Deep Transfer Learning Links Benign Glands to Prostate Cancer Progression via Transcriptomics.

Deep Transfer Learning Links Benign Glands to Prostate Cancer Progression via Transcriptomics - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Genomics Proteomics Bioinformatics . 2025 Nov 29;23(6):qzaf119. doi: 10.1093/gpbjnl/qzaf119 Deep Transfer Learning Links Benign Glands to Prostate Cancer Progression via Transcriptomics Justin L Couetil 1, 2 , Ziyu Liu 3 , Chao Chen 4 , Ahmed K Alomari 5, 6 , Kun Huang 7, 8, 9, 10 , Jie Zhang 11, 12, 13 , Travis S Johnson 14, 15, 16, 17, ✉ Editor: Shihua Zhang Author information Article notes Copyright and License information 1 Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA 2 Indiana Clinical and Translational Sciences Institute, Indianapolis, IN 46202, USA 3 Department of Statistics, Purdue University, West Lafayette, IN 47907, USA 4 Department of Biomedical Informatics, Stony Brook University, Stony Brook, NY 11794, USA 5 Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA 6 Department of Pathology, IU School of Medicine, Indianapolis, IN 46202, USA 7 Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA 8 Department of Biostatistics and Health Data Sciences, IU School of Medicine, Indianapolis, IN 46202, USA 9 Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA 10 Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA 11 Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA 12 Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA 13 Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA 14 Department of Biostatistics and Health Data Sciences, IU School of Medicine, Indianapolis, IN 46202, USA 15 Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA 16 Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA 17 Indiana Biosciences Research Institute, Indianapolis, IN 46202, USA ✉ Corresponding author: johnstrs@iu.edu (Johnson TS). Roles Shihua Zhang : Handling Editor Received 2024 Feb 19; Revised 2025 Jun 25; Accepted 2025 Nov 24; Collection date 2025 Dec. © The Author(s) 2025. Published by Oxford University Press and Science Press on behalf of the Beijing Institute of Genomics, Chinese Academy of Sciences / China National Center for Bioinformation and Genetics Society of China. This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( https://creativecommons.org/licenses/by/4.0/ ), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. PMC Copyright notice PMCID: PMC13222491  PMID: 41317378 Abstract The field effect describes the phenomena where environmental exposures, infection, and genetic predisposition result in molecular changes in cells that predispose them to developing cancer. Though this is a well-established concept in pathology, it remains underexplored in the context of high-resolution omics. We utilized the Diagnostic Evidence Gauge of Single Cells (DEGAS) deep transfer learning framework to analyze prostate cancer spatial transcriptomics to identify cells and tissues that are highly associated with cancer progression. DEGAS highlighted morphologically benign glands with reduced expression of microseminoprotein-beta ( MSMB ), a differentiation marker downregulated in aggressive tumors. These glands have upregulated genes associated with antigen presentation and aggressive neoplasms. Integration of single-cell transcriptomics and deep learning image analysis separately revealed altered immune-cell infiltration, suggesting a complex interplay in the tumor environment, facilitating aggressiveness. We used immunohistochemistry to quantify the MSMB protein (PSP-94) expression in morphologically normal and tumor tissues from patients with and without 5-year distant metastasis. Samples from patients who developed metastasis consistently showed lower fractions of positively stained cells, indicating a subtle yet significant “field effect” in seemingly benign regions. These proteomic results validate the transcriptomic findings and further underscore that inflammatory or immune-related changes in ostensibly normal tissue may contribute to aggressive disease progression. Keywords: Spatial transcriptomics, Single-cell transcriptomics, Prostate cancer, Deep learning, Tumor immune microenvironment Graphical Abstract Graphical Abstract. Open in a new tab Introduction Prostate cancer is the most common malignancy among men in the United States [ 1 ]. Fortunately, the 5-year survival rates have improved significantly for localized and regional disease, but it remains the second most common cause of cancer-related deaths [ 2 ]. Subsets of patients are much more likely to have a cancer that will spread to distant organs ( i.e. , metastasis), and there are few tools to identify and treat these patients [ 3 ]. Unfortunately, patients with distant metastases have a 5-year survival rate of just 34.1% [ 1 ]. Given the high incidence of prostate cancer and the increasing incidence of metastatic cancer patients each year [ 4 ], developing technologies to identify high-risk prostate cancer patients is of utmost importance. Clinicians have few tools to identify patients at high-risk for disease progression. Recently, the United States Preventive Services Task Force has advised against using prostate specific antigen (PSA) as a screening tool for prostate cancer diagnosis [ 5 ]. PSA is a highly sensitive biomarker for prostate cancer, but does not discriminate between high and low-risk disease. Notably, multi-center studies have demonstrated that low-risk prostate cancer is often over-treated, making the tradeoff between morbid excisional surgery and uncertain survival benefit unacceptable for many patients [ 6 ]. There is a critical need for a reliable biomarker that will identify the subset of prostate cancer that would benefit from intervention [ 7 ]. Developing prognostic models for prostate cancer has been difficult due to a high degree of genetic variability [ 8 ]. Prostate cancer has high intra- and inter-tumor genomic heterogeneity, and multiple polyclonal tumors can arise simultaneously [ 9 ]. Until recently, this heterogeneity has been difficult to study because techniques like bulk omics analyses ( e . g ., RNA-seq) and image-based techniques ( e.g. , immunohistochemistry, IHC) either lack spatial information or are very low throughput with few target molecules per sample. Single cell (SC) analysis then enabled deconvolution ( i . e ., decomposition) of the tumor and microenvironment into their constituent cells [ 10 ]. SC analysis is limited, however, because it lacks both the spatial information and the morphological information of the cells being analyzed. By the central dogma, identifying cell types based solely on gene expression may obscure cells that are phenotypically normal but transcriptionally abnormal. This is directly relevant to phenomena like the epithelial-to-mesenchyme transition, which is fundamental to cancer progression [ 11 ]. Fortunately, high-throughput spatial transcriptomics (ST) has recently become available, allowing for improved study of the heterogeneity of tumors and the microenvironment with a spatial context. ST has the potential to recognize how diverse components of the tumor and microenvironment collectively contribute to clinical outcomes. SC and ST data have been used to reveal an enormous transcriptomic diversity within and between tumors. However, identifying reliable cell gene markers in these heterogeneous and dynamic systems is a key challenge [ 12 ]. Relying on human curated knowledge of gene markers and cell morphology to assign simple labels may obscure transitional cell states of both epithelial [ 13 ] and immune cells [ 14 ]. Similarly, though ST provides the ability to simultaneously analyze transcriptomics and histopathology, traditional visual assessment by humans is semi-quantitative, and machine learning approaches are needed to detect subtle systematic differences in “normal” tissue associated with clinical outcomes [ 15 ]. To address this, we applied the Diagnostic Evidence Gauge of Single Cells (DEGAS) deep transfer learning to ST. This framework has been successfully applied to single-cell transcriptomics to identify clinically relevant molecular features in multiple cancers as well as neurodegenerative diseases [ 16 ]. DEGAS integrates bulk sequencing data with clinical outcomes to map these associations onto high-resolution transcriptomic data. It employs domain adaptation losses to align data distributions across different modalities, enabling the identification of transcriptomic signatures associated with disease progression. In this study, we extended DEGAS to ST to dissect the tumor and microenvironment of prostate cancer. This observed heterogeneity reflects the “field effect”, which has complicated the definition of “normal” tissue for decades [ 17 ]. The field effect describes the finding that chronic inflammation, driven by exposures and genetic predisposition, causes histologically normal cells to acquire genetic and epigenetic changes that contribute to the development of cancer [ 18 ]. These genetic changes reflect etiology and contribute to cancer progression [ 19–21 ]. Aligning with this, applying DEGAS to ST identified inflammatory markers and subtle morphological differences in histologically normal prostate glands that were highly associated with cancer progression. Tools like DEGAS can help in identifying field effects associated with clinical outcomes and, therefore, may be useful for screening and targeted therapies. Results Study Overview The recently developed tool DEGAS [ 16 ] was applied in ST to understand what prostate tissues contribute to aggressive cancers ( Figure 1 ), and the transcriptomics ( Figure 2 ), cell type composition ( Figure 1D ), and tissue morphology of these high-risk regions were analyzed ( Figure 1E ). The overall experimental workflow ( Figure 2 ), results of feature selection ( Figure S1 ), and results of model aggregation ( Figure S2 ) are all outlined. The resources used in this study ( Table 1 ) are reported in File S1 in compliance with STROBE guidelines to the extent that metadata were available [ 22 ]. Figure 1. Open in a new tab DEGAS workflow A . Integration of bulk and spatial transcriptomics. Patient bulk tumor RNA-seq was used to predict clinical outcomes and phenotypic information using a deep-learning framework that simultaneously matches ST gene expression distribution to that of patients. By aligning gene expression distributions across these two data types, DEGAS can then assign (“transfer”) risk scores to the spatial transcriptomic arrays, highlighting high-risk tissue. B . Comparison of RNA expression in high and low-risk tissues. C . Correlations between gene expression and risk in high and low-risk tissues. D . Morphological identification of BGs with malignant transcriptional profiles in an additional dataset. Integration with SC data allows for identification of cell-type specific expression changes associated with poor progression-free survival, and SC clusters annotated as cancer localize to histologically benign regions of ST data. E . Deep learning approaches for analyzing the histology from these different regions, identifying similarities between the high-risk glands, high-grade cancer, and inflammation. t-SNE graphic was generated from BioRender. RNA-seq, RNA sequencing; ST, spatial transcriptomics; DEGAS, Diagnostic Evidence Gauge of Single cells; BG, benign gland; SC, single cell; t-SNE, t -distributed stochastic neighbor embedding; CNN, convolutional neural network. Figure 2. Open in a new tab Experimental overview A . In the discovery workflow, a DEGAS model is trained to predict progression-free survival from the RNA-seq of prostate cancer TCGA patients. This information is mapped onto four ST samples from 10X Genomics, identifying benign morphology glands that are highly associated with poor progression-free survival. Functional enrichment shows these high-risk glands have signatures of carcinogenesis and high-grade cancer. MSMB is consistently negatively correlated with risk. B . We validated these findings in seven ST samples from Erickson et al. [ 28 ]. Using the tool RCTD, we deconvolved these samples with an SC dataset from Song et al. [ 29 ]. This demonstrates that cancer cell clusters are attributed to histologically BGs. C-SIDE of the deconvoluted ST data shows that MSMB loss is specific to high-risk luminal epithelial cells, the originators of prostate cancer. Finally, we investigated whether there might be morphologic features that are too subtle for a pathologist to interpret but could be discerned by deep-learning models. The CNN was unable to accurately differentiate the high and low-risk glands. Taking an unsupervised approach, an AE was used to compress images. These compressed representations were clustered, revealing one cluster that was disproportionally represented among the regions of the highest risk glands, highest grade cancer, and inflammation. TCGA, The Cancer Genome Atlas; RCTD, Robust Cell Type Decomposition; C-SIDE, Cell-type Specific Inference of Differentiation Expression; AE, autoencoder; UMAP, uniform manifold approximation and projection; MSMB , microseminoprotein-beta . Table 1. Overview of data sources, sample size, assay type, and additional attributes for prostate cancer datasets used in discovery and validation experiments Experiment Dataset Sample size Data type Additional attribute Discovery & validation TCGA-PRAD 493 patients RNA-seq & progression-free survival Discovery dataset 10X Genomics Normal: 2543 spots 10X Visium ST Histological labels Stage II: 3460 spots Stage III: 4371 spots Stage IV: 3043 spots Validation dataset [ 28 ] H1_2: 2775 spots 10X Visium ST Histological labels & whole-slide images H1_4: 4079 spots H1_5: 3856 spots H2_1: 3092 spots H2_5: 3554 spots V1_2: 2736 spots [ 29 ] 21,743 cells from prostate cancer and surrounding tissue Seq-Well SC Cell type annotations Open in a new tab Note : TCGA, The Cancer Genome Atlas; PRAD, prostate adenocarcinoma; RNA-seq, ribonucleic acid sequencing; ST, spatial transcriptomics; SC, single cell. Discovery cohort analysis DEGAS identifies morphologically BGs associated with diminished progression-free survival We used our newly established pipeline ( Figures 1 and 2 ) to generate DEGAS risk scores for ST data. DEGAS risk scores highlighted certain tissues in the normal prostate sample that were associated with cancer progression ( Figure 3 A ). A comprehensive review of the hematoxylin and eosin (H&E)-stained whole-slide image by a pathologist revealed that the DEGAS heatmap was highlighting glandular tissue. The highest-risk glands ( Figure 3A and E ) were devoid of any morphological indicators typically associated with malignancy, and no morphological signs of neoplasia were present ( Figure S3A ). The pathologist noted only a few instances of urothelial metaplasia, a benign phenomenon in which urothelial cells coexist with epithelia of the peripheral prostatic glands and ducts [ 23 ]. Figure 3. Open in a new tab DEGAS risk impressions in four spatial transcriptomic samples of prostate tissue A . Histologically normal prostate tissue. The black arrow indicated a particularly high-risk region, with a corresponding H&E-stained section below. B . Stage II prostate adenocarcinoma with tissue section below. This specimen was stained with immunofluorescence and thus not H&E-stained. C . Stage III prostate adenocarcinoma with right-adjacent normal tissue (blue), and an arrow indicating a particularly high-risk region. In the corresponding H&E, below, examination by a pathologist revealed that the region is a nerve. D . Stage IV prostate acinar cell carcinoma ( i.e. , adenocarcinoma) and corresponding H&E section below. E . Histological labels from normal prostate tissue shown in A. F . Scatter plot overlaid with violin plot of the hazards for all ST spots in Panels A–D, showing trend of increasing hazard with increasing stage. G . Differential gene expression plot using Wilcoxon test FDR-adjusted P values, and the log 2 mean fold change of the samples in panels A–D. High (red) and low (blue) risk scores from the DEGAS model are visualized. All samples share the same hazard/color scale. H&E, Hematoxylin and eosin. Histological characteristics associated with DEGAS risk predictions in the prostate cancer tissue samples were annotated. However, identifying correlates was not feasible for the Stage II prostate adenocarcinoma due to the absence of a diagnostic H&E-stained image ( Figure 3B ). In the Stage III prostate adenocarcinoma ( Figure 3C ), the high-risk region on the left corresponded with the bulk tumor. Conversely, on the right, both the peritumoral stroma and some prostatic glands were identified as low-risk (highlighted in blue), interspersed with a few high-risk benign glands (BGs, marked in red). Notably, an isolated nerve in the stroma of the Stage III sample was identified as high-risk ( Figure 3C ). For the Stage IV sample ( Figure 3D ), the areas identified as high-risk were consistent with regions displaying slightly higher Gleason score morphologies, characterized by closely packed glands, some dissociated cancer cells, and a prominent presence of lymphocytes ( Figure S3B ). Intuitively, DEGAS risk scores increased with increasing cancer stage ( P = 8.32E−4, P = 8.14E−26, P = 1.54E−228; Figure 3F ). Differential expression in high-risk normal prostate glands reveals signatures associated with high-grade prostate cancer and carcinogenesis Differential expression (DE) analysis was used to pinpoint genes that discriminate between high-risk and low-risk regions ( Figure 3G ). For the normal prostate sample, DE was performed between only the high and low-risk glands, thereby avoiding potential confounders from comparing against other low-risk histologic regions, like the collagenous stroma. Subsequent functional enrichment analysis of DEGs from the normal samples revealed upregulation of RNA signatures associated with pathways of carcinogenesis, congenital genetic disorders, inflammation, immune interactions, and high-grade cancers. Notably, these upregulated pathways include Triple Negative Breast Neoplasms ( P = 6.53 × 10 −9 ), Neoplasm invasiveness ( P = 9.91 × 10 −9 ), Epithelium development ( P = 7.29 × 10 −7 ), and Antigen processing and presentation of exogenous antigen via MHC Class II ( P = 8.81 × 10 −6 ; Table S1 ). Analysis across the Normal, Stage II, Stage III, and Stage IV samples revealed four common enrichment categories among the upregulated DEGs: High-grade prostatic intraepithelial neoplasia, Hormone refractory prostate cancer, Prostatic neoplasms, and Renal Cell Carcinoma ( Table 2 ; Figures S4 and S5 ). Table 2. Overlap in functional enrichment of high-risk regions in normal, stage II, III, and IV samples Enrichment term Gene Normal Stage II Stage III Stage IV High-grade prostatic intraepithelial neoplasia ACTB , GSTP1 , IGFBP3 , SQSTM1 , CTNNB1 3.78E−03 2.23E−03 4.32E−03 9.49E−03 Hormone refractory prostate cancer PSAP , HMGB1 , PABPC1 , DDX5 , GSTP1 , IGFBP3 , CLU , AHNAK , SLPI , CTNNB1 , MMP7 , NONO , HSPB1 9.46E−04 1.65E−04 1.74E−03 3.05E−04 Prostatic neoplasm LITAF , B2M , RUNX1 , GSTP1 , IGFBP3 , ZFP36L2 , CLU , EHF , CTNNB1 , CTSB , CTSD , CNN3 1.40E−03 3.47E−05 5.61E−06 3.05E−04 Renal cell carcinoma ACTB , GSTP1 , IGFBP3 , SQSTM1 , CTNNB1 2.84E−07 2.16E−11 8.37E−03 4.91E−05 Open in a new tab Note: ACTB, actin beta; GSTP1, glutathione S-transferase pi 1; IGFBP3, insulin-like growth factor binding protein 3; SQSTM1, sequestosome-1; CTNNB1, catenin beta 1; PSAP, prosaposin; HMGB1, high mobility group box 1; PABPC1, poly(A) binding protein cytoplasmic 1; DDX5, DEAD-box helicase 5; CLU, clusterin; AHNAK, AHNAK nucleoprotein; SLPI, secretory leukocyte peptidase inhibitor; MMP7, matrix metallopeptidase 7; NONO, non-POU domain containing octamer binding; HSPB1, heat shock protein family B (small) member 1; LITAF, lipopolysaccharide induced TNF factor; B2M, beta-2-microglobulin; RUNX1, RUNX family transcription factor 1; ZFP36L2, ZFP36 ring finger protein like 2; EHF, ETS homologous factor; CTSB, cathepsin B; CTSD, cathepsin D; CNN3, calponin 3. Loss of MSMB is strongly associated with high-risk BGs The Spearman’s correlation coefficient (SCC) was used to identify consistent relationships between genes and DEGAS risk scores. These values were calculated for each sample individually, providing four values, one for each of the four samples. The SCC distributions varied among these samples ( Figure 4 A ). Plotting the maximum and minimum SCC highlighted microseminoprotein-beta ( MSMB ) as the sole gene exhibiting a strong and consistent negative correlation with high-risk areas in all four samples (SCC range = [−0.339, −0.621], P value range = [< 1E−100, 7.98E−68]; Figure 4B ). Conversely, many genes, like eukaryotic translation elongation factor 2 ( EEF2 ), showed a consistent positive correlation with high-risk areas (SCC range = [0.288, 0.569], P value range = [2.96E−82, 3.64E−293]. Others demonstrated negative and positive correlations across different samples, like kallikrein related peptidase 3 ( KLK3 ) ( i . e ., PSA gene, Normal = −0.303, Stage II = −0.594, Stage III = 0.562, Stage IV = 0.450), neuropeptide Y ( NPY ) (Normal = −0.295, Stage II = −0.422, Stage III = 0.601, Stage IV = 0.449), and ubiquitin C ( UCB ) (Normal = 0.587, Stage II = 0.666, Stage III = −0.328, Stage IV = 0.457, Figure 4C ). Representative genes from each pattern are depicted in Figure 4D . Additionally, there is a relationship between DEGAS risk and total RNA expression in the normal tissue (SCC = 0.546, P = 1.15E−197), emphasizing that MSMB loss is counter to overall higher transcriptional activity in high-risk normal glands ( Figure S6 ). Figure 4. Open in a new tab Coefficients of Spearman’s correlation between genes and DEGAS hazard output across all ST spots on all prostate cancer samples A . Histogram of correlation coefficients between genes and the DEGAS hazard output for each spot. The distributions are different. For the normal sample, the SCC was not restricted to only glandular tissues (thus, the title refers to “all ST spots”). B . Concordance between SCC across the different prostate samples. It is demonstrated by plotting the range of coefficients across the four samples (maximum and minimum). MSMB is the only gene for which the SCC is consistently negative across all slides. EEF2 and others are consistently associated with high hazard. Some genes with positive and negative coefficients are KLK3 , NPY , HSPB1 , and CALD1 . C . Panel B overlaid with the FDR adjusted −log P value for each gene’s enrichment in the high-risk glandular tissues. Labeled genes have −log P value > 120. D . Expression of representative genes from panels B & C visualized on four prostate ST samples, in order from left to right: Normal, Stage II, Stage III, and Stage IV. DEGAS hazard provided top left for reference, where blue indicates hazard impression below the median value, and red is above the median. SCC, Spearman’s correlation coefficient; FDR, false discovery rate. Both MSMB and KLK3 serve as epithelial markers in prostate cancer [ 24 ]; however, their expression patterns diverged: MSMB expression was reduced in the high-risk benign-appearing glands (log 2 fold change = −1.38, P = 2.05E−67) and consistently less expressed in tumor tissue. KLK3 was expressed in both tumor and normal glands but was specifically downregulated in the highest-risk glands (log 2 fold change = −1.16, P = 2.05E−67). EEF2 , a translation elongation factor, has been implicated in promoting tumor progression across multiple malignancies, including prostate cancer. Its overexpression has been linked to increased protein synthesis, supporting cellular proliferation and metastatic potential [ 25 ]. NPY , a neuropeptide widely studied in stress responses, has been shown to influence tumor progression and the tumor microenvironment in prostate cancer, particularly through its interactions with the sympathetic nervous system and immune modulation [ 26 ]. UBC, encoding polyubiquitin, is a crucial component of the ubiquitin-proteasome system, regulating protein degradation and cellular stress responses. Dysregulation of ubiquitin pathways has been linked to oncogenesis and may contribute to tumor adaptation and progression [ 27 ]. Though these other genes have known roles in cancer, MSMB had the strongest correlation with DEGAS risk scores and was consistently negatively correlated with risk across all four samples. Moreover, it was significantly downregulated in the high-risk glands of the normal sample. Thus, we thought MSMB may serve as a specific marker for high-risk normal glands and be useful for patient risk stratification. Rather than immediately proceeding with proteomic validation, these results prompted us to first validate the high-risk gland-low MSMB relationship with an additional ST data set and investigate the immunological milieu using single-cell deconvolution methods. Validation cohort analysis DEGAS identifies high-risk glands with loss of MSMB that coincides with increased myeloid-origin inflammatory cells and decreased T-cells and fibroblasts To identify consistent and reproducible genetic findings, an additional independent prostate cancer ST dataset from Erickson et al. [ 28 ] was analyzed. This research group inferred DNA-copy number variations from ST, identifying that tumor tissue was clonally related to BGs. DEGAS also again highlighted regions of high-risk normal glands ( Figure 5 A and B ). Some of the highest-risk regions ( Figure 5A ) were identified in a sample that comes from the opposite side of the prostate from the main cancer focus ( Figure 5A ). DEGAS risk predictions varied both within and between histologic tissues, emphasizing that histological labels alone are insufficient to identify tissue heterogeneity associated with prostate cancer progression ( Figure 5C ). The benign morphology glands had large variations in risk scores ( Figure 5D ). To create categorical variables for visualization, the BGs were separated into four groups of increasing risk at roughly equally spaced DEGAS risk score intervals ( Figure 5D ). There were 1655 spots in rank 1, 1039 spots in rank 2, 1402 spots in rank 3, and 882 spots in rank 4. BG rank 1 (lowest risk) rarely overlapped with BG rank 4 (highest risk), whereas BG ranks 2 and 3 have intermediate risk and show more spatial colocalization ( Figure 5E ). Figure 5. Open in a new tab DEGAS highlights high-risk morphologically BGs with discrete spatial localization in external dataset A . DEGAS hazard impressions for seven Visium ST samples from the prostate of a single patient. B . Histology of both cancerous and benign across the seven samples. C . Violin plots of DEGAS hazards showing a trend between increasing tumor grade and hazard predictions across GG1 through GG4 regions. D . The morphologically benign tissues with different DEGAS Hazard distributions. They were divided into four ranks of increasing hazard for analysis. E . Visualization of the four ranks in the original histopathology. GG, Gleason grade. Because each ST spot contains multiple cells, the Robust Cell Type Decomposition (RCTD) cell type deconvolution algorithm was employed to discern how DEGAS risk relates to cell-type abundance. Using an annotated prostate cancer SC dataset from Song et al. [ 29 ], RCTD was able to confidently assign SC clusters to most regions of the ST data ( Figure 6 A ). Erythroblast transformation-specific related gene (ERG)-positive tumor cells from the SC were associated with the tumor-dense areas of the ST samples ( Figure 6B ). Deconvolution of the benign regions revealed a mix of both ERG-negative tumor cells and normal epithelial cells ( Figure 6C ), and a diversity of immune cell types in both tumor and normal tissue ( Figure 6D ). Figure 6. Open in a new tab ST-spot deconvolution with annotated single-cell dataset and cell-specific differential expression associated with DEGAS risk A . Cell deconvolution/decomposition at each ST spot by the RCTD attributes SCs to each ST spot. If there are multiple predominant cell types at a location, RCTD provides “doublets”. Certain locations cannot be confidently decomposed and are called “reject”. H1_5 is the majority cancer sample and is attributed mostly to a SC type. Other samples show greater heterogeneity in the number of cell types ascribed to each location. B . Both ERG-positive and negative tumor cells are mapped in both tumor and non-tumor histological regions. C . Normal tissue contains predominantly luminal epithelia, and tumors also have luminal cells. D . Immune cells, mostly myeloid and plasma cells, are scattered throughout each sample. There is a group of B-cells in the lower right corner of the V1_2 sample. E . Cell-type composition of the four morphologically BG risk shows increasing populations of myeloid cells, B-cells, and fewer luminal endothelia, fibroblasts, and smooth muscle. F . Cell-type specific differential expression associated with DEGAS risk scores shows MSMB downregulation is specific to luminal epithelia (red rectangle). Cell-type composition varied across the four different risk groups ( Figure 6E ). The Spearman’s correlations between DEGAS risk scores and RCTD cell-type enrichment among the histologically normal glands were calculated ( Figure S7 ). The three strongest positive correlations were B-cells (SCC = 0.25, P = 1.15E−72), plasma cells (SCC = 0.25, P = 1.15E−72), and myeloid cells (SCC = 0.19, P = 5.41E−42). The three strongest negative correlations were fibroblasts (SCC = −0.27, P = 5.90E−82), T-cells (SCC = −0.16, P = 3.91E−25), and basal epithelia (BE, SCC = −0.15, P = 1.44E−27). We note that ERG-positive epithelial cells increased (SCC = 0.16, P = 6.55E−30) and ERG-negative epithelial cells decreased (SCC = −0.14, P = 2E−22). The tool Cell Type-Specific Inference of Differential Expression (C-SIDE) was further used to understand how gene expression of each cell type changed with DEGAS risk, while controlling for these changing cell type proportions. This analysis showed that MSMB downregulation is specific to high-risk luminal epithelial cells, recapitulating the results of the discovery set ( Figure 6F ). Deep learning image analysis reveals subtle similarities between high-risk glands, high-grade cancer, related to increasing inflammation, features of high-grade cancer, and descreasing stroma Deep learning image analysis was used to determine whether there were subtle morphological differences in prostate tissue histology that coincided with these transcriptomic changes. The ability to recognize high-risk glands with myeloid infiltration would be a useful clinical and research tool, so a state-of-the-art convolutional neural network (CNN) was fine-tuned to classify the ST data into six categories: the four BG risk groups, cancer tissue, and a collective “other” category for all remaining histology. This CNN displayed high accuracy in identifying cancer and “other” histology but failed to differentiate the BG risk groups ( Figure 7 A ). There were significantly more “cancer” and “other” patches, which may have biased the model, so a CNN was trained to differentiate just the four BG risk groups. This model was only capable of near-random accuracy and showed overfitting to the dataset. This highlights the need for multi-omics, and specifically ST, to capture both the molecular and morphological profiles underlying the field effect ( Figure 7B ). Figure 7. Open in a new tab Clustering of image patches reveals subtle similarities between high-risk prostate glands and high-grade cancer A . Classification of sample groups by confusion matrix and training log of EfficientNetB4 CNN fine-tuned. Six groups were classified: The four BG risk groups, cancer, and all other histology grouped under the label “other”. The CNN accurately identified only the cancer and “other” regions (“C” and “Oth”). B . Classification of the four BG risk groups by a separately fine-tuned CNN. Due to the heavy class imbalance, a separate CNN was fine-tuned to classify only the four BG risk groups. CNN performance in the validation data is not better than random classification and shows evidence of overfitting. C . K-means clustering of patches based on two principal components from PCA of autoencoder embeddings. PCA of autoencoder embeddings for patches from all samples. Color represents K-means clustering based on two principal components. D . Proportion of clusters across histology. The proportion of these clusters in different histology shows that Cluster 2 (green) increases in increasingly high-risk normal glands. Cluster 2 is strongly represented in the Chronic inflammation, Inflammation, PIN/Transition state, and high-grade cancer tissue (GG4, GG4 cribriform). Cluster 2 is largely absent from the vessel, stroma, nerve, and fat regions. Cluster 1 is highly represented in Fat, stroma, nerve, and vessel, decreasing in proportion with increasing BG risk rank. E . Calculation and plotting of the mean PCA embedding. Fat, stroma, nerve, and vessels were separated from epithelial and inflammatory histology. Centroids of BG risk ranks 1 through 4 progresses sequentially in the direction of GG4 tumor, aligning with increasing DEGAS risk scores. CNN, convolutional neural network; PCA, principal component analysis; PC, principal component; val, validation. Taking an unsupervised deep learning approach, an autoencoder was used to compress the histological images. Principal component analysis (PCA) of the compressed images revealed a point cloud, and K-means was used to divide this point cloud ( Figure 7C ). PC1 and PC2 represented 10.8% and 4.6% of the variance of the ∼ 200,000 image features. The proportion of these four clusters was compared to histological labels, which provided interpretability ( Figure 7D ): Cluster 2 was strongly represented in the highest-grade tumor tissue (GG4, GG4 cribriform), chronic and acute inflammation, and intraepithelial neoplasia. Cluster 1 was strongly represented in fat and stromal tissue. Clusters 0 and 2 did not show an obvious pattern of enrichment that aligned with a histology. Crucially, as the DEGAS risk score of normal-appearing glands increased, these regions had increased proportions of Cluster 2 and decreased proportions of Cluster 1 ( P = 3.85E−28), reflecting an increase in inflammation-cancer image signature and a decrease in normal fat-stromal image signature. The standardized Chi-squared residual for Clusters 2 and 1 in the BG rank 4 glands were the largest significant residuals: 10.11 and −5.62, respectively. When a PCA centroid was calculated for each histological label, the clusters were clearly separable ( Figure 7E ). Close inspection revealed a parallel progression of BG rank 1 to BG rank 4 glands, as well as benign stromal tissues towards inflammatory tissue and high-grade Gleason grade 4 (GG4) cancer along the first principal component. These image results align directly with the DEGAS-derived transcriptomic findings, i.e. , high-risk morphologically BGs exist with an increased inflammatory and decreased stromal components. This trend was also identified when correlating the mean loading of the first 30 components (48% of variance in data) with the DEGAS risk scores ( Figure S8 ). IHC Validation Protein expression of MSMB is highly associated with 5-year distant metastasis, with normal tissue having more predictive power than tumor regions To validate our spatial transcriptomic findings at the protein level, we performed IHC staining for MSMB (PSP-94) in FFPE prostatectomy samples from patients who either developed distant metastases within five years ( n = 4) or remained metastasis-free for at least five years ( n = 4). All patients had International Society of Urological Pathology (ISUP) Grade 5 primary tumors, and three of four patients in each group had positive lymph nodes at diagnosis. Age across the two groups was not significantly different (two-tailed Wilcoxon P = 0.4651, Table S2 ). These well-balanced baseline characteristics ensured that differences in IHC staining were not driven simply by tumor grade or stage. Representative IHC images and quantification are displayed in Figure 8 . The whole-slide image (WSI) thumbnails from four representative patients revealed a clear reduction in MSMB staining within normal tissue among those with distant metastasis ( Figure 8A ). Using QuPath to quantify the cytoplasmic IHC staining intensity for every cell within the parenchyma (non-stromal tissue) showed a significant ability to stratify patient prognosis, whether treating the images as independent samples (one-tailed Wilcoxon P = 0.00235), or analyzing the IHC staining intensity summarized to the patient level (one-tailed Wilcoxon P = 0.0286, Figure 8B ). Importantly, the difference in prognostic significance of the normal tissue was greater than that of tumor tissue at the patient level (one-tailed Wilcoxon P = 0.0286 vs . P = 0.243). Figure 8. Open in a new tab Loss of MSMB protein expression in BGs predicts 5-year distant metastasis in prostate cancer patients A . Representative IHC staining for MSMB protein in tumor and benign (normal) prostate tissue from patients with (red, right) and without (blue, left) 5-year distant metastases (40× magnification). Patients who developed distant metastases exhibited lower MSMB expression, particularly in BGs. B . Quantification of MSMB staining intensity at the sample level (top) and patient level (bottom). At the sample level, MSMB expression in normal glands was significantly lower in patients with distant metastases ( P < 0.001), while tumor regions showed a weaker association ( P = 0.005). At the patient level, MSMB staining in benign regions remained predictive of metastasis ( P = 0.029), whereas tumor regions did not ( P = 0.243). C . Threshold-based analysis of IHC staining intensity, showing the fraction of cells classified as MSMB-positive at increasing intensity thresholds. MSMB expression in normal prostate tissue consistently stratified metastatic and non-metastatic patients across a wide range of intensity thresholds, with a stronger effect observed in benign regions than in tumor regions. IHC, immunohistochemical. IHC intensity scoring in pathology is typically done by using an intensity threshold to define positive cells, and the fraction of positive cells is correlated to clinical outcomes [ 30 , 31 ]. The selection of the threshold is extremely important and susceptible to bias [ 32 ], so we systematically evaluated 20 different percentiles along the entire stain intensity distribution. The prognostic power of MSMB protein IHC remained robust across a wide range of intensity thresholds ( Figure S9 ). In normal tissue, the fraction of cells classified as MSMB-positive, as defined by staining greater than the 50th percentile, clearly distinguished the metastatic from the non-metastatic group (AUC = 1, one-Tailed Wilcoxon P = 0.014). In contrast, tumor regions showed slightly weaker discriminatory power (AUC = 0.812, One-tailed Wilcoxon P = 0.1). These results strongly support the notion that MSMB loss in ostensibly normal prostate glands, both at the transcriptomic and proteomic level, is linked to aggressive disease. Furthermore, the observation that normal tissue exhibited superior predictive capacity relative to tumor tissue underscores the relevance of a “field effect,” wherein molecular alterations in ostensibly benign regions may foreshadow metastatic progression. This protein-level validation of our transcriptomic findings suggests that measuring MSMB expression in BGs could serve as a promising biomarker strategy for identifying patients at higher risk of distant recurrence. Discussion The goal of this study was to understand how different components of the prostate tumor, microenvironment, and ostensibly normal tissue in the same organ are associated with cancer progression. Utilizing over 400 RNA-seq samples, 11 ST samples across two independent datasets, single-cell data, and histopathological imaging, DEGAS repeatedly demonstrated that certain morphologically BGs are highly associated with prostate cancer progression. These high-risk glands exhibit upregulated RNA signatures of antigen presentation and neoplasms and are associated with the loss of MSMB . These molecular changes in benign-appearing glands suggest a role in cancer progression, and our future work will focus on confirming their contribution to carcinogenesis and distant spread. Our results build on RNA-seq research that has shown the prognostic value of normal-adjacent tissue in eight different cancers [ 33 , 34 ]. The growing availability of ST technologies will facilitate the continued high-resolution dissection of the microenvironment to understand how the complex interplay contributes to patient prognosis. Integrating SC and ST by cell-type decomposition demonstrated that MSMB loss occurs specifically in luminal epithelia, the putative progenitor of prostate adenocarcinoma. Additionally, the decrease of MSMB coincides with infiltration by inflammatory cells of the myeloid lineage and reduction of fibroblasts and T-cells. Validating the transcriptomic results by analyzing ST histopathology demonstrated that distinguishing high and low-risk glands is difficult, but high-risk glands share subtle imaging similarities to high-grade cancer tissue associated with increased inflammation. Our findings support a hypothesis that the observed increases in antigen presentation signatures, myeloid inflammation, and loss of epithelial differentiation markers in benign-appearing glands may play a role in the link between chronic inflammation, immune desensitization, and the progression of prostate cancer. Supporting our identification of field effects in high-resolution transcriptomics, there is substantial literature in pathology and molecular epidemiology linking MSMB [ 35 , 36 ], differentiation [ 24 , 37 ], chronic inflammation [ 38 , 39 ], and myeloid cells [ 40 , 41 ] to prostate cancer progression. Therefore, our future work will focus on leveraging these different components of the field effect as metastasis biomarkers. This has been preliminarily explored, and MSMB loss in normal glands was moderately predictive of higher prostate cancer grade and invasion outside prostate [ 42 ]. These experiments, however, relied on expert pathologists’ qualitative scoring and did not directly address metastasis. Metastasis prediction is the most important clinical need in prostate cancer management. Therefore, we are undertaking retrospective clinical validation of MSMB staining in prostatectomies from a case-control cohort of metastatic and non-metastatic patients. Additionally, the use of digital image analysis allows for direct quantification of MSMB staining, potentially improving metastasis prediction. We hope in our future studies to address the technical limitations of the currently available assays and methodologies enumerated here. Visium ST provides a snapshot in time, limiting the ability to discern dynamic processes in the microenvironment. Therefore, it cannot be directly discerned whether these high-risk glands become malignancies or whether they contribute to the progression of existing malignancies. This conundrum is especially interesting because almost all prostate cancers are polyclonal: sectioning the entire prostate will reveal several simultaneous tumors of different genetic origin [ 9 ]. Intriguingly, it has been shown that some prostate cancer metastatic foci are derived from lower-grade progenitor cancer cells, rather than the bulk, higher-grade tumors that brought on clinical suspicion [ 43 ]. This is counter to traditional pathology and bioinformatics thought. MSMB and myeloid cells have not been studied simultaneously, and their spatial colocalization may be of key importance to prognosis. Broader validation of the prognostic value of MSMB and myeloid inflammation at the proteomic and transcriptomic level is needed, especially in cohorts of diverse racial and ethnic backgrounds for whom there are known differences in outcomes [ 44 ] and tumor biology [ 45 , 46 ]. Using our deep transfer learning methods, we identified morphologically BGs that exhibited loss of differentiation markers and increased signatures for antigen presentation and aggressive disease. However, further studies are needed to confirm whether these changes contribute directly to cancer progression. These same high-risk morphologically BGs showed increased myeloid infiltration, which was only revealed via SC integration techniques. Fine-tuned deep learning image analysis revealed subtle changes in morphology related to inflammation. Markers for these high-risk, morphologically BGs, in conjunction with PSA and other demographic information, hold potential as part of the next generation of precision prostate oncology testing, pending further validation. Our IHC validation of MSMB protein loss in normal prostate glands supports these findings and underscores the importance of the “field effect”. Notably, we observed that MSMB expression had greater prognostic utility in benign tissue than in tumor regions, suggesting that transcriptional and proteomic alterations in ostensibly normal glands may precede or drive aggressive behavior. This is consistent with prior studies linking chronic inflammation and subtle epithelial changes to malignant progression. Importantly, the strength of this association is particularly striking given that our cohort consisted of just 8 patients with the same ISUP Grade 5 tumors, many of whom also had positive lymph nodes at diagnosis. These patients, under current clinical guidelines, would all be classified as high-risk and managed similarly. The fact that MSMB expression in morphologically BGs could perfectly distinguish which patients developed distant metastases — despite identical clinical staging — suggests that it captures an aspect of disease biology not currently incorporated into standard prognostic models. This highlights MSMB as a compelling biomarker candidate for refining metastatic risk stratification, and the low-expression tissue regions as important foci for targeted omics analyses. Future studies will focus on combining MSMB expression in BGs with other established clinical markers ( e.g. , PSA, Gleason grade, and genomic tests) to further refine risk stratification for metastatic progression. Such multimodal biomarker panels could be particularly valuable for identifying men who would benefit from intensified surveillance or adjuvant therapies. Ultimately, integrating ST, single-cell analysis, and advanced imaging-based pathology scoring may provide a comprehensive toolkit for precision oncology in prostate cancer. Materials and methods DEGAS framework with new applications on spatial transcriptomic data The deep transfer learning framework DEGAS [ 16 ] is constructed using TensorFlow in Python [ 47 ] and wrapped with an R interface for broader accessibility [ 48 ]. In essence, this algorithm employs maximum mean discrepancy loss to synchronize the distributions of the two different datasets within a latent space. This facilitates the transfer of labels and patient outcomes ( e.g. , mutational status and survival) between the datasets. This algorithm’s design, specifically the shared gene-level input features between ST and SC data, enables its effective application to ST data [ 49 ]. Applying DEGAS to ST data is particularly critical as it bridges gene expression and histopathology, offering a powerful tool to identify prognostic regions in the tumor and microenvironment. In this work, we trained DEGAS to predict “progression-free survival” using bulk patient tissue and mapped this information onto the ST. DEGAS can be downloaded from GitHub at https://github.com/tsteelejohnson91/DEGAS . Data pre-processing, feature selection, and analysis with DEGAS All gene expression data ( Table 1 ) were processed with DEGAS’s preprocessCounts function, which performs a log 2 transformation and Z-score normalization, then scales RNA counts to a [0, 1] range. DEGAS was used in the standard DenseNet configuration with three hidden layers and five-fold bootstrap aggregation. Risk scores for ST data were normalized against the median, assigning “high-risk” to scores above zero and “low-risk” to those below. When handling multiple prostate samples, the DEGAS model was trained on a combined matrix, allowing for consistent risk score comparison across samples. For feature selection, genes common to the ST and bulk tissue RNA-seq datasets were identified. Among these, the top 200 genes with the highest variability among The Cancer Genome Atlas (TCGA) patients served as input features for subsequent DEGAS models. To improve the generalizability of the models, ensemble learning principles were employed [ 50 ], aggregating training results from incrementally increasing feature sets, starting with 10 genes and doubling sequentially up to 200 genes to regularize the DEGAS risk predictions. Discovery cohort analysis Differential gene expression and correlation of DEGAS risk score with gene expression The DEGAS risk scores across the Normal, Stage II, Stage III, and Stage IV samples were compared using the Wilcoxon rank-sum test from the R stats package [ 48 ]. Differentially expressed genes (DEGs) between high and low-risk regions within the discovery dataset were also identified using the Wilcoxon test. The specific tissue regions under comparison for differential expression are detailed within the respective experiments. This non-parametric test statistic was chosen because it minimizes false positives without unduly sacrificing true positives [ 51 ]. P values were adjusted for the false discovery rate (FDR) utilizing the Benjamini-Hochberg (BH) method in the R package stats (v4·1·1) [ 52 ]. Volcano plots were generated using the log 2 fold change. Significance was defined as adjusted P values below 1E−2. There was no significance cut-off fold-change. Spearman’s correlation between DEGAS risk scores and gene expression was calculated for each ST prostate using the cor function from the R stats package. This metric was used to easily identify genes that were consistently correlated with risk scores across all samples. The Spearman’s correlation between genes and DEGAS risk scores was calculated in the discovery set, using all ST spots from all four samples, with no filtering by tissue type. Plotting the minimum and maximum of correlation coefficients among the four samples allowed for the identification of genes that were either consistently negative or positively correlated across all samples ( e.g. , negative correlation in all samples, positive correlation in all samples). All P values were adjusted using the FDR-BH method. Functional enrichment analysis of genes Functional enrichment analysis was performed on the most DEGs within the discovery dataset. Up to 100 genes with FDR-adjusted P values smaller than 1E−3 and finite log 2 fold change were input to Toppgene [ 53 ]. The P values reported by Toppgene were also adjusted using the Benjamini-Hochberg FDR method. Enrichment results with high similarity and near-identical gene sets were consolidated to maintain clarity. Validation cohort analysis Integrating ST and SC data to identify cell-type specific differential expression in luminal epithelial cells The RCTD tool [ 54 ] was employed to integrate SC and ST data ( Table 1 ). RCTD decomposes the expression of each ST spot as a linear combination of pre-defined cell types from the reference SC dataset. This tool was applied to the seven ST samples in the validation set. Building on the RCTD results, C-SIDE [ 55 ] was then used to identify cell type-specific differential expression associated with DEGAS risk scores. All tools were used with their default settings. RCTD, specifically, was run in “doublet mode”, assigning up to two cell types to each ST spot. The “full mode” results from RCTD were used to calculate Spearman’s correlations between DEGAS risk scores in normal glands and the enrichment scores for each cell type at each ST spot. All P values were corrected using the FDR-BH multiple testing adjustment. To explore cell type proportion changes in normal glands with varying risk levels, it was necessary to divide these tissues into different risk groups for analysis. The range of DEGAS risk scores was divided into four equal intervals, and glands were named BG rank 1, BG rank 2, BG rank 3, and BG rank 4 in order of increasing risk. This categorization enabled comparison of cell type counts assigned by RCTD doublet mode across the groups. Deep learning image analysis ST data contains both histopathological images and transcriptomic data. The 20× magnification image can be divided into 148 × 148 pixel image patches that correspond to each RNA probe. These patches were used to train two deep learning architectures: a CNN and an autoencoder. Accurate classification of high-risk glands has potential clinical significance, offering a less expensive alternative to ST data for identifying high-risk tumor regions in histopathology. The EfficientNetB4 CNN architecture was used for the image classification task [ 56 ]. This is the state-of-the-art CNN model that was selected for its balance between computational efficiency and accuracy. Images were resized from 148 × 148 to 256 × 256 pixels, as required by this architecture. Training was started with 80 epochs and early stopping based on validation loss, with patients of three epochs to mitigate overfitting. To improve generalizability, in-model data augmentation was achieved using the built-in Keras functions RandomRotation, RandomTranslation, RandomFlip, and RandomContrast sequentially. Initially, EfficientNetB4 weights were locked to ImageNet’s pre-trained values [ 57 ], with training confined to the top layers. After this first phase of training, selective unfreezing of the top 20 layers — excluding BatchNorm layers — was performed to refine the model. The learning rate was reduced to 1E−4 to allow for slower adjustments during this fine-tuning stage. An autoencoder was trained to compress patches from all histological regions. These compressed image representations were used to identify subtle similarities across the images. The AE autoencoder architecture was based on Hou et al. [ 58 ] and implemented in Keras [ 59 ]. The encoding layers were equipped with 100, 120, 240, and 320 kernels of varying sizes, followed by BatchNorm and LeakyReLU activations, with AveragePooling applied after every two convolutional layers. The corresponding decoder reconstructs the image by inversely mirroring the encoder structure. The model employed mean squared error reconstruction loss and an Adam optimizer [ 60 ]. Training was conducted for 300 epochs, incorporating early stopping with a patience of 3 epochs, a batch size of 16, and 10% of the patches reserved for validation. The autoencoder’s bottleneck layer provides a 200,000-feature representation of each image. The dimension of these extremely large vectors was reduced using PCA in Python scikit-learn (v1.1.1) [ 61 ], and two components were retained for K-means clustering and visualization. PCA was used instead of non-linear dimensionality reduction techniques like UMAP and t-SNE because PCA better retains global data structure in low-dimensional embeddings. By retaining global structure of the data, distance in the low-dimensional space better reflects distance in the high dimensional space. Thus, if two image patches are nearby in PCA space, they might also be nearby in high dimensional space. This interpretability allowed the identification of subtle similarities in the complex histological images. Two components were retained for visualization and clustering. K-means clustering was used because of its ability to partition a point cloud into equal pieces. The objective was not to identify discrete clusters but instead to define broad categories of data. Given two PCA axes, K-means of four was selected to provide four groups like a confusion matrix (−PC1 & −PC2; +PC1& −PC2; −PC1 & +PC2; +PC1 & +PC2). IHC Validation Protein-level quantification of MSMB expression and correlation with 5-year distant metastasis Four patients were identified with ISUP Grade 5 prostate cancer who suffered distant metastasis to bone or another organ, and not solely regional lymph nodes or local recurrence. These four patients were matched to four patients with identical ISUP grade and N-stage, using a cancer registry at the Indiana University School of Medicine. The matched control patients did not suffer distant recurrence within five years, and all four matched control patients had more than seven years of follow-up, also without distant recurrence. The ages of patients identified in two outcome groups were compared to evaluate age-related biases. ISUP grading involves combining the two highest Gleason grades of the tumor mass of the prostatectomy specimen. The diagnostic H&E-stained specimens were used to identify the tissue with the largest quantities of the primary and secondary grades. In addition, two sections of normal tissue were selected. One of the pieces of normal tissue was adjacent to the primary tumor mass, and an effort was made to select a piece of normal tissue that was contralateral to the side of the prostate containing the primary tumor mass, and thus, potentially the least affected by local neoplastic inflammation. Selection of all tissues was performed by our expert pathologist. The protein product of the MSMB gene has several aliases, including “PSP-94”. Rabbit monoclonal antibodies for this protein were purchased from Abcam ® (EPR7346, Catalog No. ab133296) for IHC. These antibodies were validated with endometrium serving as a negative control and non-cancerous glandular prostate tissue as a positive control. The dilution of 1:5000 was the least dilute aliquot that had no staining of background stroma and retained strong staining of benign prostatic epithelium. The generated IHC slides of representative sections were digitized with a Leica Aperio CS2 scanner at 40× magnification. The digital WSIs were analyzed with the software QuPath, which enabled us to extract the mean cytoplasmic staining from all cells within the parenchymal tissue. The analysis of this image feature, called IHC intensity herein, is described in the relevant sections of the results. Supplementary Material qzaf119_Supplementary_Data qzaf119_supplementary_data.zip (78.9MB, zip) Acknowledgments This research was supported by the National Institute of General Medical Sciences, USA (Grant No. 1RO1GM148970 to CC, JZ, and TSJ), the National Cancer Institute, USA (Grant No. 1R21CA264339 to JZ and TSJ), the American Cancer Society, USA (Grant No. 19-144-32 to TSJ), and the Indiana CTSI, USA (Grant No. UL1-TR-002529, NIH, S. Moe, S. Wiehe to JLC). We would like to acknowledge the support of the IU School of Medicine Center for Computational Biology and Bioinformatics for their technical expertise and the IU University Information Technology Services for their high-performance computing resources. We would like to thank the IU Cancer Registry for their assistance in identifying prostate cancer patients, the IUH Pathology Laboratory Histotechnology core personnel for their assistance in validating the MSMB antibody, and the laboratory of Dr. George Sandusky for generating whole-slide images. This study was in part supported by Lilly Endowment, Inc., through its support for the Indiana University Pervasive Technology Institute and its high-performance computing infrastructure. Contributor Information Justin L Couetil, Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA; Indiana Clinical and Translational Sciences Institute, Indianapolis, IN 46202, USA. Ziyu Liu, Department of Statistics, Purdue University, West Lafayette, IN 47907, USA. Chao Chen, Department of Biomedical Informatics, Stony Brook University, Stony Brook, NY 11794, USA. Ahmed K Alomari, Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA; Department of Pathology, IU School of Medicine, Indianapolis, IN 46202, USA. Kun Huang, Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA; Department of Biostatistics and Health Data Sciences, IU School of Medicine, Indianapolis, IN 46202, USA; Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA; Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA. Jie Zhang, Department of Medical and Molecular Genetics, IU School of Medicine, Indianapolis, IN 46202, USA; Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA; Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA. Travis S Johnson, Department of Biostatistics and Health Data Sciences, IU School of Medicine, Indianapolis, IN 46202, USA; Melvin and Bren Simon Comprehensive Cancer Center, Indianapolis, IN 46202, USA; Center for Computational Biology and Bioinformatics, IU School of Medicine, Indianapolis, IN 46202, USA; Indiana Biosciences Research Institute, Indianapolis, IN 46202, USA. Ethical statement This research was conducted in accordance with Indiana University School of Medicine and federal research standards, United States (Approval No. IRB 21981). Code availability Code for both DEGAS and deep-learning image analysis in this project is available within the Mendeley repository ( https://data.mendeley.com/ ; DOI: 10.17632/97s9fjn5bb.1). CRediT author statement Justin L. Couetil: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Ziyu Liu: Formal analysis, Methodology, Software, Writing – review & editing. Chao Chen: Funding acquisition, Writing – review & editing. Ahmed K. Alomari: Funding acquisition, Resources, Supervision, Validation, Writing – review & editing. Kun Huang: Funding acquisition, Resources, Supervision, Writing – review & editing. Jie Zhang: Conceptualization, Funding acquisition, Resources, Supervision, Writing – review & editing. Travis S. Johnson: Conceptualization, Data curation, Funding acquisition, Project administration, Resources, Software, Supervision, Writing – review & editing. All authors have read and approved the final manuscript. Competing interests The authors have declared no competing interests. Supplementary material Supplementary material is available at Genomics, Proteomics & Bioinformatics online ( https://doi.org/10.1093/gpbjnl/qzaf119 ). ORCID 0000-0002-9421-2738 (Justin L. Couetil) 0009-0006-5465-3527 (Ziyu Liu) 0000-0003-1703-6483 (Chao Chen) 0000-0002-3776-5053 (Ahmed K. Alomari) 0000-0002-8530-370X (Kun Huang) 0000-0001-6939-7905 (Jie Zhang) 0000-0002-4628-2256 (Travis S. Johnson) References [1]. National Cancer Institute. SEER*Explorer: an interactive website for SEER cancer statistics [Internet]. Bethesda, MD: National Cancer Institute; 2023, https://seer.cancer.gov/statistics-network/explorer/ . [2]. Brawley OW.  Trends in prostate cancer in the United States. J Natl Cancer Inst Monogr  2012;2012:152–6. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [3]. San Francisco IF, Rojas PA, Bravo JC, Díaz J, Ebel L, Urrutia S, et al.  Can we predict prostate cancer metastasis based on biomarkers? Where are we now?  Int J Mol Sci  2023;24:12508. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [4]. Desai MM, Cacciamani GE, Gill K, Zhang J, Liu L, Abreu A, et al.  Trends in incidence of metastatic prostate cancer in the US. JAMA Netw Open  2022;5:e222246. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [5]. U.S. Preventive Services Task Force. Prostate cancer: screening. Final recommendation statement [Internet]. Rockville, MD: U.S. Preventive Services Task Force; 2018, https://www.uspreventiveservicestaskforce.org/uspstf/recommendation/prostate-cancer-screening . [6]. Fenton JJ, Weyrich MS, Durbin S, Liu Y, Bang H, Melnikow J.  Prostate-specific antigen-based screening for prostate cancer: Evidence report and systematic review for the US Preventive Services Task Force. JAMA  2018;319:1914–31. [ DOI ] [ PubMed ] [ Google Scholar ] [7]. Roobol MJ, Carlsson SV.  Risk stratification in prostate cancer screening. Nat Rev Urol  2013;10:38–48. [ DOI ] [ PubMed ] [ Google Scholar ] [8]. Cyll K, Ersvær E, Vlatkovic L, Pradhan M, Kildal W, Avranden Kjær M, et al.  Tumour heterogeneity poses a significant challenge to cancer biomarker research. Br J Cancer  2017;117:367–75. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [9]. Haffner MC, Zwart W, Roudier MP, True LD, Nelson WG, Epstein JI, et al.  Genomic and phenotypic heterogeneity in prostate cancer. Nat Rev Urol  2021;18:79–92. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [10]. Tuong ZK, Loudon KW, Berry B, Richoz N, Jones J, Tan X, et al.  Resolving the immune landscape of human prostate at a single-cell level in health and cancer. Cell Rep  2021;37:110132. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [11]. Lamouille S, Xu J, Derynck R.  Molecular mechanisms of epithelial-mesenchymal transition. Nat Rev Mol Cell Biol  2014;15:178–96. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [12]. Lähnemann D, Köster J, Szczurek E, McCarthy DJ, Hicks SC, Robinson MD, et al.  Eleven grand challenges in single-cell data science. Genome Biol  2020;21:31. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [13]. Burdziak C, Alonso-Curbelo D, Walle T, Reyes J, Barriga FM, Haviv D, et al.  Epigenetic plasticity cooperates with cell-cell interactions to direct pancreatic tumorigenesis. Science  2023;380:eadd5327. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [14]. Hegde S, Leader AM, Merad M.  MDSC: markers, development, states, and unaddressed complexity. Immunity  2021;54:875–84. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [15]. Veltri RW, Khan MA, Miller MC, Epstein JI, Mangold LA, Walsh PC, et al.  Ability to predict metastasis based on pathology findings and alterations in nuclear structure of normal-appearing and cancer peripheral zone epithelium in the prostate. Clin Cancer Res  2004;10:3465–73. [ DOI ] [ PubMed ] [ Google Scholar ] [16]. Johnson TS, Yu CY, Huang Z, Xu S, Wang T, Dong C, et al.  Diagnostic Evidence GAuge of Single cells (DEGAS): a flexible deep transfer learning framework for prioritizing cells in relation to disease. Genome Med  2022;14:11. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [17]. Slaughter DP, Southwick HW, Smejkal W.  Field cancerization in oral stratified squamous epithelium; clinical implications of multicentric origin. Cancer  1953;6:963–8. [ DOI ] [ PubMed ] [ Google Scholar ] [18]. Baker KT, Salk JJ, Brentnall TA, Risques RA.  Precancer in ulcerative colitis: the role of the field effect and its clinical implications. Carcinogenesis  2018;39:11–20. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [19]. Lochhead P, Chan AT, Nishihara R, Fuchs CS, Beck AH, Giovannucci E, et al.  Etiologic field effect: reappraisal of the field effect concept in cancer predisposition and progression. Mod Pathol  2015;28:14–29. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [20]. Risk MC, Knudsen BS, Coleman I, Dumpit RF, Kristal AR, LeMeur N, et al.  Differential gene expression in benign prostate epithelium of men with and without prostate cancer: evidence for a prostate cancer field effect. Clin Cancer Res  2010;16:5414–23. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [21]. Chai H, Brown RE.  Field effect in cancer — an update. Ann Clin Lab Sci  2009;39:331–7. [ PubMed ] [ Google Scholar ] [22]. von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP.  The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. J Clin Epidemiol  2008;61:344–9. [ DOI ] [ PubMed ] [ Google Scholar ] [23]. Iczkowski KA. Prostate gland & seminal vesicles: nonneoplastic metaplasia [Internet]. PathologyOutlines.com; 2016, https://www.pathologyoutlines.com/topic/prostateurothelialmet.html . [24]. Chan PS, Chan LW, Xuan JW, Chin JL, Choi HL, Chan FL.  In situ hybridization study of PSP94 (prostatic secretory protein of 94 amino acids) expression in human prostates. Prostate  1999;41:99–109. [ DOI ] [ PubMed ] [ Google Scholar ] [25]. Grzmil M, Hemmings BA.  Translation regulation as a therapeutic target in cancer. Cancer Res  2012;72:3891–900. [ DOI ] [ PubMed ] [ Google Scholar ] [26]. Ding Y, Lee M, Gao Y, Bu P, Coarfa C, Miles B, et al.  Neuropeptide Y nerve paracrine regulation of prostate cancer oncogenesis and therapy resistance. Prostate  2021;81:58–71. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [27]. Bianchi M, Crinelli R, Giacomini E, Carloni E, Radici L, Scarpa ES, et al.  A negative feedback mechanism links UBC gene expression to ubiquitin levels by affecting RNA splicing rather than transcription. Sci Rep 2019;9:18556. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [28]. Erickson A, He M, Berglund E, Marklund M, Mirzazadeh R, Schultz N, et al.  Spatially resolved clonal copy number alterations in benign and malignant tissue. Nature  2022;608:360–7. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [29]. Song H, Weinstein HNW, Allegakoen P, Wadsworth MH 2nd, Xie J, Yang H, et al.  Single-cell analysis of human primary prostate cancer reveals the heterogeneity of tumor-associated epithelial cell states. Nat Commun  2022;13:141. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [30]. Regitnig P, Reiner A, Dinges HP, Höfler G, Müller-Holzner E, Lax SF, et al.  Quality assurance for detection of estrogen and progesterone receptors by immunohistochemistry in Austrian pathology laboratories. Virchows Arch  2002;441:328–34. [ DOI ] [ PubMed ] [ Google Scholar ] [31]. College of American Pathologists. Template for reporting results of biomarker testing of specimens from patients with carcinoma of the breast [Internet]. Northfield, IL: College of American Pathologists; 2018, https://documents.cap.org/protocols/cp-breast-18biomarker-1201.pdf . [32]. Yi M, Huo L, Koenig KB, Mittendorf EA, Meric-Bernstam F, Kuerer HM, et al.  Which threshold for ER positivity? A retrospective study based on 9639 patients. Ann Oncol  2014;25:1004–11. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [33]. Aran D, Camarda R, Odegaard J, Paik H, Oskotsky B, Krings G, et al.  Comprehensive analysis of normal adjacent to tumor transcriptomes. Nat Commun  2017;8:1077. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [34]. Aran D, Sirota M, Butte AJ.  Systematic pan-cancer analysis of tumour purity. Nat Commun  2015;6:8971. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [35]. Waters KM, Le Marchand L, Kolonel LN, Monroe KR, Stram DO, Henderson BE, et al.  Generalizability of associations from prostate cancer genome-wide association studies in multiple populations. Cancer Epidemiol Biomarkers Prev  2009;18:1285–9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [36]. Wang X, Hayes JE, Xu X, Gao X, Mehta D, Lilja HG, et al.  Validation of prostate cancer risk variants rs10993994 and rs7098889 by CRISPR/Cas9 mediated genome editing. Gene  2021;768:145265. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [37]. Anklesaria JH, Mhatre DR, Mahale SD.  Structural and molecular biology of PSP94: its significance in prostate pathophysiology. Front Biosci (Landmark Ed)  2018;23:535–62. [ DOI ] [ PubMed ] [ Google Scholar ] [38]. Ou T, Zhou Z, Turner DP, Zhu B, Lilly M, Jiang W.  Increased preoperative plasma level of microbial 16S rDNA translocation is associated with relapse after prostatectomy in prostate cancer patients. Front Oncol  2019;9:1532. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [39]. Gurel B, Lucia MS, Thompson IM Jr, Goodman PJ, Tangen CM, Kristal AR, et al.  Chronic inflammation in benign prostate tissue is associated with high-grade prostate cancer in the placebo arm of the prostate cancer prevention trial. Cancer Epidemiol Biomarkers Prev  2014;23:847–56. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [40]. Källberg E, Vogl T, Liberg D, Olsson A, Björk P, Wikström P, et al.  S100A9 interaction with TLR4 promotes tumor growth. PLoS One  2012;7:e34207. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [41]. Tidehag V, Hammarsten P, Egevad L, Granfors T, Stattin P, Leanderson T, et al.  High density of S100A9 positive inflammatory cells in prostate cancer stroma is associated with poor outcome. Eur J Cancer  2014;50:1829–35. [ DOI ] [ PubMed ] [ Google Scholar ] [42]. Bergström SH, Järemo H, Nilsson M, Adamo HH, Bergh A.  Prostate tumors downregulate microseminoprotein-beta (MSMB) in the surrounding benign prostate epithelium and this response is associated with tumor aggressiveness. Prostate  2018;78:257–65. [ DOI ] [ PubMed ] [ Google Scholar ] [43]. Haffner MC, Mosbruger T, Esopi DM, Fedor H, Heaphy CM, Walker DA, et al.  Tracking the clonal origin of lethal prostate cancer. J Clin Invest  2013;123:4918–22. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [44]. Chowdhury-Paulino IM, Ericsson C, Vince R Jr, Spratt DE, George DJ, Mucci LA.  Racial disparities in prostate cancer among black men: epidemiology and outcomes. Prostate Cancer Prostatic Dis  2022;25:397–402. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [45]. Okobia MN, Zmuda JM, Ferrell RE, Patrick AL, Bunker CH.  Chromosome 8q24 variants are associated with prostate cancer risk in a high risk population of African ancestry. Prostate  2011;71:1054–63. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [46]. Robbins CM, Hooker S, Kittles RA, Carpten JD.  EphB2 SNPs and sporadic prostate cancer risk in African American men. PLoS One  2011;6:e19494. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [47]. Abadi M, Barham P, Chen J, Chen Z, Davis A, Dean J, et al. TensorFlow: a system for large-scale machine learning. Proceedings of the 12th USENIX Conference on Operating Systems Design and Implementation 2016:265–83. [ Google Scholar ] [48]. The R Core Team. R: a language and environment for statistical computing. 2021, https://www.r-project.org/ . [49]. Alsaleh L, Li C, Couetil JL, Ye Z, Huang K, Zhang J, et al.  Spatial transcriptomic analysis reveals associations between genes and cellular topology in breast and prostate cancers. Cancers (Basel)  2022;14:4856. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [50]. Dietterich TG, editor. Ensemble methods in machine learning. In: Multiple Classifier Systems. Lecture Notes in Computer Science, vol 1857. Berlin Heidelberg: Springer; 2000, p.1–15. [ Google Scholar ] [51]. Li Y, Ge X, Peng F, Li W, Li JJ.  Exaggerated false positives by popular differential expression methods when analyzing human population samples. Genome Biol  2022;23:79. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [52]. The R Core Team. stats: the R stats package. 2023, https://www.r-project.org/ . [53]. Chen J, Bardes EE, Aronow BJ, Jegga AG.  ToppGene Suite for gene list enrichment analysis and candidate gene prioritization. Nucleic Acids Res  2009;37:W305–11. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [54]. Cable DM, Murray E, Zou LS, Goeva A, Macosko EZ, Chen F, et al.  Robust decomposition of cell type mixtures in spatial transcriptomics. Nat Biotechnol  2022;40:517–26. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [55]. Cable DM, Murray E, Shanmugam V, Zhang S, Zou LS, Diao M, et al.  Cell type-specific inference of differential expression in spatial transcriptomics. Nat Methods  2022;19:1076–87. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [56]. Tan M, Le QV. Efficientnet: rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning 2019:6105–14. [ Google Scholar ] [57]. Deng J, Dong W, Socher R, Li LJ, Kai L, Li FF. ImageNet: a large-scale hierarchical image database. Proceedings of the 2009 IEEE Conference on Computer Vision and Pattern Recognition 2009:248–55. [ Google Scholar ] [58]. Hou L, Nguyen V, Kanevsky AB, Samaras D, Kurc TM, Zhao T, et al.  Sparse autoencoder for unsupervised nucleus detection and representation in histopathology images. Pattern Recognit  2019;86:188–200. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] [59]. Chollet Fao. Keras. 2015, https://keras.io . [60]. Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv 2017;1412.6980. [61]. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al.  Scikit-learn: machine learning in Python. J Mach Learn Res  2011;12:2825–30. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials qzaf119_Supplementary_Data qzaf119_supplementary_data.zip (78.9MB, zip) Articles from Genomics, Proteomics & Bioinformatics are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (14.1 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• HyLnc: a hybrid deep learning and feature-based approach for long non-coding RNA prediction.#976662
• Leveraging proteomics and machine learning for mechanism and biomarker discovery on glioma progression and transformation: from LGG to GBM.#30891
• RegNetAgents: A Multi-Agent Framework for Cross-Network Regulatory Driver Identification in Cancer Genomics#1110298
• DORA: a dose-response autoencoder for interpretable transcriptome-to-viability prediction#314152
• Advancements in brain tumor analysis: a comprehensive review of machine learning, hybrid deep learning, and transfer learning approaches for MRI-based classification and segmentation.#489495