Multiparametric magnetic resonance imaging (mpMRI) has dramatically enhanced the management of localized prostate cancer, providing an opportunity to improve diagnosis and risk stratification while simultaneously reducing unnecessary and risky needle biopsies . However, because ~ 20% of clinically significant tumors remain invisible to mpMRI , there is limited consensus on when a biopsy can be safely avoided upon a negative mpMRI. The reasons for prostate cancer mpMRI invisibility are largely unknown, despite mpMRI-visible tumors harboring more adverse pathological and biological features [3,4,5,6]. Within International Society of Urological Pathology (ISUP) Grade Group 2, mpMRI visibility is associated with increased genomic instability, presence of intraductal carcinoma and/or cribriform architecture (IDC/CA) histology and hypoxia, a constellation of features termed nimbosus [3, 7]. Given the role of cellular density and perfusion in mpMRI, differences in stromal organization in non-malignant tissue  are hypothesized to affect water diffusion and thus mediate tumor microenvironmental influence on mpMRI visibility.
To understand the biological underpinnings of tumor visibility on mpMRI, we performed global proteomics on twenty mpMRI-invisible (Prostate Imaging Reporting and Data System version 2 [PI-RADSv2] 1–2) and twenty mpMRI-visible (PI-RADSv2 5) tumors, all from patients with a solitary pathological ISUP Grade Group 2 lesion larger than 1.5 cm . We analyzed both tumor and adjacent histologically normal tissue (NAT) from all patients, leading to 81 proteomes (Fig. 1A, Additional file 3: Table S1). A detailed description of the methods can be found in Additional file 1: Methods (available online).
We quantified 4772 proteins (Additional file 4: Table S2), of which 2309 were detected in all 81 samples (Fig. 1B). Clustering by protein abundance yielded four protein subtypes and four sample subtypes (Additional file 2: Fig. S1A). The sample subtypes were driven by differences between tumors and NATs (Adjusted Rand Index [ARI] = 0.22, p = 0.001) and not mpMRI visibility (ARI = − 0.01, p = 0.64). The protein subtypes reflected specific biological pathways. For example, P1 genes were associated with immune response and extracellular matrix organization and were more abundant in tumors than NATs (Additional file 5: Table S3).
To test the important and widespread hypothesis that the tumor microenvironment influences visibility on mpMRI [3, 8], we compared protein abundances between NATs from patients with mpMRI-visible and mpMRI-invisible tumors. To our surprise, not a single protein differed between the two groups (Fig. 1C). Similarly, differences in the proteomes of mpMRI-visible and mpMRI-invisible tumors were also small and not statistically significant, albeit with larger effect sizes compared to the result from NATs (Fig. 1D). In contrast, we observed the expected large, statistically significant differences between the proteomes of tumors and NATs (Fig. 1E). Similarly, large differences were observed at the transcriptome level (Additional file 1: Methods, Additional file 2: Fig. S1B), where most tumor/NAT proteomic differences were corroborated (Spearman’s ρ = 0.57, p < 2.2 × 10–16, Fig. 1F).
Given these modest differences between mpMRI-visible and mpMRI-invisible tumor proteomes, we hypothesized that mpMRI-invisible tumors might reflect an intermediate state between NATs and mpMRI visibility. Consistent with this, protein abundance differences associated with tumor mpMRI visibility were correlated with NAT-tumor differences (Spearman’s ρ = 0.46, p < 1 × 10–16, Fig. 1G). These associations were diminished in the NAT proteomes (Spearman’s ρ = 0.13, p = 7.01 × 10–11, Additional file 2: Fig. S1C), and in the matched tumor transcriptomes  (Spearman’s ρ = 0.00, p = 0.79, Additional file 2: Fig. S1D). The proteome of mpMRI-invisible tumors was more similar to that of NATs compared to the proteome of mpMRI-visible tumors (Fig. 1H), likely contributing to their invisibility. Consistently, normoxic tumors and tumors lacking IDC/CA histology were more similar to NATs (Fig. 1H). Altered pathways in mpMRI-visible tumors vs. mpMRI-invisible tumors overlapped substantially with those distinguishing tumors from NATs (hypergeometric test p = 5.5 × 10–14, Fig. 1I). Epithelial-to-mesenchymal transition and myogenesis genes were enriched in mpMRI-invisible tumors compared to mpMRI-visible tumors, consistent with reports that stromal and extracellular matrix genes were enriched in mpMRI-invisible tumors . mpMRI-visible tumors were enriched in pathways associated with advanced disease, including androgen response, DNA repair, and MYC and TGF-β signaling . Taken together, these data help explain the aggressive clinical behavior of mpMRI-visible tumors, concordant with increased PTEN loss , higher Oncotype and Decipher genomic classifier scores , and elevated nimbosus hallmarks .
To identify protein-coding RNAs and proteins associated with mpMRI visibility and disease aggression, we next focused on the nimbosus hallmarks [3, 7] and small nucleolar RNAs (snoRNA) that are associated with mpMRI visibility [3, 7]. These hallmarks were previously shown to be associated with mpMRI visibility and disease aggression at the genomic and transcriptomic level . An independent discovery cohort of 144 National Comprehensive Cancer Network (NCCN) intermediate-risk tumors was used to discover associations between RNA abundance and each hallmark (Additional file 1: Methods) [11, 12]. We identified 14,044 protein-coding RNAs and 1,622 proteins associated with at least one nimbosus hallmark in this cohort (Fig. 2A, Additional file 1: Methods). Proportion of the genome with a copy number aberration (PGA) and IDC/CA status showed the largest effects on the transcriptome and proteome. Proteins more abundant in mpMRI-invisible tumors were also negatively correlated with these hallmarks (Fig. 2B). Proteins associated with high PGA were preferentially associated with mpMRI visibility (hypergeometric test p = 3.3 × 10–2; Fig. 2C). mpMRI visibility was also strongly associated with aggressive hallmarks such as hypoxia, presence of IDC/CA, and SChLAP1 expression through proteins, rather than protein-coding RNAs (Fig. 2D).
Finally, we employed a machine learning approach to find proteins that best differentiate mpMRI-visible and mpMRI-invisible tumors in our cohort. Following feature selection, we created a three-protein logistic regression model (LDHB, GNA11, SRD5A2) that classified mpMRI visibility status with an AUC of 0.88 (95% CI = 0.77–0.98, Fig. 2E, Additional file 1: Methods). This model was associated with worse biochemical recurrence-free survival in an independent cohort of 76 predominantly NCCN intermediate-risk tumors (HR = 1.79, 95% CI = 0.92–3.51, p = 0.089, median follow-up 6.02 years, Fig. 2E, inset) , further supporting the association between proteomic determinants of mpMRI visibility and tumor aggressiveness.
These data establish that mpMRI visibility is largely independent of the molecular features of tumor-adjacent stromal cells in the prostate. Rather, the proteome of mpMRI-invisible tumors is more similar to that of normal tissues [4, 10], suggesting that mpMRI visibility reflects the degree of proteomic dysregulation. Caveats of this study include uncertain generalization beyond ISUP Grade Group 2 tumors, the Caucasian ancestry of most patients, and study of only PI-RADSv2 scores of 1–2 and 5. These data suggest that tumors are invisible to mpMRI because their proteome does not differ sufficiently from normal prostate.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.