Oncogenic translation directs spliceosome dynamics revealing an integral role for SF3A3 in breast cancer.

Splicing is a central RNA-based process commonly altered in human cancers; however, how spliceosomal components are co-opted during tumorigenesis remains poorly defined. Here we unravel the core splice factor SF3A3 at the nexus of a translation-based program that rewires splicing during malignant transformation. Upon MYC hyperactivation, SF3A3 levels are modulated translationally through an RNA stem-loop in an eIF3D-dependent manner. This ensures accurate splicing of mRNAs enriched for mitochondrial regulators. Altered SF3A3 translation leads to metabolic reprogramming and stem-like properties that fuel MYC tumorigenic potential in vivo. Our analysis reveals that SF3A3 protein levels predict molecular and phenotypic features of aggressive human breast cancers. These findings unveil a post-transcriptional interplay between splicing and translation that governs critical facets of MYC-driven oncogenesis.

Author Correction: Tertiary lymphoid structures improve immunotherapy and survival in melanoma.

An Amendment to this paper has been published and can be accessed via a link at the top of the paper.

Tertiary lymphoid structures improve immunotherapy and survival in melanoma.

Checkpoint blockade therapies that reactivate tumour-associated T cells can induce durable tumour control and result in the long-term survival of patients with advanced cancers1. Current predictive biomarkers for therapy response include high levels of intratumour immunological activity, a high tumour mutational burden and specific characteristics of the gut microbiota2,3. Although the role of T cells in antitumour responses has thoroughly been studied, other immune cells remain insufficiently explored. Here we use clinical samples of metastatic melanomas to investigate the role of B cells in antitumour responses, and find that the co-occurrence of tumour-associated CD8+ T cells and CD20+ B cells is associated with improved survival, independently of other clinical variables. Immunofluorescence staining of CXCR5 and CXCL13 in combination with CD20 reveals the formation of tertiary lymphoid structures in these CD8+CD20+ tumours. We derived a gene signature associated with tertiary lymphoid structures, which predicted clinical outcomes in cohorts of patients treated with immune checkpoint blockade. Furthermore, B-cell-rich tumours were accompanied by increased levels of TCF7+ naive and/or memory T cells. This was corroborated by digital spatial-profiling data, in which T cells in tumours without tertiary lymphoid structures had a dysfunctional molecular phenotype. Our results indicate that tertiary lymphoid structures have a key role in the immune microenvironment in melanoma, by conferring distinct T cell phenotypes. Therapeutic strategies to induce the formation of tertiary lymphoid structures should be explored to improve responses to cancer immunotherapy.

Development and prognostic validation of a three-level NHG-like deep learning-based model for histological grading of breast cancer.

BACKGROUND: Histological grade is a well-known prognostic factor that is routinely assessed in breast tumours. However, manual assessment of Nottingham Histological Grade (NHG) has high inter-assessor and inter-laboratory variability, causing uncertainty in grade assignments. To address this challenge, we developed and validated a three-level NHG-like deep learning-based histological grade model (predGrade). The primary performance evaluation focuses on prognostic performance. METHODS: This observational study is based on two patient cohorts (SöS-BC-4, N = 2421 (training and internal test); SCAN-B-Lund, N = 1262 (test)) that include routine histological whole-slide images (WSIs) together with patient outcomes. A deep convolutional neural network (CNN) model with an attention mechanism was optimised for the classification of the three-level histological grading (NHG) from haematoxylin and eosin-stained WSIs. The prognostic performance was evaluated by time-to-event analysis of recurrence-free survival and compared to clinical NHG grade assignments in the internal test set as well as in the fully independent external test cohort. RESULTS: We observed effect sizes (hazard ratio) for grade 3 versus 1, for the conventional NHG method (HR = 2.60 (1.18-5.70 95%CI, p-value = 0.017)) and the deep learning model (HR = 2.27, 95%CI 1.07-4.82, p-value = 0.033) on the internal test set after adjusting for established clinicopathological risk factors. In the external test set, the unadjusted HR for clinical NHG 2 versus 1 was estimated to be 2.59 (p-value = 0.004) and clinical NHG 3 versus 1 was estimated to be 3.58 (p-value < 0.001). For predGrade, the unadjusted HR for predGrade 2 versus 1 HR = 2.52 (p-value = 0.030), and 4.07 (p-value = 0.001) for preGrade 3 versus 1 was observed in the independent external test set. In multivariable analysis, HR estimates for neither clinical NHG nor predGrade were found to be significant (p-value > 0.05). We tested for differences in HR estimates between NHG and predGrade in the independent test set and found no significant difference between the two classification models (p-value > 0.05), confirming similar prognostic performance between conventional NHG and predGrade. CONCLUSION: Routine histopathology assessment of NHG has a high degree of inter-assessor variability, motivating the development of model-based decision support to improve reproducibility in histological grading. We found that the proposed model (predGrade) provides a similar prognostic performance as clinical NHG. The results indicate that deep CNN-based models can be applied for breast cancer histological grading.

Improved detection of clinically relevant fusion transcripts in cancer by machine learning classification.

BACKGROUND: Genomic rearrangements in cancer cells can create fusion genes that encode chimeric proteins or alter the expression of coding and non-coding RNAs. In some cancer types, fusions involving specific kinases are used as targets for therapy. Fusion genes can be detected by whole genome sequencing (WGS) and targeted fusion panels, but RNA sequencing (RNA-Seq) has the advantageous capability of broadly detecting expressed fusion transcripts. RESULTS: We developed a pipeline for validation of fusion transcripts identified in RNA-Seq data using matched WGS data from The Cancer Genome Atlas (TCGA) and applied it to 910 tumors from 11 different cancer types. This resulted in 4237 validated gene fusions, 3049 of them with at least one identified genomic breakpoint. Utilizing validated fusions as true positive events, we trained a machine learning classifier to predict true and false positive fusion transcripts from RNA-Seq data. The final precision and recall metrics of the classifier were 0.74 and 0.71, respectively, in an independent dataset of 249 breast tumors. Application of this classifier to all samples with RNA-Seq data from these cancer types vastly extended the number of likely true positive fusion transcripts and identified many potentially targetable kinase fusions. Further analysis of the validated gene fusions suggested that many are created by intrachromosomal amplification events with microhomology-mediated non-homologous end-joining. CONCLUSIONS: A classifier trained on validated fusion events increased the accuracy of fusion transcript identification in samples without WGS data. This allowed the analysis to be extended to all samples with RNA-Seq data, facilitating studies of tumor biology and increasing the number of detected kinase fusions. Machine learning could thus be used in identification of clinically relevant fusion events for targeted therapy. The large dataset of validated gene fusions generated here presents a useful resource for development and evaluation of fusion transcript detection algorithms.

Matched analysis of circulating selenium with the breast cancer selenotranscriptome: a multicentre prospective study.

INTRODUCTION: Low serum selenium and altered tumour RNA expression of certain selenoproteins are associated with a poor breast cancer prognosis. Selenoprotein expression stringently depends on selenium availability, hence circulating selenium may interact with tumour selenoprotein expression. However, there is no matched analysis to date. METHODS: This study included 1453 patients with newly diagnosed breast cancer from the multicentric prospective Sweden Cancerome Analysis Network - Breast study. Total serum selenium, selenoprotein P and glutathione peroxidase 3 were analysed at time of diagnosis. Bulk RNA-sequencing was conducted in matched tumour tissues. Fully adjusted Cox regression models with an interaction term were employed to detect dose-dependent interactions of circulating selenium with the associations of tumour selenoprotein mRNA expression and mortality. RESULTS: 237 deaths were recorded within ~ 9 years follow-up. All three serum selenium biomarkers correlated positively (p < 0.001). All selenoproteins except for GPX6 were expressed in tumour tissues. Single cell RNA-sequencing revealed a heterogeneous expression pattern in the tumour microenvironment. Circulating selenium correlated positively with tumour SELENOW and SELENON expression (p < 0.001). In fully adjusted models, the associations of DIO1, DIO3 and SELENOM with mortality were dose-dependently modified by serum selenium (p < 0.001, p = 0.020, p = 0.038, respectively). With increasing selenium, DIO1 and SELENOM associated with lower, whereas DIO3 expression associated with higher mortality. Association of DIO1 with lower mortality was only apparent in patients with high selenium [above median (70.36 µg/L)], and the HR (95%CI) for one-unit increase in log(FPKM + 1) was 0.70 (0.50-0.98). CONCLUSIONS: This first unbiased analysis of serum selenium with the breast cancer selenotranscriptome identified an effect-modification of selenium on the associations of DIO1, SELENOM, and DIO3 with prognosis. Selenium substitution in patients with DIO1-expressing tumours merits consideration to improve survival.

Regulatory networks and 5' partner usage of miRNA host gene fusions in breast cancer.

Genomic rearrangements in cancer cells can create gene fusions where the juxtaposition of two different genes leads to the production of chimeric proteins or altered gene expression through promoter-swapping. We have previously shown that fusion transcripts involving microRNA (miRNA) host genes contribute to deregulation of miRNA expression regardless of the protein-coding potential of these transcripts. Many different genes can also be used as 5' partners by a miRNA host gene in what we named recurrent miRNA-convergent fusions. Here, we have explored the properties of 5' partners in fusion transcripts that involve miRNA hosts in breast tumours from The Cancer Genome Atlas (TCGA). We hypothesised that firstly, 5' partner genes should belong to pathways and transcriptional programmes that reflect the tumour phenotype and secondly, there should be a selection for fusion events that shape miRNA expression to benefit the tumour cell through the known hallmarks of cancer. We found that the set of 5' partners in miRNA host fusions is non-random, with overrepresentation of highly expressed genes in pathways active in cancer including epithelial-to-mesenchymal transition, translational regulation and oestrogen signalling. Furthermore, many miRNAs were upregulated in samples with host gene fusions, including established oncogenic miRNAs such as mir-21 and the mir-106b~mir-93~mir-25 cluster. To the list of mechanisms for deregulation of miRNA expression, we have added fusion transcripts that change the promoter region. We propose that this adds material for genetic selection and tumour evolution in cancer cells and that miRNA host fusions can act as tumour 'drivers'.

Performance of gene expression-based single sample predictors for assessment of clinicopathological subgroups and molecular subtypes in cancers: a case comparison study in non-small cell lung cancer.

The development of multigene classifiers for cancer prognosis, treatment prediction, molecular subtypes or clinicopathological groups has been a cornerstone in transcriptomic analyses of human malignancies for nearly two decades. However, many reported classifiers are critically limited by different preprocessing needs like normalization and data centering. In response, a new breed of classifiers, single sample predictors (SSPs), has emerged. SSPs classify samples in an N-of-1 fashion, relying on, e.g. gene rules comparing expression values within a sample. To date, several methods have been reported, but there is a lack of head-to-head performance comparison for typical cancer classification problems, representing an unmet methodological need in cancer bioinformatics. To resolve this need, we performed an evaluation of two SSPs [k-top-scoring pair classifier (kTSP) and absolute intrinsic molecular subtyping (AIMS)] for two case examples of different magnitude of difficulty in non-small cell lung cancer: gene expression-based classification of (i) tumor histology and (ii) molecular subtype. Through the analysis of ~2000 lung cancer samples for each case example (n = 1918 and n = 2106, respectively), we compared the performance of the methods for different sample compositions, training data set sizes, gene expression platforms and gene rule selections. Three main conclusions are drawn from the comparisons: both methods are platform independent, they select largely overlapping gene rules associated with actual underlying tumor biology and, for large training data sets, they behave interchangeably performance-wise. While SSPs like AIMS and kTSP offer new possibilities to move gene expression signatures/predictors closer to a clinical context, they are still importantly limited by the difficultness of the classification problem at hand.

Molecular analyses of triple-negative breast cancer in the young and elderly.

BACKGROUND: Breast cancer in young adults has been implicated with a worse outcome. Analyses of genomic traits associated with age have been heterogenous, likely because of an incomplete accounting for underlying molecular subtypes. We aimed to resolve whether triple-negative breast cancer (TNBC) in younger versus older patients represent similar or different molecular diseases in the context of genetic and transcriptional subtypes and immune cell infiltration. PATIENTS AND METHODS: In total, 237 patients from a reported population-based south Swedish TNBC cohort profiled by RNA sequencing and whole-genome sequencing (WGS) were included. Patients were binned in 10-year intervals. Complimentary PD-L1 and CD20 immunohistochemistry and estimation of tumor-infiltrating lymphocytes (TILs) were performed. Cases were analyzed for differences in patient outcome, genomic, transcriptional, and immune landscape features versus age at diagnosis. Additionally, 560 public WGS breast cancer profiles were used for validation. RESULTS: Median age at diagnosis was 62 years (range 26-91). Age was not associated with invasive disease-free survival or overall survival after adjuvant chemotherapy. Among the BRCA1-deficient cases (82/237), 90% were diagnosed before the age of 70 and were predominantly of the basal-like subtype. In the full TNBC cohort, reported associations of patient age with changes in Ki67 expression, PIK3CA mutations, and a luminal androgen receptor subtype were confirmed. Within DNA repair deficiency or gene expression defined molecular subgroups, age-related alterations in, e.g., overall gene expression, immune cell marker gene expression, genetic mutational and rearrangement signatures, amount of copy number alterations, and tumor mutational burden did, however, not appear distinct. Similar non-significant associations for genetic alterations with age were obtained for other breast cancer subgroups in public WGS data. Consistent with age-related immunosenescence, TIL counts decreased linearly with patient age across different genetic TNBC subtypes. CONCLUSIONS: Age-related alterations in TNBC, as well as breast cancer in general, need to be viewed in the context of underlying genomic phenotypes. Based on this notion, age at diagnosis alone does not appear to provide an additional layer of biological complexity above that of proposed genetic and transcriptional phenotypes of TNBC. Consequently, treatment decisions should be less influenced by age and more driven by tumor biology.

Association between breast cancer risk and disease aggressiveness: Characterizing underlying gene expression patterns.

The association between breast cancer risk defined by the Tyrer-Cuzick score (TC) and disease prognosis is not well established. Here, we investigated the relationship between 5-year TC and disease aggressiveness and then characterized underlying molecular processes. In a case-only study (n = 2474), we studied the association of TC with molecular subtypes and tumor characteristics. In a subset of patients (n = 672), we correlated gene expression to TC and computed a low-risk TC gene expression (TC-Gx) profile, that is, a profile expected to be negatively associated with risk, which we used to test for association with disease aggressiveness. We performed enrichment analysis to pinpoint molecular processes likely to be altered in low-risk tumors. A higher TC was found to be inversely associated with more aggressive surrogate molecular subtypes and tumor characteristics (P < .05) including Ki-67 proliferation status (P < 5 × 10-07 ). Our low-risk TC-Gx, based on the weighted sum of 37 expression values of genes strongly correlated with TC, was associated with basal-like (P < 5 × 10-13 ), HER2-enriched subtype (P < 5 × 10-07 ) and worse 10-year breast cancer-specific survival (log-rank P < 5 × 10-04 ). Associations between low-risk TC-Gx and more aggressive molecular subtypes were replicated in an independent cohort from The Cancer Genome Atlas database (n = 975). Gene expression that correlated with low TC was enriched in proliferation and oncogenic signaling pathways (FDR < 0.05). Moreover, higher proliferation was a key factor explaining the association with worse survival. Women who developed breast cancer despite having a low risk were diagnosed with more aggressive tumors and had a worse prognosis, most likely driven by increased proliferation. Our findings imply the need to establish risk factors associated with more aggressive breast cancer subtypes.

Analysis of fusion transcripts indicates widespread deregulation of snoRNAs and their host genes in breast cancer.

Genomic rearrangements in cancer can join the sequences of two separate genes. Studies of such gene fusion events have mainly focused on identification of fusion proteins from the chimeric transcripts. We have previously investigated how fusions instead can affect the expression of intronic microRNA (miRNA) genes that are encoded within fusion gene partners. Here, we extend our analysis to small nucleolar RNAs (snoRNAs) that also are embedded within protein-coding or noncoding host genes. We found that snoRNA hosts are selectively enriched in fusion transcripts, like miRNA host genes, and that this enrichment is associated with all snoRNA classes. These structural changes may have functional consequences for the cell; proteins involved in the protein translation machinery are overrepresented among snoRNA host genes, a gene architecture assumed to be needed for closely coordinated expression of snoRNAs and host proteins. Our data indicate that this structure is frequently disrupted in cancer. We furthermore observed that snoRNA genes involved in fusions tend to associate with stronger promoters than the natural host, suggesting a mechanism that selects for snoRNA overexpression. In summary, we highlight a previously unexplored frequent structural change in cancer that affects important components of cellular physiology.

Alternative splicing and ACMG-AMP-2015-based classification of PALB2 genetic variants: an ENIGMA report.

BACKGROUND: PALB2 monoallelic loss-of-function germ-line variants confer a breast cancer risk comparable to the average BRCA2 pathogenic variant. Recommendations for risk reduction strategies in carriers are similar. Elaborating robust criteria to identify loss-of-function variants in PALB2-without incurring overprediction-is thus of paramount clinical relevance. Towards this aim, we have performed a comprehensive characterisation of alternative splicing in PALB2, analysing its relevance for the classification of truncating and splice site variants according to the 2015 American College of Medical Genetics and Genomics-Association for Molecular Pathology guidelines. METHODS: Alternative splicing was characterised in RNAs extracted from blood, breast and fimbriae/ovary-related human specimens (n=112). RNAseq, RT-PCR/CE and CloneSeq experiments were performed by five contributing laboratories. Centralised revision/curation was performed to assure high-quality annotations. Additional splicing analyses were performed in PALB2 c.212-1G>A, c.1684+1G>A, c.2748+2T>G, c.3113+5G>A, c.3350+1G>A, c.3350+4A>C and c.3350+5G>A carriers. The impact of the findings on PVS1 status was evaluated for truncating and splice site variant. RESULTS: We identified 88 naturally occurring alternative splicing events (81 newly described), including 4 in-frame events predicted relevant to evaluate PVS1 status of splice site variants. We did not identify tissue-specific alternate gene transcripts in breast or ovarian-related samples, supporting the clinical relevance of blood-based splicing studies. CONCLUSIONS: PVS1 is not necessarily warranted for splice site variants targeting four PALB2 acceptor sites (exons 2, 5, 7 and 10). As a result, rare variants at these splice sites cannot be assumed pathogenic/likely pathogenic without further evidences. Our study puts a warning in up to five PALB2 genetic variants that are currently reported as pathogenic/likely pathogenic in ClinVar.

Agreement between molecular subtyping and surrogate subtype classification: a contemporary population-based study of ER-positive/HER2-negative primary breast cancer.

PURPOSE: Oestrogen receptor-positive (ER+) and human epidermal receptor 2-negative (HER2-) breast cancers are classified as Luminal A or B based on gene expression, but immunohistochemical markers are used for surrogate subtyping. The aims of this study were to examine the agreement between molecular subtyping (MS) and surrogate subtyping and to identify subgroups consisting mainly of Luminal A or B tumours. METHODS: The cohort consisted of 2063 patients diagnosed between 2013-2017, with primary ER+/HER2- breast cancer, analysed by RNA sequencing. Surrogate subtyping was performed according to three algorithms (St. Gallen 2013, Maisonneuve and our proposed Grade-based classification). Agreement (%) and kappa statistics (κ) were used as concordance measures and ROC analysis for luminal distinction. Ki67, progesterone receptor (PR) and histological grade (HG) were further investigated as surrogate markers. RESULTS: The agreement rates between the MS and St. Gallen 2013, Maisonneuve and Grade-based classifications were 62% (κ = 0.30), 66% (κ = 0.35) and 70% (κ = 0.41), respectively. PR did not contribute to distinguishing Luminal A from B tumours (auROC = 0.56). By classifying HG1-2 tumours as Luminal A-like and HG3 as Luminal B-like, agreement with MS was 80% (κ = 0.46). Moreover, by combining HG and Ki67 status, a large subgroup of patients (51% of the cohort) having > 90% Luminal A tumours could be identified. CONCLUSIONS: Agreement between MS and surrogate classifications was generally poor. However, a post hoc analysis showed that a combination of HG and Ki67 could identify patients very likely to have Luminal A tumours according to MS.

An integrated genomics analysis of epigenetic subtypes in human breast tumors links DNA methylation patterns to chromatin states in normal mammary cells.

BACKGROUND: Aberrant DNA methylation is frequently observed in breast cancer. However, the relationship between methylation patterns and the heterogeneity of breast cancer has not been comprehensively characterized. METHODS: Whole-genome DNA methylation analysis using Illumina Infinium HumanMethylation450 BeadChip arrays was performed on 188 human breast tumors. Unsupervised bootstrap consensus clustering was performed to identify DNA methylation epigenetic subgroups (epitypes). The Cancer Genome Atlas data, including methylation profiles of 669 human breast tumors, was used for validation. The identified epitypes were characterized by integration with publicly available genome-wide data, including gene expression levels, DNA copy numbers, whole-exome sequencing data, and chromatin states. RESULTS: We identified seven breast cancer epitypes. One epitype was distinctly associated with basal-like tumors and with BRCA1 mutations, one epitype contained a subset of ERBB2-amplified tumors characterized by multiple additional amplifications and the most complex genomes, and one epitype displayed a methylation profile similar to normal epithelial cells. Luminal tumors were stratified into the remaining four epitypes, with differences in promoter hypermethylation, global hypomethylation, proliferative rates, and genomic instability. Specific hyper- and hypomethylation across the basal-like epitype was rare. However, we observed that the candidate genomic instability drivers BRCA1 and HORMAD1 displayed aberrant methylation linked to gene expression levels in some basal-like tumors. Hypomethylation in luminal tumors was associated with DNA repeats and subtelomeric regions. We observed two dominant patterns of aberrant methylation in breast cancer. One pattern, constitutively methylated in both basal-like and luminal breast cancer, was linked to genes with promoters in a Polycomb-repressed state in normal epithelial cells and displayed no correlation with gene expression levels. The second pattern correlated with gene expression levels and was associated with methylation in luminal tumors and genes with active promoters in normal epithelial cells. CONCLUSIONS: Our results suggest that hypermethylation patterns across basal-like breast cancer may have limited influence on tumor progression and instead reflect the repressed chromatin state of the tissue of origin. On the contrary, hypermethylation patterns specific to luminal breast cancer influence gene expression, may contribute to tumor progression, and may present an actionable epigenetic alteration in a subset of luminal breast cancers.

Erythropoietin suppresses the activation of pro-apoptotic genes in head and neck squamous cell carcinoma xenografts exposed to surgical trauma.

BACKGROUND: Several studies on the use of erythropoietin (Epo) to treat anaemia in patients undergoing cancer treatment have shown adverse effects on tumour control and survival. Experimental studies indicate that this could be linked to an interaction with wound healing processes and not an effect on tumour cells per se. We have previously shown that erythropoietin in combination with surgical trauma stimulates tumour growth. In the present study, we investigated the effect of surgery and Epo on gene expression. METHODS: Human tumours from oral squamous cell cancer were xenotransplanted to nude mice treated with Epo. The tumours were then transected in a standardised procedure to mimic surgical trauma and the change in gene expression of the tumours was investigated by microarray analysis. qRT-PCR was used to measure the levels of mRNAs of pro-apoptotic genes. The frequency of apoptosis in the tumours was assessed using immunohistochemistry for caspase-3. RESULTS: There was little change in the expression of genes involved in tumour growth and angiogenesis but a significant down-regulation of the expression of genes involved in apoptosis. This effect on apoptosis was confirmed by a general decrease in the expression of mRNA for selected pro-apoptotic genes. Epo-treated tumours had a significantly lower frequency of apoptosis as measured by immunohistochemistry for caspase 3. CONCLUSIONS: Our results suggest that the increased tumour growth during erythropoietin treatment might be due to inhibition of apoptosis, an effect that becomes significant during tissue damage such as surgery.This further suggests that the decreased survival during erythropoietin treatment might be due to inhibition of apoptosis.

Recruitment of HIF-1alpha and HIF-2alpha to common target genes is differentially regulated in neuroblastoma: HIF-2alpha promotes an aggressive phenotype.

In neuroblastoma specimens, HIF-2alpha but not HIF-1alpha is strongly expressed in well-vascularized areas. In vitro, HIF-2alpha protein was stabilized at 5% O2 (resembling end capillary oxygen conditions) and, in contrast to the low HIF-1alpha activity at this oxygen level, actively transcribed genes like VEGF. Under hypoxia (1% O2), HIF-1alpha was transiently stabilized and primarily mediated acute responses, whereas HIF-2alpha protein gradually accumulated and governed prolonged hypoxic gene activation. Knockdown of HIF-2alpha reduced growth of neuroblastoma tumors in athymic mice. Furthermore, high HIF-2alpha protein levels were correlated with advanced clinical stage and high VEGF expression and predicted poor prognosis in a clinical neuroblastoma material. Our results demonstrate the relevance of HIF-2alpha in neuroblastoma progression and have general tumor biological implications.

Caveolin-1 gene expression provides additional prognostic information combined with PAM50 risk of recurrence (ROR) score in breast cancer.

Combining information from the tumor microenvironment (TME) with PAM50 Risk of Recurrence (ROR) score could improve breast cancer prognostication. Caveolin-1 (CAV1) is a marker of an active TME. CAV1 is a membrane protein involved in cell signaling, extracellular matrix organization, and tumor-stroma interactions. We sought to investigate CAV1 gene expression in relation to PAM50 subtypes, ROR score, and their joint prognostic impact. CAV1 expression was compared between PAM50 subtypes and ROR categories in two cohorts (SCAN-B, n = 5326 and METABRIC, n = 1980). CAV1 expression was assessed in relation to clinical outcomes using Cox regression and adjusted for clinicopathological predictors. Effect modifications between CAV1 expression and ROR categories on clinical outcome were investigated using multiplicative and additive two-way interaction analyses. Differential gene expression and gene set enrichment analyses were applied to compare high and low expressing CAV1 tumors. All samples expressed CAV1 with the highest expression in the Normal-like subtype. Gene modules consistent with epithelial-mesenchymal transition (EMT), hypoxia, and stromal activation were associated with high CAV1 expression. CAV1 expression was inversely associated with ROR category. Interactions between CAV1 expression and ROR categories were observed in both cohorts. High expressing CAV1 tumors conferred worse prognosis only within the group classified as ROR high. ROR gave markedly different prognostic information depending on the underlying CAV1 expression. CAV1, a potential mediator between the malignant cells and TME, could be a useful biomarker that enhances and further refines PAM50 ROR risk stratification in patients with ROR high tumors and a potential therapeutic target.

Perturbation and stability of PAM50 subtyping in population-based primary invasive breast cancer.

PAM50 gene expression subtypes represent a cornerstone in the molecular classification of breast cancer and are included in risk prediction models to guide therapy. We aimed to illustrate the impact of included genes and biological processes on subtyping while considering a tumor's underlying clinical subgroup defined by ER, PR, and HER2 status. To do this we used a population-representative and clinically annotated early-stage breast tumor cohort of 6233 samples profiled by RNA sequencing and applied a perturbation strategy of excluding co-expressed genes (gene sets). We demonstrate how PAM50 nearest-centroid classification depends on biological processes present across, but also within, ER/PR/HER2 subgroups and PAM50 subtypes themselves. Our analysis highlights several key aspects of PAM50 classification. Firstly, we demonstrate the tight connection between a tumor's nearest and second-nearest PAM50 centroid. Additionally, we show that the second-best subtype is associated with overall survival in ER-positive, HER2-negative, and node-negative disease. We also note that ERBB2 expression has little impact on PAM50 classification in HER2-positive disease regardless of ER status and that the Basal subtype is highly stable in contrast to the Normal subtype. Improved consciousness of the commonly used PAM50 subtyping scheme will aid in our understanding and interpretation of breast tumors that have seemingly conflicting PAM50 classification when compared to clinical biomarkers. Finally, our study adds further support in challenging the common misconception that PAM50 subtypes are distinct classes by illustrating that PAM50 subtypes in tumors represent a continuum with prognostic implications.

Transcriptional intra-tumour heterogeneity predicted by deep learning in routine breast histopathology slides provides independent prognostic information.

BACKGROUND: Intra-tumour heterogeneity (ITH) causes diagnostic challenges and increases the risk for disease recurrence. Quantification of ITH is challenging and has not been demonstrated in large studies. It has previously been shown that deep learning can enable spatially resolved prediction of molecular phenotypes from digital histopathology whole slide images (WSIs). Here we propose a novel method (Deep-ITH) to predict and measure ITH, and we evaluate its prognostic performance in breast cancer. METHODS: Deep convolutional neural networks were used to spatially predict gene-expression (PAM50 set) from WSIs. For each predicted transcript, 12 measures of heterogeneity were extracted in the training data set (N = 931). A prognostic score to dichotomise patients into Deep-ITH low- and high-risk groups was established using an elastic-net regularised Cox proportional hazards model (recurrence-free survival). Prognostic performance was evaluated in two independent data sets: SöS-BC-1 (N = 1358) and SCAN-B-Lund (N = 1262). RESULTS: We observed an increase in risk of recurrence in the high-risk group with hazard ratio (HR) 2.11 (95%CI:1.22-3.60; p = 0.007) using nested cross-validation. Subgroup analyses confirmed the prognostic performance in oestrogen receptor (ER)-positive, human epidermal growth factor receptor 2 (HER2)-negative, grade 3, and large tumour subgroups. The prognostic value was confirmed in the independent SöS-BC-1 cohort (HR=1.84; 95%CI:1.03-3.3; p = 3.99 ×10-2). In the other external cohort, significant HR was observed in the subgroup of histological grade 2 patients, as well as in the subgroup of patients with small tumours (<20 mm). CONCLUSION: We developed a novel method for an automated, scalable, and cost-efficient measure of ITH from WSIs that provides independent prognostic value for breast cancer. SIGNIFICANCE: Transcriptional ITH predicted by deep learning models enables prediction of patient survival from routine histopathology WSIs in breast cancer.