Single-cell Spatial Proteomic Revelations on the Multiparametric MRI Heterogeneity of Clinically Significant Prostate Cancer

As prostate cancer screening with PSA has come under considerable controversy due to its poor diagnostic accuracy (1), the use of prostate multiparametric MRI (mpMRI) has rapidly increased. Owing to improvements in technology, discrete intraprostatic lesions can be visualized using mpMRI (2–5). The interpretation of these prostate mpMRI images and the degree of suspicion for prostate cancer has been standardized with the Prostate Imaging-Reporting and Data System (PI-RADS; ref. 6). However, false-positive (e.g., PI-RADS classification 4–5 with benign biopsy) and false-negative [e.g., PI-RADS classification 1–2 with clinically significant prostate cancer (csPCa)] prostate mpMRI are not uncommon (7, 8). Although a number of confounders of accurate prostate mpMRI interpretation exist (9), and some of the false-positive and false-negative mpMRI results may be attributable to radiologist variability in interpretation (10, 11), several prostate cancer tumors are recognized as inherently undetectable on mpMRI (e.g., “mpMRI-invisible”).

The potential link between mpMRI visibility and oncologic outcomes is only recently being explored (12–15), and due to the relatively contemporary use of mpMRI, intermediate and longer term oncologic outcomes are very limited. Although the molecular features of mpMRI visibility have been reported previously (12, 16, 17), the underlying molecular and cellular mechanisms connecting mpMRI visibility with clinical outcome have yet to be determined. Moreover, inflammation has been associated with MRI visibility and false-positive detection by mpMRI as inflammation mimics prostate cancer (18). To better characterize prostate cancer mpMRI visibility, we molecularly and microexamined tissue from clinically matched patients with mpMRI-visible and mpMRI-invisible prostate cancer who underwent radical prostatectomy using multiplex immunofluorescence (MxIF) and gene expression profiling. By marrying MxIF with transcriptomics, we delved into the role that spatial architecture and the tumor microenvironment play in prostate cancer mpMRI visibility at single-cell resolution.

Here, we show that mpMRI-visible prostate cancer tumors display a more complex tumor architecture compared with mpMRI-invisible tumors. Major differences were discovered in stromal organization within the malignant zones between mpMRI-invisible and mpMRI-visible prostate cancer tumors, whereas no significant correlation between key immunologic features and mpMRI visibility was identified. Stroma architecture similarities were found between mpMRI-invisible, but not mpMRI-visible prostate cancer, and normal prostate tissue. Molecularly, mpMRI-invisible tumors were enriched for gene expression related to the stroma and cell adhesion, approximating the stromal gene expression profile of normal prostate tissue. Prostate cancer The Cancer Genome Atlas (TCGA) analysis showed that prostate cancer carrying the stromal signature enriched in mpMRI-invisible prostate cancer exhibited better clinical outcome. Interestingly, mpMRI-visible prostate cancer experienced more recurrence within our small cohort.

This study uses MxIF imaging analyzed by a unique artificial intelligence (AI)–based segmentation pipeline in conjunction with gene expression analysis to determine the underlying cellular, molecular, and structural reasons for prostate cancer mpMRI invisibility. We collected prostate cancer tissue (only one lesion observed per patient) from clinically and morphologically matched patients (≥ Gleason score 7) with mpMRI-visible (n = 8) and mpMRI-invisible prostate cancer (n = 8) who underwent radical prostatectomy and only had one mpMRI prior to prostatectomy. These tissue samples were collected following the ethical guidelines pertaining to the Declaration of Helsinki after the provision of written informed consent and approval by the Institutional Review Board at Washington University in St. Louis (St. Louis, MO). Then, 14 prostate cancer tissues of the 16 prostate cancer tissues (n = 7 mpMRI-visible and n = 7 mpMRI-invisible) were subjected to MxIF using a 14-marker panel, and the images were analyzed using a customized AI-segmentation pipeline to create images and masks. No statistical methods were used to predetermine the sample size, and the samples were chosen based on availability, known visibility status, and with significant follow-up time to determine oncologic outcome. The overall analysis consisted of the calculation of malignant cell density, the number of malignant cell neighbors, and the differences between stromal distribution between the malignant and nonmalignant compartments (e.g., Wasserstein distance), immune cell composition, and correlation with patient relapse. The same resected prostate cancer tissue (n = 8 mpMRI-invisible and n = 8 mpMRI-visible) also underwent gene expression analysis using the HTG EdgeSeq Oncology Biomarker Panel (2,549 genes; HTG Molecular Diagnostics, Inc.) to determine potential underlying molecular differences between visible and invisible tumors. To confirm this analysis, publicly available TCGA gene expression data of normal prostate tissue plus prostate cancer tissue were assessed to determine similarities between invisible prostate cancer tissue and normal prostate tissue. Prostate cancer patient relapse and clinical outcomes were correlated within our dataset as well as with TCGA dataset.

Fresh prostate tissue specimens were carefully handled and fixed after dissection with 10% formalin for 48 hours at room temperature. Tissue specimens were dehydrated by immersion in a series of ethanol solutions of increasing concentration until water-free ethanol: 70% (1×), 80% (1×), 95% (1×), 100% (2×) for 1 hour each. The specimens were then immersed in fresh xylene (or xylene substitute) three times for 1.5 hours each and in paraffin wax (58°C–60°C) twice for 2 hours each. Next, the tissue specimens were embedded into paraffin blocks.

Tissue sections (5 μm) were cut from formalin-fixed paraffin-embedded prostate cancer tissue blocks using a microtome and mounted onto Superfrost Ultra Plus adhesion slides (Thermo Fisher Scientific) for morphologic examination and analysis. Slide tissue sections were baked at 60°C in an oven for 1 hour. Tissue sections were deparaffinized with two washes of fresh xylene and were then rehydrated with ethanol 100% (2×), 95% (2×), 70% (2×), 50% (2×), 1× PBS (1×) and 0.3% Triton X100 in 1× PBS (1x) washes and subjected to a two-step antigen retrieval process. Next, the tissue sections underwent repeated cycles of staining, imaging, and signal removal. The sections were stained with antibodies directly conjugated with either Cy3 or Cy5 dye at a previously optimized concentration (Supplementary Table S2). All antibody mixes used for seven incubation rounds were incubated at room temperature for 1 hour in a humid chamber. Slides were washed in 1× PBS for 5 minutes (3×). The tissue sections were then stained with DAPI solution (1 μg/mL) for 15 minutes. The slides were washed with 1× PBS, and the coverslip was immediately added using mounting media.

Immunofluorescence confocal imaging of human prostate tissue slides was performed using an IN Cell Analyzer 2200 equipped with 20× objective (GE Healthcare). Images were captured with a CMOS camera. MxIF imaging was performed with high-efficiency fluorochrome-specific filter sets (DAPI, Cy3, Cy5). Image processing and deconvolution were performed with NIS-Elements imaging software (Nikon Instruments, Inc).

After preprocessing, a large image dataset was obtained with different shapes but at the same resolution and at 16-bit depth. The images were cropped into small regions of the same size, while maintaining marker names and numbering of regions for the purpose of original image restoration at the final steps.

The creation of stroma and endothelial masks enabled the determination of the degree of tissue transformation and vascularization. To create the binary masks, PCK26 (pan-cytokeratin), S6 (stroma), and CD31 (blood vessels) markers were utilized. All processing steps used the OpenCV library in Python language. For the stroma, mean tissue sections were analyzed between the epithelial compartments (normal and malignant), including smooth muscle cells, fibrosis, immune infiltrates, and others. The first step included thresholding, binarization, filling holes, and inversion of the PCK26 marker image, which aided in the identification of nonepithelial compartments. The S6 marker image was used for correction. S6 marker analysis was performed for every cell on each slide for the identification of “black holes” for exclusion in the stroma mask. If starches or secretion products obscured the imaging process, the masks were manually corrected with ImageJ (19). Vascular masks were obtained employing the same process.

Cell segmentation was performed using the U-Net semantic segmentation neural network (20) and watershed post-processing of the identified cell masks to reduce the undersegmented cell count. To instill the correct logic for the cell segmentation neural network, two markers were utilized in the training set as follows: (i) a region was designated a cell only in the presence of one nucleus; (ii) a closed NAKATPASE border around the nucleus was designated as a membrane; (iii) if no NAKATPASE border was not found, the cell membrane was defined as an area at a distance up to 15 pixels from the nucleus, depending on the proximity of neighbors.

Manually segmented cell masks generated for the training set were supplemented with cell masks created in CellProfiler (Broad Institute of MIT and Harvard in Cambridge, MA; https://personal.broadinstitute.org/anne/). The training data were generated with DAPI as a nucleus marker and NAKATPASE for cell membrane identification. Next, border masks were predicted in parallel, and border masks were excluded from the cell masks. Finally, the training data consisted of three image types: (i) fluorescent images combined to a two-channel array, (ii) cells, and (iii) cell contact border masks. Every image was cropped to 224 × 224 and augmented by a random rotation of 0 to 270 degrees, with reflection about the vertical and horizontal axes. Training images were normalized by subtracting the average pixel value over the channel divided by the SD.

U-Net architecture, which consists of encoders and decoders, each with 4-layer depth, was used. Each layer included 2-fold convolution with a kernel size of 3 × 3, padding 1 and stride 1, processed through batch normalization and ReLU activation function. Between the layers, max pooling was performed with a kernel size of 2 × 2 in an encoder and bilinear upsampling in each decoder.

The data were divided into training (85%) and test (15%) sets. All of the training processes were performed with the NVIDIA Tesla P100 platform (NVIDIA Corp), with batch size at 25, and the run at 100 epochs. Dice loss/intersection was preferred over union, a parameter used to calculate error, in combination with binary cross-entropy as the loss function and Adam as the optimization algorithm (21), with the learning rate set to 0.01.

The trained network was run on the images of all patients in the cohort. The images were normalized using the same technique as the training images. Masks, predicted by convolution neural network, were processed through the watershed function and split undivided cells. For each cell, a mask was obtained, and, for each marker, the mean intensity per cell was computed. Values were divided by possible maximum for 16-bit pixel range and the arcsinh was transformed to perform clustering.

Cell subtype assignment was performed using FlowSOM (22) cell clustering using mean marker intensity in a given cell mask. Cell subtype assignment was guided by tSNE projection plots, allowing similar cell subpopulation clustering, resulting in 20 clusters. The clusters were then classified with a marker signature into the following seven types: CD8⁺CD45⁺ cytotoxic T cells; CD4⁺CD45⁺ T-cell helpers; CD4⁺FoxP3⁺ regulatory T cells (Treg); CD68⁺ macrophages; PCK26⁺ epithelial cells; PCK26⁺KI67⁺ malignant cells in mitosis; and unclassified (every other).

To determine the differences between stromal distribution in nonmalignant and tumor zones, each zone was divided into small squares of the same size, and the percentage of the stromal mask in each square was calculated. Square size was chosen to maintain maximum variability of values to distinguish areas with large and dense stromal elements from rarefied elements, enabling the formation of a uniform thin web on the slide. Next, the histograms of two distributions were compared in different zones with Wasserstein distance using custom codes developed in Python using OpenCV, SciPy libraries, and others.

RNA library preparation and expression analysis was performed by HTG Molecular Diagnostics, Inc. using their DESeq2 Pipeline (23). Resulting expression was normalized to one million. The expression of 556 TCGA samples (502 tumor, 54 normal) were calculated using kallisto v0.42.4 with GENCODE v23 transcripts. Only protein coding transcripts were left for normalization, excluding IGH/K/L and TCR transcripts, mitochondrial transcripts and histones, resulting in 20,062 protein-coding transcripts to analyze. We transformed gene expression values as log₂(1+TPM).

All data associated with this study are present in the article or Supplementary Materials and Methods.

To uncover the differences in tumor architecture between mpMRI-invisible and mpMRI-visible prostate cancer tumors collected from matched patients (clinically and pathologically) with clinically significant prostate cancer (≥ Gleason score 7) with one single lesion who underwent radical prostatectomy (Fig. 1A and B; Supplementary Table S1), we employed an AI-based segmentation platform for MxIF to analyze the entire processed tissue slide (n = 14, seven patients each group), approximating 70 regions of interest (ROI)/slide, demonstrating the ability of this pipeline to robustly process and examine entire tumor tissue (Fig. 2A). The MxIF analysis comprised 14 different markers (Supplementary Table S2), including the epithelial cell markers PCK26 and NaKATPase, the endothelial cell marker CD31, the commonly used T-cell markers CD4 and CD8, FOXP3 to identify Tregs, and the macrophage marker CD68. This AI-based segmentation pipeline enabled the comprehensive and automated evaluation of prostate cancer tumors, including visualization at single-cell resolution, segmentation, cell typing, modeling of tumor architecture and spatial interactions, and the employment of software-generated masks (Fig. 2B). On the basis of the utilization of these markers and this AI-based segmentation pipeline, the single-cell spatial architecture was investigated to reveal the molecular, cellular, and structural mechanisms that may render prostate cancer tumors invisible to mpMRI.

As shown in Fig. 2C–E, automated cell typing was performed for every mpMRI-invisible and mpMRI-visible prostate cancer tumor (n = 7 each). A pathologist review of every tumor ROI led to each ROI annotated as “glandular” or “compact and complex.” Each tumor included eight to 50 (mean = 23) tumor ROIs, with the resulting content of the two types of tumor structures shown in Fig. 2F. mpMRI-visible prostate cancer tumors had a higher content of tumor ROIs classified as complex tumor structures, characterized by a lack of clear boundaries between the stromal and glandular components. In contrast, mpMRI-invisible prostate cancer tumors were characterized by tumor ROIs with rounded glandular structures, denoted by relatively clear and mostly uniform contours and less closely spaced glands interspersed with stroma (Fig. 2F). Two patients of seven with mpMRI-invisible tumors possessed small areas (< 30% of tumor ROIs) with compact and complex tumor architecture, whereas five patients of seven with mpMRI-visible tumors had large areas in their tumors with compact and complex tumor architecture (range, 40%–100% of tumor ROIs; Fig. 2F). This analysis demonstrates that the application of comprehensive imaging using MxIF coupled with a bioinformatics pipeline enabled the accurate measure of malignant cell density and cell–cell contacts. Next, malignant cell density and spatial neighboring (Fig. 3A) were examined to determine whether these parameters may explain mpMRI invisibility. Malignant cell density was calculated as the total number of malignant cells divided by tumor mask area. The mpMRI-visible prostate cancer tumors had increased neighboring malignant cells quantified as the number of malignant cells in close proximity to one malignant cell within an 80 μm radius (Fig. 3B; P = 0.06) and higher malignant cell density (Fig. 3B; P = 0.1). The number of neighboring malignant cells strongly correlated with tumor cell density (Fig. 3C; Spearman r_s = 0.87, P < 0.001 for both plots). The density and complexity of the prostate cancer tumor architecture may underlie their detection through mpMRI. Moreover, mpMRI-visible patients with prostate cancer were found to have a higher rate of biochemical disease recurrence compared with the mpMRI-invisible patients with prostate cancer (71% vs. 0%, P = 0.03) with a mean follow-up of almost 5 years. Higher malignant cell density and neighboring malignant cells, two mpMRI visibility characteristics, were associated with patient relapse (Fig. 3C), suggesting that imaging visibility, specifically PI-RADS 5 visibility, may be linked to poor prostate cancer clinical outcome.

The stromal architecture in the malignant and nonmalignant (normal) compartments were evaluated for each tumor as the malignant stroma is known to impact prostate cancer aggressiveness (Fig. 3D, representative image), demonstrating the influence of the prostate cancer microenvironment on tumor phenotype. Overall, 62% of mpMRI-visible prostate cancer tumors (n = 7) showed more compact and complex tumor architecture represented by tumor acini fusion into poorly formed glandular structures without intervening stroma. However, the stromal mask area in the malignant and nonmalignant compartments showed no significant differences between mpMRI-invisible and mpMRI-visible prostate cancer, suggesting similarities in stroma organization between the two groups (Fig. 3E). To further evaluate stromal organization, Wasserstein distances (WD), also known as the Earth Mover's Distance (EMD), which provides the measure of similarity between stromal organization in the nonmalignant (normal) and malignant compartments (24), were examined (Supplementary Fig. S1). The distribution of stromal cells in the malignant and nonmalignant zones of the mpMRI-invisible and mpMRI-visible tumors was assessed using stromal mask area. To avoid sample heterogeneity, we compared individual distribution in the malignant zone with distribution in the normal zone, averaged across all samples. Notable differences in stromal distributions in the malignant zones were observed between the two groups (Supplementary Fig. S1). Moreover, the mpMRI-invisible prostate cancer stroma appeared more similar to normal tissue as denoted by the lower WD values of the mpMRI-invisible tissue. To neutralize the contribution of the individual characteristics of each patient, the WD between distributions of the stroma in the tumor and nonmalignant zones, a reflection of the degree of stromal deformation during malignancy, was calculated. For mpMRI-invisible prostate cancer cases, the mean WD between the tumor and normal regions approximated 0.15 (Supplementary Fig. S2A), and in the mpMRI-visible prostate cancer cases, the mean WD was approximately 0.21 (P = 0.3; Supplementary Fig. S2A), suggesting a greater difference in stromal transformation during malignancy. Samples with higher WD values and/or higher malignant cell density were more often mpMRI visible and prostate cancer recurrent, with only nonrelapsed prostate cancer present in the bottom left quadrant (Supplementary Fig. S2B). No correlation was observed between WD values and malignant cell density (Spearman r_s = 0.03, P = 0.9). These data suggest a role for prostate cancer stromal organization in mpMRI detection based on the ability of the stroma and microenvironment to influence tumor phenotype.

Immune composition of the mpMRI-invisible and mpMRI-visible prostate cancer tumors was examined to identify variations in the immune microenvironments and inflammation. Specifically, using the MxIF markers, the cell types and their locations within the mpMRI-invisible and mpMRI-visible prostate cancer tumors were determined (Fig. 4A and B). The cell populations were visualized using tSNE plots across the prostate cancer tumors, with confirmation via MxIF imaging (Fig. 4B and C). MxIF visualization enabled the identification of the cell composition of the prostate cancer tumors in nonmalignant and malignant zones, including the epithelium, with a representative tumor image and magnification of its various subcompartments (Fig. 5A). To account for each sample possessing different sizes of malignant and nonmalignant (normal) areas, with different cell numbers/area, cell count was normalized by zone area to compare cell densities. Cell typing revealed a higher density of immune cells, especially CD8 T cells, in the normal epithelium versus the malignant epithelium (Fig. 5B; Supplementary Fig. S3A). Notably, the malignant epithelium of all the prostate cancer tumors lacked immune cells (Fig. 5B), which may underlie the moderate success of immune checkpoint blockade in prostate cancer (25). Prostate cancer is also known to have low tumor mutational burden, directly reducing neoantigen expression (26), rendering the prostate cancer tumors cold and impenetrable to immunotherapies. Moreover, mpMRI-visible prostate cancer tumors displayed a higher content of CD4 T cells in the malignant stroma of mpMRI-visible prostate cancer (Fig. 5C), and the content of stromal cells [i.e., potential fibroblasts, including cancer-associated fibroblasts (CAF)] was higher in the malignant epithelium and stroma of mpMRI-invisible prostate cancer tumors (Fig. 5B; Supplementary Fig. S3A). Immune infiltration did not correlate with mpMRI visibility or prostate cancer recurrence, suggesting that the other molecular and cellular characteristics such as stromal organization, higher WD, and the number of malignant cells are better indicators of mpMRI visibility. However, slightly elevated levels of macrophages and slightly lower levels of stromal cells (i.e., potential fibroblasts and CAFs) were observed across the relapsed prostate cancer versus nonrelapsed tumors (Fig. 5D; Supplementary Fig. S3A).

Gene expression analysis was performed on the mpMRI-invisible and mpMRI-visible prostate cancer tumors with the HTG EdgeSeq Oncology Biomarker Panel. In addition to the 14 prostate cancer tumor samples with MxIF data, we also had the expression data for one mpMRI-invisible (8394) and one mpMRI-visible (8403) sample, totaling 16 prostate cancer tumors with expression data. Overall, 24 genes were identified as the most differentially expressed genes (DEG) between the mpMRI-invisible and mpMRI-visible groups, with a fold change of ≥ 2 or ≤ −2, P < 0.05 and the FDR (q-value) ≤ 0.5 (Fig. 6A). Of these DEGs, filamin C (FLNC) (FDR < 0.1) was the most differentially expressed gene, with significant downregulation in the mpMRI-visible prostate cancer tumors (Fig. 6A–C). FLNC encodes an actin-binding protein, is an important part of the cytoskeleton and plays a role in cell migration in prostate cancer (27). In addition, FLNC is highly expressed in fibroblasts, muscle cells, and endothelium and is not expressed in immune cells (Supplementary Fig. S4A). Changes in FLNC expression have been associated with clinical outcomes in various cancers such as prostate cancer, gastric cancer, hepatocellular carcinoma, and glioblastoma multiforme (28, 29, 30). mpMRI-invisible tumors showed significantly higher expression of FLNC (FDR = 0.018) and other cellular adhesion-related genes, including the integrin-encoding genes ITGA9 (FDR = 0.19) and ITGA7 (FDR = 0.2) as well as DES (desmin) (FDR = 0.2; Fig. 6A). Also, DNAJB5 (FDR = 0.03) had significantly lower expression in the mpMRI-visible prostate cancer tumors (Fig. 6A and B). DNAJB5 belongs to the large and diverse heatshock protein (HSP) family whose members act as tumor promoters or suppressors (31). We next confirmed whether the selected genes were significantly enriched in biologically relevant pathways. Cell adhesion pathways, including focal adhesion and integrin-related pathways, were among the most enriched pathways in the mpMRI-invisible group (Table 1).

TCGA dataset comprising normal prostate (n = 52) and prostate cancer (n = 504) tissue was analyzed to assess our findings in a larger cohort. The signature scores of the expression of the 24 stromal genes (Fig. 6A) were calculated, and the prostate cancer tissue samples were divided into “high” and “low” stromal signature groups by medium value (Fig. 6D). The expression profiles of the normal prostate and prostate cancer tissue were compared, and the normal prostate tissue displayed the high stromal gene expression signature in our cohort as also evidenced by higher FLNC expression (Fig. 6E), suggesting that the mpMRI-invisible prostate cancer tissue more closely resembles normal prostate tissue at the gene expression and molecular levels. The enriched expression of FLNC and other cellular adhesion-related genes supports the spatial imaging findings where stromal organization of mpMRI-invisible prostate cancer tissue resembled normal prostate tissue. Moreover, the high stromal gene expression tissues, which are similar to mpMRI-invisible prostate cancer tissue molecularly, had a higher CAF gene signature score in the tumor region compared with the low stromal gene expression signature prostate cancer tissue (P < 0.001; Supplementary Fig. S4B). CAFs play complex roles in prostate cancer, and while most data suggest they play a role in promotion of tumorigenesis, they have been shown to inhibit tumor growth in certain settings (32–34). To further confirm our findings, we assessed the histology of two randomly selected high stromal gene expression signature prostate cancer tissues compared with two randomly selected low stromal gene expression signature prostate cancer tissues from TCGA (Fig. 6F). The high stromal gene expression tissues had a greater area of fibrosis, their fibrous connective tissue was evenly distributed, and the tumor cells mostly had a single-layer epithelium organization (Fig. 6F). Importantly, FLNC was previously identified as one of seven genes that significantly correlated with biochemical recurrence (BCR)-free survival in two human prostate cancer cohorts (n = 188 total), with higher expression of FLNC correlating with significantly improved BCR-free survival (35). Prostate cancer tumors expressing the high stromal signature [i.e., the expression of the 24 genes resembling mpMRI-invisible prostate cancer tissue (Fig. 6A)], had better progression-free survival (PFS) and BCR-free survival compared with low signature samples (P = 0.002 and P = 0.2, respectively; Fig. 6G). Prostate cancer tumors with higher FLNC expression, which would be suggestive of mpMRI-invisible prostate cancer, had superior PFS compared with prostate cancer tumors with lower FLNC expression (Supplementary Fig. S4C; P = 0.02). Indeed, higher FLNC expression has been linked to better recurrence-free survival in prostate cancer (35, 36), further supporting the notion that mpMRI-invisible prostate cancer has better clinical outcomes compared with mpMRI-visible prostate cancer. Finally, integrative analysis of FLNC expression and with spatial organization represented by malignant cell density yielded the best separation of samples based on mpMRI visibility and prostate cancer recurrence (Fig. 6H; 14 samples with both expression and MxIF data are shown).

mpMRI has become an integral imaging tool for prostate cancer; however, prostate cancer invisibility poses a diagnostic challenge. Intriguingly, mpMRI invisibility has been reported to carry a high negative predictive value for csPCa (37). The structural/molecular characteristics underlying mpMRI heterogeneity are unclear, and the value of mpMRI imaging properties beyond diagnosis remain unknown (38). Therefore, we sought to determine the underlying differences between mpMRI-invisible and mpMRI-visible prostate cancer, potentially impacting management strategies post-mpMRI. To our knowledge, our studies here represent the first integrated multi-omics analyses of clinically and pathologically matched mpMRI-invisible and mpMRI-visible prostate cancer. Specifically, the two groups were relatively well-balanced regarding tumor volume and Gleason pattern 4 variants. All but two samples in the mpMRI-invisible group contained cribiform pattern, often correlated with poorer prognosis (39, 40). Glomeruloid pattern, sometimes thought to be a precursor of cribiform—although with some studies showing correlation with improved clinical outcomes and BCR-free survival (41, 42)—was found in both groups at equivalent incidence (n = 3 in each group). Total extent of tumor has also been associated with poorer clinical outcomes (43). The estimated mean tumor volumes (mpMRI-invisible: 18% ± 12%; mpMRI-invisible: 13% ± 8%) and Gleason pattern 4 (of total tumor volume) were relatively similar in both groups (4.3% vs. 4.8%). Only one sample in the mpMRI-invisible group showed no evidence of any of these patterns (e.g., cribiform, poorly formed, ductal, glomeruloid), suggesting that the presence or absence of these histologic characteristics were not primarily driving mpMRI visibility.

An automated AI-based segmentation pipeline was employed for complete tissue MxIF imaging to extract numerous data points for the accurate determination of the tissue architecture differences between mpMRI-invisible and mpMRI-visible prostate cancer, such as cell positioning and neighboring as well as the software-based generation of masks. These spatial analyses were then supported and confirmed at the transcriptional and histologic levels, showing that multi-omics analyses, including MxIF, transcriptomics, and histology, enabled the deeper understanding of MRI visibility. The AI-based pipeline uncovered that the mpMRI-visible prostate cancer tumors displayed more compact and complex structural features compared with the mpMRI-invisible prostate cancer tumors, which were primarily composed of glandular structures and better resembled normal prostate tissue. PI-RADS is heavily influenced by diffusion-weighted imaging that examines molecule diffusion. Therefore, the compact and complex structural features present in the mpMRI-visible prostate cancer suggest that there is more cellular crowding and less diffusion, enabling detection with mpMRI. Indeed, WD was used to calculate stromal differences between the malignant and nonmalignant compartments of the mpMRI-visible and mpMRI-invisible prostate cancer tumors, and the WDs showed dissimilar stromal organization patterns between the two groups. These noted differences in the physical composition of the prostate cancer tumors may explain why the mpMRI-invisible tumors are rendered undetectable via non-PET conventional imaging.

Underlying these disparities in tumor architecture between the visible and invisible prostate cancer tumors was distinct gene expression profiles. The mpMRI-visible tumors had lower expression levels of stroma-related and cell adhesion–related genes such as FLNC. Interestingly, FLNC expression was high in normal prostate tissue, similarly to mpMRI-invisible tumors, indicating that invisible prostate cancer tumors share both molecular and structural similarities with normal tissue, which may make image detection more difficult. Notably, FLNC may be associated with scarring and fibrosis induced by the stromal response to the presence of the tumor. In prostate cancer with low FLNC expression, a robust anticancer response by the stromal microenvironment may be limited or absent, which may underlie the poor outcomes and inferior survival observed in low FLNC prostate cancer and mpMRI-visible prostate cancer. Ultimately, the host stromal response to the cancer may be responsible for local and metastatic control of the tumor. In further support of stromal differences mediating variations in imaging detection, no notable differences were observed in the immune composition of the two groups of tumors. However, in addition to the significant differential expression of FLNC observed between the mpMRI-invisible and mpMRI-visible groups, DNAJB5, a member of the HSP40/DNAJ family, was also significantly differentially expressed between the two groups. HSP40/DNAJ family members are known to have either antitumor or protumor biological roles, including the regulation of tumor invasion and mesenchymal and stromal phenotypes (31), potentially suggesting that DNAJB5 may influence the stromal prostate cancer microenvironment. Moreover, increased DNAJB5 expression may result in less protein aggregation (44), which may then alter the prostate cancer tumors cellularly and molecularly, leading to the inability to detect the tumors via imaging. Overall, the combination of the molecular and structural differences in the stromal compartments of the mpMRI-visible and mpMRI-invisible prostate cancer tumors indicate that a stroma-related phenomenon is responsible for the inability to detect certain prostate cancer tumors by mpMRI.

Associations between MRI visibility and prostate cancer clinical outcomes were also revealed through this multi-omics analysis. Characteristics linked with mpMRI visibility, including higher malignant cell density and number of neighboring malignant cells, were associated with prostate cancer recurrence. Moreover, lower FLNC expression and low expression of a “stromal” signature was found in mpMRI-visible prostate cancer tumors compared with mpMRI-invisible tumors. Our data support other published data showing the correlation between low FLNC expression and poorer clinical outcomes in prostate cancer. The low expression of the “stromal” signature was associated with inferior recurrence-free survival, suggesting a potential link between mpMRI visibility, specifically PI-RADS 5 visibility, with poor clinical outcome and an aggressive clinical course.

Our study has a number of limitations, including a small sample size for both the prostate cancer MxIF and transcriptomic analyses. Moreover, the MxIF imaging and gene expression analyses were dependent on defined panels, limiting our analyses to the exploratory phase, necessitating validation in larger cohorts to truly discern the link between MRI visibility and clinical outcome. Nevertheless, we believe that this is one of the first studies linking imaging variations with significant molecular disparities that may ultimately underlie notable differences in clinical outcome and demonstrated the need for tumor architecture to be examined at single-cell resolution via complex imaging analysis, which may lead to the development of biomarkers through less invasive measures. For example, we linked FLNC expression to stromal structural differences between mpMRI-invisible and mpMRI-visible tumors and superior survival, suggesting that FLNC expression and transcriptomics can be potentially applied in a clinical setting. Moreover, the association between mpMRI visibility and clinical outcomes indicates that mpMRI visibility alone could potentially be utilized as a biomarker in prostate cancer treatment decisions, although this should be validated in larger cohorts. In addition, if the observed molecular differences are a result of a potential host response to the tumor, mpMRI visibility could be used to inform possible adjuvant therapies—especially those targeting stroma—to improve clinical outcomes.

N. Miheecheva reports employment with BostonGene. N. Kotlov is employed by BostonGene and receives compensation as part of his employment. E. Postovalova is employed by BostonGene and receives compensation as part of her employment, and is an inventor on patent applications related to the imaging analysis technology. I. Galkin is employed by BostonGene and receives compensation as part of his employment, and is an inventor on patent applications related to the imaging analysis technology. V. Svekolkin is employed by BostonGene and receives compensation as part of his employment, and is an inventor on patent applications related to the imaging analysis technology. J.P. Gaut reports grants from NIH and grants from Mid-America Transplant Foundation outside the submitted work. A. Bagaev is employed by BostonGene and receives compensation as part of his employment, and is an inventor on patent applications related to the imaging analysis technology. M. Bruttan is employed by BostonGene and receives compensation as a part of her employment. O. Gancharova reports employment with BostonGene. K. Nomie is employed by BostonGene and receives compensation as part of her employment. M. Tsiper is employed by BostonGene and receives compensation as a part of her employment. R. Ataullakhanov is employed by BostonGene and receives compensation as part of his employment, and is an inventor on patent applications related to the imaging analysis technology. J.J. Hsieh reports grants and personal fees from BostonGene during the conduct of the study and personal fees from Eisai outside the submitted work. No disclosures were reported by the other authors.

R.K. Pachynski: Conceptualization, resources, data curation, formal analysis, supervision, investigation, writing–original draft, writing–review and editing. E.H. Kim: Conceptualization, resources, data curation, formal analysis, writing–original draft, project administration, writing–review and editing. N. Miheecheva: software, formal analysis, validation, investigation, writing–original draft, writing–review and editing. N. Kotlov: Formal analysis, supervision, investigation, methodology. A. Ramachandran: Resources, data curation, formal analysis, writing–review and editing. E. Postovalova: Data curation, formal analysis, supervision, visualization, methodology, writing–review and editing. I. Galkin: Formal analysis, investigation, visualization, methodology. V. Svekolkin: Formal analysis, investigation, visualization, methodology. Y. Lyu: Data curation, formal analysis, investigation. Q. Zou: Data curation, formal analysis, investigation. D. Cao: Resources, data curation, formal analysis. J. Gaut: Formal analysis, investigation. J.E. Ippolito: Data curation, formal analysis, investigation. A. Bagaev: Data curation, formal analysis, supervision, investigation, visualization, methodology, project administration, writing–review and editing. M. Bruttan: Data curation, investigation, visualization. O. Gancharova: Formal analysis, investigation, visualization. K. Nomie: Writing–original draft, project administration, writing–review and editing. M. Tsiper: Resources, supervision, project administration, writing–review and editing. G.L. Andriole: Conceptualization, data curation, investigation, writing–review and editing. R. Ataullakhanov: Conceptualization, supervision, investigation, project administration, writing–review and editing. J.J. Hsieh: Conceptualization, resources, investigation, writing–original draft, project administration, writing–review and editing.

The results shown here are in whole or part based upon data generated by TCGA Research Network: http://cancergenome.nih.gov/ and The Cancer Imaging Archive (45). This work was funded in part by a grant to Russell Pachynski (principal investigator) from the Midwest Stone Institute.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.