdash;using F1-score and Jaccard under conservative and permissive criteria. Detailed results for each scenario are presented in prior subsections. Here, we synthesize overall performance across all eight metrics \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003eD, E\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e \u003cp\u003eReCIDE emerged as the top performer, achieving an average ranking of 1.79 across all test datasets and metrics, reflecting its robust ability to detect and reproduce DP and PR cell types. Bisque followed with an average ranking of 2.78, demonstrating strong performance particularly in DP-focused tasks, while BayesPrism ranked third at 3.04, notable for PR cell type identification (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003eE). These rankings aggregate method consistency, reproducibility, and prognostic relevance, aligning with our benchmark\u0026rsquo;s emphasis on biologically meaningful cell types. Therefore, for users in need of cancer data deconvolution analysis, we highly recommend using ReCIDE, followed by Bisque ad BayesPrism.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003ePan-cancer prognostic analysis based on deconvolution results\u003c/h2\u003e \u003cp\u003eIn this section, we constructed a pan-cancer atlas by leveraging all the cancer scRNA-seq data collected in this study, which includes two cancer types (KIRC and STAD) whose cell annotations from original research did not meet the requirements for benchmark evaluation. The constructed pan-cancer atlas encompasses 1.4\u0026nbsp;million cells including 10 major cancer types: BRCA, COAD, ESCA, HNSC, KIRC, LIHC, NSCLC, PAAD, PRAD, and STAD. Among these, three cancer types contain distinct subtypes: BRCA includes TNBC and ER\u0026thinsp;+\u0026thinsp;BC; HNSC comprises HPV- HNSC and HPV\u0026thinsp;+\u0026thinsp;HNSC; and NSCLC consists of LUSC and LUAD. Based on cell type marker genes obtained from published studies\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e, this comprehensive atlas has been systematically annotated into eight major cell types and 36 distinct cell subtypes (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003eA, B; \u003cb\u003eMethods\u003c/b\u003e), establishing a robust foundation for pan-cancer analysis.\u003c/p\u003e \u003cp\u003eUsing ReCIDE, Bisque, and BayesPrism\u0026mdash;top performers from our benchmark\u0026mdash;we deconvolved 14 cohorts (10 TCGA, 4 GEO, Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e2\u003c/span\u003e) with survival data across 10 cancer types. For cancer types with distinct subtypes, we conducted independent analyses for each subtype (e.g., BRCA, HNSC, NSCLC), excluding HPV\u0026thinsp;+\u0026thinsp;HNSC from TCGA due to the absence of mortality events, resulting in a total of 12 distinct cancer entities analyzed. Univariate Cox Regression analysis was performed to assess the correlation between the proportions of cell subtypes and patient survival outcomes within each cancer entity, where hazard ratios (HR) were calculated, with -ln(HR)\u0026thinsp;\u0026gt;\u0026thinsp;0 indicating favorable prognosis and -ln(HR)\u0026thinsp;\u0026lt;\u0026thinsp;0 denoting poor prognosis (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e6\u003c/span\u003eA). Futhermore, we applied a one-sided Wilcoxon Signed-Rank Test with FDR correction to the median -ln(HR) values across all 12 entities to identify cell subtypes that are consistently associated with positive or negative cancer prognosis at the pan-cancer level, defining their pan-cancer significance \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e6\u003c/span\u003eB\u003cb\u003e)\u003c/b\u003e. ReCIDE revealed matrix cancer-associated fibroblasts (mCAF) as a poor prognosis marker (p\u0026thinsp;=\u0026thinsp;0.0081), showing a negative correlation with survival in ER\u0026thinsp;+\u0026thinsp;BC, HPV- HNSC, and KIRC (\u003cb\u003eSupplementary Fig.\u0026nbsp;2A\u003c/b\u003e). BayesPrism identified cDC_CLEC9A as a favorable prognosis marker (p\u0026thinsp;=\u0026thinsp;0.016), showing a positive correlation in TNBC, LIHC, LUAD, and PAAD (\u003cb\u003eSupplementary Fig.\u0026nbsp;2B\u003c/b\u003e). Bisque detected no significant prognosis markers.\u003c/p\u003e \u003cp\u003eGO enrichment analysis of genes specifically expressed in mCAF and cDC_CLEC9A populations revealed distinct functional characteristics. Beyond the general features shared across the fibroblast class, mCAF showed enrichment in pathways including 'cellular response to transforming growth factor beta stimulus' and 'cellular response to Vascular Endothelial Growth Factor (VEGF) stimulus' \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eA\u003cb\u003e)\u003c/b\u003e. The enrichment of VEGF response pathways suggests that mCAFs participate in angiogenesis, which not only supports tumor growth but also facilitates tumor spread and metastasis\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. The GO enrichment analysis of cDC_CLEC9A-specific genes showed significant enrichment in biological processes related to 'peptide antigen assembly with MHC class II protein complex' and 'antigen processing and presentation of exogenous peptide antigen via MHC class II' \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eB\u003cb\u003e)\u003c/b\u003e, indicating its primary function involves assembling antigen-MHC complexes to execute antigen presentation\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eBuilding on these findings, we explored whether combining mCAF or cDC_CLEC9A cell proportions with other cell types could generate prognostic indicators applicable across at least five TCGA cancer entities, with subsequent validation in GEO cohorts. The results demonstrated that for the deconvolution results of ReCIDE, the prognostic evaluation metric calculated by subtracting mCAF from antibody-secreting cell (ASC) proportions was identified as the optimal indicator, which correlated with a positive prognosis in five TCGA cancer entities \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eC\u003cb\u003e)\u003c/b\u003e. For BayesPrism deconvolution results, the addition of cDC_CLEC9A to either Mast_KIT or hypoxic cell proportions showed significant correlation with positive prognosis across five TCGA cancer entities \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eD, E\u003cb\u003e)\u003c/b\u003e. To further validate these findings, we tested the constructed prognostic indicators across four GEO cohorts (five cancer entities, LUAD and LUSC in one cohort). ReCIDE\u0026rsquo;s (ASC% - mCAF%) indicator exhibited significant positive correlation with prognosis across three GEO cancer entities \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eF\u003cb\u003e)\u003c/b\u003e, while BayesPrism\u0026rsquo;s (Hypoxia% + cDC_CLEC9A%) indicator showed significant prognostic relevance in one GEO cancer entity \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e7\u003c/span\u003eG\u003cb\u003e)\u003c/b\u003e. In contrast, BayesPrism\u0026rsquo;s (Mast_KIT% + cDC_CLEC9A%) indicator failed to demonstrate any significant correlation with cancer prognosis in the GEO cohorts \u003cb\u003e(Supplementary Fig.\u0026nbsp;3A)\u003c/b\u003e. Overall, ReCIDE\u0026rsquo;s (ASC% - mCAF%) indicator demonstrated the highest reliability for patient stratification across multiple cancer types, emphasizing its superior clinical utility in pan-cancer prognostic studies.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eDeconvolution methods are pivotal bioinformatics tools for dissecting the cellular composition of tumor microenvironments, widely applied in cancer research\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Unlike prior pseudobulk-based benchmarks, our study introduces a novel evaluation framework using real bulk RNA-seq and microarray data from 16 clinical cohorts, emphasizing biologically and clinically relevant metrics over artificial simulations.\u003c/p\u003e \u003cp\u003eCentral to our approach is the use of relative changes in differentially proportioned (DP) cell types between conditions as the evaluation standard, rather than absolute proportions prone to biological and technical biases. For example, in TCGA-COAD dataset, ReCIDE, Bisque, and DWLS consistently detected dendritic cell upregulation in tumors versus normal tissue, aligning with trends observed in scRNA-seq data despite discrepancies in absolute proportions. This directional consistency underscores the reliability of relative shifts, forming the basis for three evaluation scenarios: (1) consistency with scRNA-seq, (2) reproducibility across cohorts (internal robustness), and (3) identification of prognosis-related (PR) cell types using survival data (external utility). These scenarios, detailed in \u003cb\u003eResults\u003c/b\u003e, progress from biological fidelity to clinical applicability, offering a comprehensive assessment of deconvolution performance.\u003c/p\u003e \u003cp\u003eAcross these scenarios, ReCIDE demonstrated consistent advantages over other methods, particularly in recall and Jaccard, making it a potentially valuable deconvolution method for minimizing false negatives in cancer studies. Bisque showed strength in precision, effectively reducing false positives, while BayesPrism provided a balanced performance across metrics. Notably, DWLS and MuSiC, which have shown strong results in previous pseudobulk benchmarks\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e, exhibited limited performance with real-world data, presumably due to their sensitivity to inherent noise. This discrepancy underscores the importance of evaluating methods using real-world data rather than relying solely on simulated benchmarks.\u003c/p\u003e \u003cp\u003eBuilding on this framework, we conducted a pan-cancer prognostic analysis across 12 cancer entities using ReCIDE, Bisque, and BayesPrism. ReCIDE identified matrix cancer-associated fibroblasts (mCAF) as a consistent poor prognosis marker (p\u0026thinsp;=\u0026thinsp;0.0081), while BayesPrism linked cDC_CLEC9A to favorable outcomes (p\u0026thinsp;=\u0026thinsp;0.016). The most robust prognostic indicator, ReCIDE\u0026rsquo;s subtraction of mCAF from antibody-secreting cell (ASC) proportions, correlated with survival in five TCGA cohorts and validated in three GEO cohorts, demonstrating superior utility for patient stratification compared to prognostic indicators based on BayesPrism.\u003c/p\u003e \u003cp\u003eDespite these advances, limitations persist. Cohort heterogeneity\u0026mdash;despite matching tumor subtypes and stages\u0026mdash;may introduce variability in DP and PR cell types, uniformly affecting all methods but potentially undermining reproducibility. Additionally, our reliance on disease grouping and survival data, due to limited clinical annotations, excludes other relevant metrics like therapy response (e.g., antibody treatments\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e, \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e, an emerging field). Future studies should address cohort variability and expand clinical dimensions to enhance generalizability.\u003c/p\u003e \u003cp\u003eIn summary, our benchmark provides a real-data-driven framework that surpasses pseudobulk limitations, rigorously evaluating six deconvolution methods. This work complements existing research and sets a foundation for refining deconvolution tools in translational cancer studies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003escRNA-seq references\u003c/h2\u003e \u003cp\u003eWe sourced scRNA-seq datasets for 10 cancer types from TISH2\u003csup\u003e53\u003c/sup\u003e, CellXGene\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e, and Curated Cancer Cell Atlas databases\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e, applying these criteria: (1) datasets included\u0026thinsp;\u0026gt;\u0026thinsp;5 samples, (2) samples spanned distinct clinical conditions (e.g., tumor vs. adjacent normal, cancer subtypes), and (3) cell type annotations were provided. If suitable datasets were unavailable, we searched PubMed using the pubmed.mine\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e R package with keywords \"single cell\" and the disease\u0026rsquo;s full name. Details are in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003escRNA-seq datasets used in this research.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDisease\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNumber of samples (N: normal, T: tumor)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003escRNA_source_article\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBRCA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e15ER\u0026thinsp;+\u0026thinsp;BC and 11TNBC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBassez A et al.\u0026nbsp;2021\u003csup\u003e28\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCOAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e35N and 53T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePelka et al. 2021\u003csup\u003e29\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eESCA\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4N and 60T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eZhang X et al. 2021\u003csup\u003e30\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHNSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12HPV- HNSC and 6HPV\u0026thinsp;+\u0026thinsp;HNSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eK\u0026uuml;rten et al. 2021\u003csup\u003e31\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKIRC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10N and 12T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMei S et al.2022\u003csup\u003e32\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLIHC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9N and 14T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSharma et al. 2020\u003csup\u003e33\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNSCLC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18LUAD and 16LUSC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSalcher S et al. 2022\u003csup\u003e34\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePAAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11N and 24T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePeng et al. 2019\u003csup\u003e35\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePRAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e13N and 18T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHirz T et al.2023\u003csup\u003e36\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSTAD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9N and 23T\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eKumar V et al.2022\u003csup\u003e37\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eBulk RNA-seq and microarray cohorts\u003c/h2\u003e \u003cp\u003eBulk RNA-expression cohorts from TCGA and GEO were selected based on: (1) histopathological alignment with scRNA-seq; (2) primary tumor origin; and (3) preference for cohorts with the largest sample size. TCGA expression matrices were retrieved via TCGAbiolinks\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e v2.30.4, with survival data from UCSC Xena (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://xenabrowser.net/\u003c/span\u003e\u003cspan address=\"https://xenabrowser.net/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). GEO matrices were obtained using GEOquery\u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e v2.70.0. See Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e2\u003c/span\u003e for details.\u003c/p\u003e \u003cp\u003e \u003cb\u003eReference selection for deconvolution\u003c/b\u003e \u003c/p\u003e \u003cp\u003eTo ensure reliable deconvolution, we matched scRNA-seq references to bulk data by condition. For cancerous tissues, scRNA-seq from tumor samples served as the reference; for non-cancerous tissues, scRNA-seq from normal samples served as the reference. Similarly, for cancer subtypes (e.g., HPV\u0026thinsp;+\u0026thinsp;vs. HPV- HNSC), subtype-specific scRNA-seq references deconvolved corresponding bulk cohorts.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation metrics\u003c/h2\u003e \u003cp\u003eWe assessed deconvolution performance across three scenarios using metrics detailed below, with formulas in \u003cb\u003eSupplementary Note 1\u003c/b\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eScenario 1: Consistency with scRNA-seq\u003c/h2\u003e \u003cp\u003eConservative and permissive F1-scores evaluated consistency between deconvolution results and scRNA-seq-defined differentially proportioned (DP) cell types. Recall and precision were calculated, with F1-scores as their weighted harmonic mean. Recall and conservative precision required statistical significance in scRNA-seq for true positives (TP), while permissive precision additionally incorporated consistent trends. Here, \"consistent trends\" refer to cell types that, although not reaching statistical significance in scRNA-seq analysis, exhibited proportional changes in the same direction as the deconvolution results, with directionality determined by median cell proportions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eScenario 2: Reproducibility across cohorts\u003c/h2\u003e \u003cp\u003eConservative and permissive Jaccard measured DP cell type overlap between paired bulk cohorts, with conservative Jaccard requiring significance in both and permissive Jaccard additionally incorporated consistent trends. Here, \"consistent trends\" refer to cell types that, although not reaching statistical significance in one cohort, exhibited proportional changes in the same direction as the other cohort, with directionality determined by median cell proportions. F1-scores further assessed shared DP cell types against scRNA-seq, mirroring scenario 1\u0026rsquo;s permissive approach.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eScenario 3: Prognostic relevance\u003c/h2\u003e \u003cp\u003eFor cohorts with survival data, conservative and permissive Jaccard evaluated consistency of prognosis-related (PR) cell types across bulk datasets, using the same significance and trend criteria as scenario 2. Here, \"consistent trends\" were defined as cell types that, although not reaching statistical significance in one cohort, exhibited the same directionality of prognostic correlation as observed in another cohort. Directionality was determined by whether -ln(HR) was greater than 0 (indicating favorable prognosis) or less than 0 (denoting poor prognosis).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003ePan-cancer atlas construction and visualization\u003c/h2\u003e \u003cp\u003eThe pan-cancer atlas was built separately for five major cell types (T \u0026amp; NK_cells, myeloid_cells, fibroblasts, endothelial_cells, B_cells). For each major cell types, we normalized scRNA-seq data using Seurat\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e v4.4.0\u0026rsquo;s NormalizeData(), selected 2,000 variable genes with FindVariableFeatures(), and removed batch effects via IntegrateData(). Then, we performed PCA with RunPCA() (top 30 components), followed by unsupervised clustering (FindNeighbors(), FindClusters()) and UMAP visualization (RunUMAP()). Clusters were annotated using published marker genes\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eSurvival analysis\u003c/h2\u003e \u003cp\u003eKaplan-Meier survival analysis, implemented with survival v3.5 and survminer v0.4.9 R packages, used median cell proportions or prognostic indicators as high/low-risk cutoffs. The pairwise_survdiff() and ggsurvplot() functions from the survminer v0.4.9 R package were used for plotting Kaplan-Meier plots and log-rank tests. Cox regression was conducted using coxph().\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eGO enrichment analysis\u003c/h2\u003e \u003cp\u003eMarker genes for mCAF and cDC_CLEC9A were identified with Seurat\u0026rsquo;s FindMarkers() (|log2FC| \u0026gt; 1, adj.p.val\u0026thinsp;\u0026lt;\u0026thinsp;0.05), followed by enrichment analysis using clusterProfiler\u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e v4.10.0\u0026rsquo;s enrichGO().\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eStatistical tests\u003c/h2\u003e \u003cp\u003eWilcoxon tests used ggpubr v0.6.0\u0026rsquo;s stat_compare_means(), Fisher\u0026rsquo;s tests used stats v4.3.3\u0026rsquo;s fisher.test(), and hierarchical clustering used pheatmap v1.0.12\u0026rsquo;s pheatmap().\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eDeconvolution methods\u003c/h2\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eCIBERSORT\u003c/h2\u003e \u003cp\u003eCIBERSORT was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Moonerss/CIBERSORT\u003c/span\u003e\u003cspan address=\"https://github.com/Moonerss/CIBERSORT\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and utilized following the instructions provided at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Moonerss/CIBERSORT/blob/main/README.md\u003c/span\u003e\u003cspan address=\"https://github.com/Moonerss/CIBERSORT/blob/main/README.md\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The inputs included the bulk data and MGM (Marker Gene Matrix). MGM construction began by initially selecting 100 marker genes for each cell type using the cosg() function from the COSG R package\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e. Marker genes with a SecondFC (Second Fold Change) greater than 1.5 were retained\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, and the mean expression values of marker genes for each cell type were utilized to construct the MGM. All other parameters were left at their default settings.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eDWLS\u003c/h2\u003e \u003cp\u003eThe DWLS R package v0.1.0 was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/dtsoucas/DWLS\u003c/span\u003e\u003cspan address=\"https://github.com/dtsoucas/DWLS\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and utilized by following the tutorials provided at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/dtsoucas/DWLS/blob/master/Manual.docx\u003c/span\u003e\u003cspan address=\"https://github.com/dtsoucas/DWLS/blob/master/Manual.docx\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Inputs for DWLS included the bulk data, raw counts gene expression matrix of reference scRNA-seq data, and cell type labels, with the latter being annotations from the reference scRNA-seq dataset. To enhance computational efficiency, the findmarker() function from the Seurat package, invoked by the buildSignatureMatrixUsingSeurat() function in DWLS, was substituted with the cosg() function from the COSG R package\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e. All other parameters were maintained at their default settings.\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003eBayesPrism\u003c/h2\u003e \u003cp\u003eThe BayesPrism R package v2.0 was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Danko-Lab/BayesPrism\u003c/span\u003e\u003cspan address=\"https://github.com/Danko-Lab/BayesPrism\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and utilized by following the tutorial available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/Danko-Lab/BayesPrism/blob/main/tutorial_deconvolution.pdf\u003c/span\u003e\u003cspan address=\"https://github.com/Danko-Lab/BayesPrism/blob/main/tutorial_deconvolution.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Inputs for BayesPrism included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and cell state labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset. In accordance with the benchmark paper by Tran et al.\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e, we opted not to utilize the cell state labels option to ensure comparability with other methods, specifying only cell types for BayesPrism. When tumor cells were present, the key parameters in the new.prism() function were configured as tumor cells, while all other parameters were set to the default options of the algorithm.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e \u003ch2\u003eBisqueRNA (Bisque)\u003c/h2\u003e \u003cp\u003eThe Bisque R package v1.0.5 was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/cozygene/bisque\u003c/span\u003e\u003cspan address=\"https://github.com/cozygene/bisque\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and utilized by following the guidelines provided in the tutorials available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/cozygene/bisque/blob/master/vignettes/bisque.Rmd\u003c/span\u003e\u003cspan address=\"https://github.com/cozygene/bisque/blob/master/vignettes/bisque.Rmd\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Inputs for Bisque included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were obtained from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were configured to the default settings of the algorithm.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e \u003ch2\u003eMuSiC\u003c/h2\u003e \u003cp\u003eThe MuSiC R package v1.0.0 was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/xuranw/MuSiC\u003c/span\u003e\u003cspan address=\"https://github.com/xuranw/MuSiC\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and implemented by following the instructions provided in the tutorials at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://xuranw.github.io/MuSiC/articles/MuSiC.html\u003c/span\u003e\u003cspan address=\"https://xuranw.github.io/MuSiC/articles/MuSiC.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Inputs to MuSiC included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were left at their default settings.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003eReCIDE\u003c/h2\u003e \u003cp\u003eThe ReCIDE R package v1.0.0 was obtained from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/limingham/ReCIDE\u003c/span\u003e\u003cspan address=\"https://github.com/limingham/ReCIDE\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e and implemented by following the instructions provided in the tutorials at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/limingham/ReCIDE/blob/main/README.md\u003c/span\u003e\u003cspan address=\"https://github.com/limingham/ReCIDE/blob/main/README.md\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Inputs to ReCIDE included the bulk data, the raw counts gene expression matrix of reference scRNA-seq data, cell type labels, and reference sample labels. Cell type labels were derived from annotations in the reference scRNA-seq dataset, while reference sample labels corresponded to sample labels from the reference dataset. All other parameters were left at their default settings.\u003c/p\u003e \u003c/div\u003e "},{"header":"Declarations","content":"\u003ch2\u003eData available\u003c/h2\u003e\n\u003cp\u003eThis study did not generate any new data. All scRNA-seq and bulk data used in this study are listed in \u003cstrong\u003eTable 1\u003c/strong\u003e and \u003cstrong\u003eTable 2\u003c/strong\u003e, and the download links for all datasets can be accessed at\u0026nbsp;https://github.com/TianLab-Bioinfo/Benchmark-realdata.\u003c/p\u003e\n\u003ch2\u003eCode availability\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eThe R scripts used for analyses in this study is available at https://github.com/TianLab-Bioinfo/Benchmark-realdata.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis work was supported by the National Natural Science Foundation of China [32170667, 32370719].\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eL. and S. contributed equally to this work. L. and T. designed the study. L. and S. performed the analyses and wrote the manuscript. T. supervised the project, provided critical revisions, and approved the final version. All authors reviewed and approved the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eDu W, Elemento O. Cancer systems biology: embracing complexity to develop better anticancer therapeutic strategies. Oncogene. 2015;34:3215\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCha YJ, Park EJ, Baik SH, Lee KY, Kang J. Clinical significance of tumor-infiltrating lymphocytes and neutrophil-to-lymphocyte ratio in patients with stage III colon cancer who underwent surgery followed by FOLFOX chemotherapy. Sci Rep. 2019;9:11617.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIdos GE, Kwok J, Bonthala N, Kysh L, Gruber SB, Qu C. The Prognostic Implications of Tumor Infiltrating Lymphocytes in Colorectal Cancer: A Systematic Review and Meta-Analysis. Sci Rep. 2020;10:3360.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaibach F, Sadozai H, Seyed Jafari SM, Hunger RE, Schenk M. Tumor-Infiltrating Lymphocytes and Their Prognostic Value in Cutaneous Melanoma. Front Immunol. 2020;11:2105.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang D, et al. Advances in single-cell RNA sequencing and its applications in cancer research. J Hematol Oncol. 2023;16:98.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAissa AF, et al. Single-cell transcriptional changes associated with drug tolerance and response to combination therapies in cancer. Nat Commun. 2021;12:1628.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang F, et al. mRNA-Seq whole-transcriptome analysis of a single cell. Nat Methods. 2009;6:377\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJovic D, Liang X, Zeng H, Lin L, Xu F, Luo Y. Single-cell RNA sequencing technologies and applications: A brief overview. Clin Transl Med. 2022;12:e694.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang S et al. The Evolution of Single-Cell RNA Sequencing Technology and Application: Progress and Perspectives. Int J Mol Sci 24, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClough E, Barrett T. The Gene Expression Omnibus Database. Methods Mol Biol. 2016;1418:93\u0026ndash;110.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHutter C, Zenklusen JC. The Cancer Genome Atlas: Creating Lasting Value beyond Its Data. Cell. 2018;173:283\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEdgar R, Domrachev M, Lash AE. Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res. 2002;30:207\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarrett T, et al. NCBI GEO: archive for functional genomics data sets\u0026ndash;update. Nucleic Acids Res. 2013;41:D991\u0026ndash;995.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIm Y, Kim Y. A Comprehensive Overview of RNA Deconvolution Methods and Their Application. Mol Cells. 2023;46:99\u0026ndash;105.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAvila Cobos F, Vandesompele J, Mestdagh P, De Preter K. Computational deconvolution of transcriptomics data from mixed cell populations. Bioinformatics. 2018;34:1969\u0026ndash;79.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNewman AM, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods. 2015;12:453\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTsoucas D, Dong R, Chen H, Zhu Q, Guo G, Yuan G-C. Accurate estimation of cell-type composition from gene expression data. Nat Commun. 2019;10:2975.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang X, Park J, Susztak K, Zhang NR, Li M. Bulk tissue cell type deconvolution with multi-subject single-cell expression reference. Nat Commun. 2019;10:380.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJew B, et al. Accurate estimation of cell composition in bulk expression through robust integration of single-cell information. Nat Commun. 2020;11:1971.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChu T, Wang Z, Pe\u0026rsquo;er D, Danko CG. Cell type and gene expression deconvolution with BayesPrism enables Bayesian integrative analysis across bulk and single-cell RNA sequencing in oncology. Nat Cancer. 2022;3:505\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi M, Su Y, Gao Y, Tian W. ReCIDE: robust estimation of cell type proportions by integrating single-reference-based deconvolutions. Brief Bioinform 25, (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNewman AM, Gentles AJ, Liu CL, Diehn M, Alizadeh AA. Data normalization considerations for digital tumor dissection. Genome Biol. 2017;18:128.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAvila Cobos F, Alquicira-Hernandez J, Powell JE, Mestdagh P, De Preter K. Benchmarking of cell type deconvolution pipelines for transcriptomics data. Nat Commun. 2020;11:5650.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTran KA, et al. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures. Nat Commun. 2023;14:5758.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu M, Chikina M. Heterogeneous pseudobulk simulation enables realistic benchmarking of cell-type deconvolution methods. Genome Biol. 2024;25:169.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarmire LX, et al. Challenges and perspectives in computational deconvolution of genomics data. Nat Methods. 2024;21:391\u0026ndash;400.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHippen AA, et al. Performance of computational algorithms to deconvolve heterogeneous bulk ovarian tumor tissue depends on experimental factors. Genome Biol. 2023;24:239.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBassez A, et al. A single-cell map of intratumoral changes during anti-PD1 treatment of patients with breast cancer. Nat Med. 2021;27:820\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePelka K, et al. Spatially organized multicellular immune hubs in human colorectal cancer. Cell. 2021;184:4734\u0026ndash;e47524720.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, et al. Dissecting esophageal squamous-cell carcinoma ecosystem by single-cell transcriptomic analysis. Nat Commun. 2021;12:5291.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK\u0026uuml;rten CHL, et al. Investigating immune and non-immune cell interactions in head and neck tumors by single-cell RNA sequencing. Nat Commun. 2021;12:7338.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMei S, et al. Single-cell analysis of immune and stroma cell remodeling in clear cell renal cell carcinoma primary tumors and bone metastatic lesions. Genome Med. 2024;16:1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharma A, et al. Onco-fetal Reprogramming of Endothelial Cells Drives Immunosuppressive Macrophages in Hepatocellular Carcinoma. Cell. 2020;183:377\u0026ndash;e394321.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSalcher S, et al. High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Cancer Cell. 2022;40:1503\u0026ndash;e15201508.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeng J, et al. Single-cell RNA-seq highlights intra-tumoral heterogeneity and malignant progression in pancreatic ductal adenocarcinoma. Cell Res. 2019;29:725\u0026ndash;38.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHirz T, et al. Dissecting the immune suppressive human prostate tumor microenvironment via integrated single-cell and spatial transcriptomic analyses. Nat Commun. 2023;14:663.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar V, et al. Single-Cell Atlas of Lineage States, Tumor Microenvironment, and Subtype-Specific Expression Programs in Gastric Cancer. Cancer Discov. 2022;12:670\u0026ndash;91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee H-O, et al. Lineage-dependent gene expression programs influence the immune landscape of colorectal cancer. Nat Genet. 2020;52:594\u0026ndash;603.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFonseca MAS, et al. Single-cell transcriptomic analysis of endometriosis. Nat Genet. 2023;55:255\u0026ndash;67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang Y, et al. Pan-cancer single-cell dissection reveals phenotypically distinct B cell subtypes. Cell. 2024;187:4790\u0026ndash;e48114722.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQian J, et al. A pan-cancer blueprint of the heterogeneous tumor microenvironment revealed by single-cell profiling. Cell Res. 2020;30:745\u0026ndash;62.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi W, et al. A single-cell pan-cancer analysis to show the variability of tumor-infiltrating myeloid cells in immune checkpoint blockade. Nat Commun. 2024;15:6142.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa C, et al. Pan-cancer spatially resolved single-cell analysis reveals the crosstalk between cancer-associated fibroblasts and tumor microenvironment. Mol Cancer. 2023;22:170.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi J, Wang D, Tang F, Ling X, Zhang W, Zhang Z. Pan-cancer integrative analyses dissect the remodeling of endothelial cells in human cancers. Natl Sci Rev. 2024;11:nwae231.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurphy TL, Murphy KM. Dendritic cells in cancer immunology. Cell Mol Immunol. 2022;19:3\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu P, Zhao L, Kroemer G, Kepp O. Conventional type 1 dendritic cells (cDC1) in cancer immunity. Biol Direct. 2023;18:71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchelker M, et al. Estimation of immune cell content in tumour tissue using single-cell RNA-seq data. Nat Commun. 2017;8:2032.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTirosh I, et al. Dissecting the multicellular ecosystem of metastatic melanoma by single-cell RNA-seq. Science. 2016;352:189\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNguyen H, Nguyen H, Tran D, Draghici S, Nguyen T. Fourteen years of cellular deconvolution: methodology, applications, technical evaluation and outstanding challenges. Nucleic Acids Res. 2024;52:4761\u0026ndash;83.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Y, et al. Single-cell analyses reveal key immune cell subsets associated with response to PD-L1 blockade in triple-negative breast cancer. Cancer Cell. 2021;39:1578\u0026ndash;e15931578.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHargadon KM, Johnson CE, Williams CJ. Immune checkpoint blockade therapy for cancer: An overview of FDA-approved immune checkpoint inhibitors. Int Immunopharmacol. 2018;62:29\u0026ndash;39.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYin S, et al. Patient-derived tumor-like cell clusters for personalized chemo- and immunotherapies in non-small cell lung cancer. Cell Stem Cell. 2024;31:717\u0026ndash;e733718.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHan Y, et al. TISCH2: expanded datasets and new tools for single-cell transcriptome analyses of the tumor microenvironment. Nucleic Acids Res. 2022;51:D1425\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdulla S, et al. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Res. 2025;53:D886\u0026ndash;900.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTyler M et al. The Curated Cancer Cell Atlas: comprehensive characterisation of tumours at single-cell resolution. \u003cem\u003ebioRxiv\u003c/em\u003e, 2024.2010.2011.617836 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRani J, Shah AR, Ramachandran S. pubmed.mineR: An R package with text-mining algorithms to analyse PubMed abstracts. J Biosci. 2015;40:671\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eColaprico A, et al. TCGAbiolinks: an R/Bioconductor package for integrative analysis of TCGA data. Nucleic Acids Res. 2015;44:e71\u0026ndash;71.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavis S, Meltzer PS. GEOquery: a bridge between the Gene Expression Omnibus (GEO) and BioConductor. Bioinformatics. 2007;23:1846\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHao Y, et al. Integrated analysis of multimodal single-cell data. Cell. 2021;184:3573\u0026ndash;e35873529.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu G, Wang L-G, Han Y, He Q-Y. clusterProfiler: an R Package for Comparing Biological Themes Among Gene Clusters. OMICS. 2012;16:284\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDai M, Pei X, Wang XJ. Accurate and fast cell marker gene identification with COSG. Brief Bioinform 23, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu SZ, et al. A single-cell and spatially resolved atlas of human breast cancers. Nat Genet. 2021;53:1334\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarisa L, et al. Gene expression classification of colon cancer into molecular subtypes: characterization, validation, and prognostic value. PLoS Med. 2013;10:e1001453.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCao W, et al. Multi-faceted epigenetic dysregulation of gene expression promotes esophageal squamous cell carcinoma. Nat Commun. 2020;11:3675.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWichmann G, et al. The role of HPV RNA transcription, immune response-related gene expression and disruptive TP53 mutations in diagnostic and prognostic profiling of head and neck cancer. Int J Cancer. 2015;137:2846\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003evon Roemeling CA, et al. Neuronal pentraxin 2 supports clear cell renal cell carcinoma by activating the AMPA-selective glutamate receptor-4. Cancer Res. 2014;74:4796\u0026ndash;810.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWoo HG, et al. Integrative analysis of genomic and epigenomic regulation of the transcriptome in liver cancer. Nat Commun. 2017;8:839.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHight SK, et al. An in vivo functional genomics screen of nuclear receptors and their co-regulators identifies FOXA1 as an essential gene in lung tumorigenesis. Neoplasia. 2020;22:294\u0026ndash;310.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang H, et al. A 12-gene set predicts survival benefits from adjuvant chemotherapy in non-small cell lung cancer patients. Clin Cancer Res. 2013;19:1577\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoffitt RA, et al. Virtual microdissection identifies distinct tumor- and stroma-specific subtypes of pancreatic ductal adenocarcinoma. Nat Genet. 2015;47:1168\u0026ndash;78.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePenney KL, et al. Association of prostate cancer risk variants with gene expression in normal and tumor tissue. Cancer Epidemiol Biomarkers Prev. 2015;24:255\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLabb\u0026eacute; DP, et al. High-fat diet fuels prostate cancer progression by rewiring the metabolome and amplifying the MYC program. Nat Commun. 2019;10:4358.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOh SC, et al. Clinical and genomic landscape of gastric cancer with a mesenchymal phenotype. Nat Commun. 2018;9:1777.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"genome-biology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gbio","sideBox":"Learn more about [Genome Biology](https://genomebiology.biomedcentral.com/)","snPcode":"13059","submissionUrl":"https://submission.springernature.com/new-submission/13059/3","title":"Genome Biology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6389993/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6389993/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDeconvolution of bulk RNA-expression data unlocks the cellular complexity of cancer, yet traditional pseudobulk benchmarks falter in real-world settings where absolute cell proportions are unknown. Here, we introduce a novel real-data framework, leveraging 16 real bulk RNA-expression cohorts (4,576 samples) across eight cancer types to evaluate six deconvolution methods based on relative changes in differentially proportioned (DP) cell types\u0026mdash;an impartial and reliable metric. Across three innovative benchmark scenarios\u0026mdash;consistency with scRNA-seq, reproducibility across cohorts, and prognostic relevance\u0026mdash;ReCIDE, Bisque, and BayesPrism have been demonstrated to be the three most robust deconvolution methods. analysis of ten cancer types revealed matrix cancer-associated fibroblasts (mCAF) as a poor prognosis marker (p\u0026thinsp;=\u0026thinsp;0.0081) and CLEC9A\u0026thinsp;+\u0026thinsp;dendritic cells (cDC_CLEC9A) as a favorable one (p\u0026thinsp;=\u0026thinsp;0.016). Furthermore, a prognostic indicator (ASC% - mCAF%) developed using ReCIDE was validated across five TCGA and three GEO cohorts. This study broadens deconvolution benchmarking, offering actionable tools for precision oncology and guiding method selection for translational research.\u003c/p\u003e","manuscriptTitle":"Evaluating Deconvolution Methods Using Real Bulk RNA-expression Data for Robust Prognostic Insights in Pan-Cancer","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-12 09:47:12","doi":"10.21203/rs.3.rs-6389993/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-06-18T11:03:52+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-06-13T16:35:10+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-05-20T16:37:39+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"7725974741721227815465868347439785570","date":"2025-05-08T17:55:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"239887806792043840464776105045924368295","date":"2025-05-06T18:18:24+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-05-06T15:55:54+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-10T11:11:52+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-07T09:35:09+00:00","index":"","fulltext":""},{"type":"submitted","content":"Genome Biology","date":"2025-04-07T04:11:35+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"genome-biology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"gbio","sideBox":"Learn more about [Genome Biology](https://genomebiology.biomedcentral.com/)","snPcode":"13059","submissionUrl":"https://submission.springernature.com/new-submission/13059/3","title":"Genome Biology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"139c836a-658c-4eb7-941f-2c7810359641","owner":[],"postedDate":"May 12th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-01-26T16:07:04+00:00","versionOfRecord":{"articleIdentity":"rs-6389993","link":"https://doi.org/10.1186/s13059-026-03942-1","journal":{"identity":"genome-biology","isVorOnly":false,"title":"Genome Biology"},"publishedOn":"2026-01-21 15:58:30","publishedOnDateReadable":"January 21st, 2026"},"versionCreatedAt":"2025-05-12 09:47:12","video":"","vorDoi":"10.1186/s13059-026-03942-1","vorDoiUrl":"https://doi.org/10.1186/s13059-026-03942-1","workflowStages":[]},"version":"v1","identity":"rs-6389993","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6389993","identity":"rs-6389993","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}