Thomas Stone

Novel molecular subgroups for clinical classification and outcome prediction in childhood medulloblastoma: a cohort study

Summary Background International consensus recognises four medulloblastoma molecular subgroups: WNT (MBWNT), SHH (MBSHH), group 3 (MBGrp3), and group 4 (MBGrp4), each defined by their characteristic genome-wide transcriptomic and DNA methylomic profiles. These subgroups have distinct clinicopathological and molecular features, and underpin current disease subclassification and initial subgroup-directed therapies that are underway in clinical trials. However, substantial biological heterogeneity and differences in survival are apparent within each subgroup, which remain to be resolved. We aimed to investigate whether additional molecular subgroups exist within childhood medulloblastoma and whether these could be used to improve disease subclassification and prognosis predictions. Methods In this retrospective cohort study, we assessed 428 primary medulloblastoma samples collected from UK Childrens Cancer and Leukaemia Group (CCLG) treatment centres (UK), collaborating European institutions, and the UKCCSG-SIOP-PNET3 European clinical trial. An independent validation cohort (n=276) of archival tumour samples was also analysed. We analysed samples from patients with childhood medulloblastoma who were aged 0–16 years at diagnosis, and had central review of pathology and comprehensive clinical data. We did comprehensive molecular profiling, including DNA methylation microarray analysis, and did unsupervised class discovery of test and validation cohorts to identify consensus primary molecular subgroups and characterise their clinical and biological significance. We modelled survival of patients aged 3–16 years in patients (n=215) who had craniospinal irradiation and had been treated with a curative intent. Findings Seven robust and reproducible primary molecular subgroups of childhood medulloblastoma were identified. MBWNT remained unchanged and each remaining consensus subgroup was split in two. MBSHH was split into age-dependent subgroups corresponding to infant (<4·3 years; MBSHH-Infant; n=65) and childhood patients (≥4·3 years; MBSHH-Child; n=38). MBGrp3 and MBGrp4 were each split into high-risk (MBGrp3-HR [n=65] and MBGrp4-HR [n=85]) and low-risk (MBGrp3-LR [n=50] and MBGrp4-LR [n=73]) subgroups. These biological subgroups were validated in the independent cohort. We identified features of the seven subgroups that were predictive of outcome. Cross-validated subgroup-dependent survival models, incorporating these novel subgroups along with secondary clinicopathological and molecular features and established disease risk-factors, outperformed existing disease risk-stratification schemes. These subgroup-dependent models stratified patients into four clinical risk groups for 5-year progression-free survival: favourable risk (54 [25%] of 215 patients; 91% survival [95% CI 82–100]); standard risk (50 [23%] patients; 81% survival [70–94]); high-risk (82 [38%] patients; 42% survival [31–56]); and very high-risk (29 [13%] patients; 28% survival [14–56]). Interpretation The discovery of seven novel, clinically significant subgroups improves disease risk-stratification and could inform treatment decisions. These data provide a new foundation for future research and clinical investigations. Funding Cancer Research UK, The Tom Grahame Trust, Star for Harris, Action Medical Research, SPARKS, The JGW Patterson Foundation, The INSTINCT network (co-funded by The Brain Tumour Charity, Great Ormond Street Childrens Charity, and Children with Cancer UK).

PredictionIO: a distributed machine learning server for practical software development

One of the biggest challenges for software developers to build real-world predictive applications with machine learning is the steep learning curve of data processing frameworks, learning algorithms and scalable system infrastructure. We present PredictionIO, an open source machine learning server that comes with a step-by-step graphical user interface for developers to (i) evaluate, compare and deploy scalable learning algorithms, (ii) tune hyperparameters of algorithms manually or automatically and (iii) evaluate model training status. The system also comes with an Application Programming Interface (API) to communicate with software applications for data collection and prediction retrieval. The whole infrastructure of PredictionIO is horizontally scalable with a distributed computing component based on Hadoop. The demonstration shows a live example and workflows of building real-world predictive applications with the graphical user interface of PredictionIO, from data collection, algorithm tuning and selection, model training and re-training to real-time prediction querying.

An empirical study of top-n recommendation for venture finance

This paper concerns the task of top-N investment opportunity recommendation in the domain of venture finance. By venture finance, specifically, we are interested in the investment activity of venture capital (VC) firms and their investment partners. We have access to a dataset of recorded venture financings (i.e., investments) by VCs and their investment partners in private US companies. This research was undertaken in partnership with Correlation Ventures, a venture capital firm who are pioneering the use of predictive analytics in order to better inform investment decision making. This paper undertakes a detailed empirical study and data analysis then demonstrates the efficacy of recommender systems in this novel application domain.

Tumour compartment transcriptomics demonstrates the activation of inflammatory and odontogenic programmes in human adamantinomatous craniopharyngioma and identifies the MAPK/ERK pathway as a novel therapeutic target

Adamantinomatous craniopharyngiomas (ACPs) are clinically challenging tumours, the majority of which have activating mutations in CTNNB1. They are histologically complex, showing cystic and solid components, the latter comprised of different morphological cell types (e.g. β-catenin-accumulating cluster cells and palisading epithelium), surrounded by a florid glial reaction with immune cells. Here, we have carried out RNA sequencing on 18 ACP samples and integrated these data with an existing ACP transcriptomic dataset. No studies so far have examined the patterns of gene expression within the different cellular compartments of the tumour. To achieve this goal, we have combined laser capture microdissection with computational analyses to reveal groups of genes that are associated with either epithelial tumour cells (clusters and palisading epithelium), glial tissue or immune infiltrate. We use these human ACP molecular signatures and RNA-Seq data from two ACP mouse models to reveal that cell clusters are molecularly analogous to the enamel knot, a critical signalling centre controlling normal tooth morphogenesis. Supporting this finding, we show that human cluster cells express high levels of several members of the FGF, TGFB and BMP families of secreted factors, which signal to neighbouring cells as evidenced by immunostaining against the phosphorylated proteins pERK1/2, pSMAD3 and pSMAD1/5/9 in both human and mouse ACP. We reveal that inhibiting the MAPK/ERK pathway with trametinib, a clinically approved MEK inhibitor, results in reduced proliferation and increased apoptosis in explant cultures of human and mouse ACP. Finally, we analyse a prominent molecular signature in the glial reactive tissue to characterise the inflammatory microenvironment and uncover the activation of inflammasomes in human ACP. We validate these results by immunostaining against immune cell markers, cytokine ELISA and proteome analysis in both solid tumour and cystic fluid from ACP patients. Our data support a new molecular paradigm for understanding ACP tumorigenesis as an aberrant mimic of natural tooth development and opens new therapeutic opportunities by revealing the activation of the MAPK/ERK and inflammasome pathways in human ACP.

Abstract 1804: Expression analysis of adamantinomatous craniopharyngioma suggests two subtypes associated with CTNNB1 mutational frequency and highlights potential therapeutic targets

We conducted RNA-sequencing analysis of Adamantinomatous Craniopharnygioma (ACP), including Laser capture micro-dissection (LCM) of individual cellular compartments, to characterise both disease heterogeneity and targetable deregulated molecular pathways. RNA sequencing was performed on 18 samples of ACP (17 pediatric, 1 adult) and six control samples (three fetal pituitaries, three non-functioning pituitary adenomas). LCM and RNA sequencing were performed on a subset of cases selecting nuclear β catenin accumulating clusters, and paired non-cluster tumour tissue (palisading epithelium). Mouse ACP models indicate that these clusters initiate tumorigenesis in a non-cell autonomous manner. Consensus clustering analysis revealed two subgroups of human ACP, one associated with a high CTNNB1 mutational frequency (>30%), a relatively high tumour content and fibrous reactive tissue (Group H) and another with a lower mutation frequency ( The meta-gene signature of Group H defined by non-negative matrix factorization included secreted factors, such as Sonic Hedgehog (SHH) and Fibroblast Growth Factors (FGFs) 3, 4 and 19, Keratins and Matrix Metalloproteinases and was enriched for genes related to odontogenesis. Weighted gene co-expression network analysis independently showed correlation of expression of such genes with mutational frequency. Glial markers and neural differentiation genes were up-regulated in Group L tumors. Differential expression showed up-regulation in ACPs of potential therapeutic targets including SHH (32x), EGFR (8.9x) and TNF (10x). Similarly, differential expression analysis of LCM dissected nuclear β catenin accumulating clusters identified up-regulation of SHH, FGFs and other secreted factors including as WNTs and BMPs. Analysis of the SHH pathway found that SHH is over expressed in clusters (9x) whilst its downstream targets Ptch1 (1.7x), Gli1 (2.1x), Gli3 (2.9x) are over expressed in palisading epithelium, indicating the presence of paracrine signaling. These findings were independently confirmed by in situ hybridisation in human samples and the mouse model. Clustering of murine expression data with those genes differentially expressed in the human clusters revealed murine and human clusters group together, suggesting they are similar structures functionally and biologically. Together, these results suggest that a subset of cells, carrying CTNNB1 mutations and showing nuclear β-catenin accumulation, express secreted pro-oncogenic factors and the response may be observed in neighboring cell types. Targeting such signalling offers an attractive therapeutic opportunity, for which the results suggest the mouse model is well placed for testing. The clinical significance of the two putative ACP subgroups remains to be determined. Citation Format: John R. Apps, Nital Jani, Gabriela Carreno, Jose M. Gonzalez-Meljem, Kyoko Tossell, Thomas J. Stone, Mark A. Ungless, Hywel J. Williams, Thomas S. Jacques, Juan-Pedro Martinez-Barbera. Expression analysis of adamantinomatous craniopharyngioma suggests two subtypes associated with CTNNB1 mutational frequency and highlights potential therapeutic targets. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr 1804.

Medulloblastoma: selecting children for reduced treatment

A major challenge in paediatric neuro-oncology is to accurately predict the outcome for an individual patient. Particularly, so those patients with a good prognosis can avoid life-long complications of intensive treatment. In a paper in this issue of the journal, Goschzik and colleagues investigate the best way to do this in medulloblastoma, the commonest malignant brain tumour in children [1]. Survival rates for children with medulloblastoma are impressive compared with many malignant brain tumours but complications of treatment are frequent and the children carry the burden of this disability, potentially for the many years of their adult life [2–5]. For example, one study found significant impairment on neuropsychological testing in more than 50% of medulloblastoma survivors [5]. Therefore, there is a pressing need to be able to identify children for whom treatment might be reduced without increasing the risk of tumour relapse. David Ellison, Steve Clifford and colleagues first showed in a trial cohort that activation of the WNT signalling pathway could be used to identify a subgroup of paediatric medulloblastoma with an excellent prognosis [6]. They found that patients with WNT activation had a 5-year overall survival of 92.3%, compared with 65.2% in patients without WNT activation. This result has subsequently been replicated in several studies [7–11]. However, the data available suggest that adults with WNTgroup medulloblastoma may not benefit from the same survival advantage as children [12]. The almost universally good prognosis in WNT-group medulloblastoma in children has led to the suggestion that these patients could be offered less intensive treatment. This hypothesis is being tested in the next set of clinical trials both in Europe and the USA and it is likely that some patients will seek reduced treatment off trial. A significant problem is how to identify this low-risk group with confidence. This is critical so that no child risks relapse due to insufficient treatment and none suffer as a consequence of unnecessarily aggressive treatment. Interestingly, in one focus group study, most parents of medulloblastoma patients stated a preference for ‘survival at all costs’ with aggressive treatment regardless of subsequent quality of life [13]. Finally, the success of trials to test the effectiveness of reduced treatment depends on the accurate ascertainment of WNT-group cases. WNT-group medulloblastoma can be identified by a variety of techniques: nuclear accumulation of β-catenin on immunohistochemistry, a mutation in the β-catenin gene (CTNNB1), monosomy for chromosome 6, or by gene expression or methylation signatures [14– 19]. However, at present, it is unclear which of these, if any, represent the best way to identify a WNT-group tumour for clinical diagnosis or for trial design. For example, WNT-tumours have been reported without monosomy 6 or CTNNB1 mutations [6,16]. A particular problem with staining for nuclear β-catenin is determining the percentage threshold of positive cells at which a diagnosis of WNT subtype is made, an issue confounded by variability in staining protocols and the subjective nature of interpretation. In this issue, Goschzik et al. have addressed these uncertainties by comparing three standard approaches: immunohistochemistry for nuclear β-catenin, sequencing of CTNNB1 and chromosome 6 copy number analysis. In addition, they undertook tumour subtyping on methylation arrays. In a large cohort of 186 cases, they found that immunohistochemistry for nuclear β-catenin (at a threshold of 5% of the cells) identified all the WNTgroup tumours (when compared with the other tests), suggesting that this a very sensitive screening test for WNT-group tumours. However, their data suggest that immunohistochemistry cannot be used definitively in isolation. In particular, they found that 3 out of the 21 tumours that were positive for nuclear β-catenin belonged to non-WNT subgroups as judged by methylation profiling. Interestingly, all three of these cases had β-catenin staining that was close to the 5% threshold. The authors found that all of the remaining 18 cases with >5% nuclear β-catenin positivity had mutations in CTNNB1 but in contrast, monosomy 6 was only present in 15 of the cases. On the basis of this data, the authors recommend that patients are Conflict of interest: The authors declare no conflict of interest.