Early detection and diagnosis of cancer with interpretable machine learning to uncover cancer-specific DNA methylation patterns.

Cancer, a collection of more than two hundred different diseases, remains a leading cause of morbidity and mortality worldwide. Usually detected at the advanced stages of disease, metastatic cancer accounts for 90% of cancer-associated deaths. Therefore, the early detection of cancer, combined with current therapies, would have a significant impact on survival and treatment of various cancer types. Epigenetic changes such as DNA methylation are some of the early events underlying carcinogenesis. Here, we report on an interpretable machine learning model that can classify 13 cancer types as well as non-cancer tissue samples using only DNA methylome data, with 98.2% accuracy. We utilize the features identified by this model to develop EMethylNET, a robust model consisting of an XGBoost model that provides information to a deep neural network that can generalize to independent data sets. We also demonstrate that the methylation-associated genomic loci detected by the classifier are associated with genes, pathways and networks involved in cancer, providing insights into the epigenomic regulation of carcinogenesis.

Accurate detection of benign and malignant renal tumor subtypes with MethylBoostER: An epigenetic marker-driven learning framework.

Current gold standard diagnostic strategies are unable to accurately differentiate malignant from benign small renal masses preoperatively; consequently, 20% of patients undergo unnecessary surgery. Devising a more confident presurgical diagnosis is key to improving treatment decision-making. We therefore developed MethylBoostER, a machine learning model leveraging DNA methylation data from 1228 tissue samples, to classify pathological subtypes of renal tumors (benign oncocytoma, clear cell, papillary, and chromophobe RCC) and normal kidney. The prediction accuracy in the testing set was 0.960, with class-wise ROC AUCs >0.988 for all classes. External validation was performed on >500 samples from four independent datasets, achieving AUCs >0.89 for all classes and average accuracies of 0.824, 0.703, 0.875, and 0.894 for the four datasets. Furthermore, consistent classification of multiregion samples (N = 185) from the same patient demonstrates that methylation heterogeneity does not limit model applicability. Following further clinical studies, MethylBoostER could facilitate a more confident presurgical diagnosis to guide treatment decision-making in the future.

The renal lineage factor PAX8 controls oncogenic signalling in kidney cancer.

Large-scale human genetic data1-3 have shown that cancer mutations display strong tissue-selectivity, but how this selectivity arises remains unclear. Here, using experimental models, functional genomics and analyses of patient samples, we demonstrate that the lineage transcription factor paired box 8 (PAX8) is required for oncogenic signalling by two common genetic alterations that cause clear cell renal cell carcinoma (ccRCC) in humans: the germline variant rs7948643 at 11q13.3 and somatic inactivation of the von Hippel-Lindau tumour suppressor (VHL)4-6. VHL loss, which is observed in about 90% of ccRCCs, can lead to hypoxia-inducible factor 2α (HIF2A) stabilization6,7. We show that HIF2A is preferentially recruited to PAX8-bound transcriptional enhancers, including a pro-tumorigenic cyclin D1 (CCND1) enhancer that is controlled by PAX8 and HIF2A. The ccRCC-protective allele C at rs7948643 inhibits PAX8 binding at this enhancer and downstream activation of CCND1 expression. Co-option of a PAX8-dependent physiological programme that supports the proliferation of normal renal epithelial cells is also required for MYC expression from the ccRCC metastasis-associated amplicons at 8q21.3-q24.3 (ref. 8). These results demonstrate that transcriptional lineage factors are essential for oncogenic signalling and that they mediate tissue-specific cancer risk associated with somatic and inherited genetic variants.

Locus-specific induction of gene expression from heterochromatin loci during cellular senescence.

Senescence is a fate-determined state, accompanied by reorganization of heterochromatin. Although lineage-appropriate genes can be temporarily repressed through facultative heterochromatin, stable silencing of lineage-inappropriate genes often involves the constitutive heterochromatic mark, histone H3 lysine 9 trimethylation (H3K9me3). The fate of these heterochromatic genes during senescence is unclear. In the present study, we show that a small number of lineage-inappropriate genes, exemplified by the LCE2 skin genes, are derepressed during senescence from H3K9me3 regions in fibroblasts. DNA FISH experiments reveal that these gene loci, which are condensed at the nuclear periphery in proliferative cells, are decompacted during senescence. Decompaction of the locus is not sufficient for LCE2 expression, which requires p53 and C/EBPß signaling. NLRP3, which is predominantly expressed in macrophages from an open topologically associated domain (TAD), is also derepressed in senescent fibroblasts due to the local disruption of the H3K9me3-rich TAD that contains it. NLRP3 has been implicated in the amplification of inflammatory cytokine signaling in senescence and aging, highlighting the functional relevance of gene induction from 'permissive' H3K9me3 regions in senescent cells.

Sequential inverse dysregulation of the RNA helicases DDX3X and DDX3Y facilitates MYC-driven lymphomagenesis.

DDX3X is a ubiquitously expressed RNA helicase involved in multiple stages of RNA biogenesis. DDX3X is frequently mutated in Burkitt lymphoma, but the functional basis for this is unknown. Here, we show that loss-of-function DDX3X mutations are also enriched in MYC-translocated diffuse large B cell lymphoma and reveal functional cooperation between mutant DDX3X and MYC. DDX3X promotes the translation of mRNA encoding components of the core translational machinery, thereby driving global protein synthesis. Loss-of-function DDX3X mutations moderate MYC-driven global protein synthesis, thereby buffering MYC-induced proteotoxic stress during early lymphomagenesis. Established lymphoma cells restore full protein synthetic capacity by aberrant expression of DDX3Y, a Y chromosome homolog, the expression of which is normally restricted to the testis. These findings show that DDX3X loss of function can buffer MYC-driven proteotoxic stress and highlight the capacity of male B cell lymphomas to then compensate for this loss by ectopic DDX3Y expression.

SGK1 mutations in DLBCL generate hyperstable protein neoisoforms that promote AKT independence.

Serum and glucocorticoid-regulated kinase 1 (SGK1) is one of the most frequently mutated genes in diffuse large B-cell lymphoma (DLBCL). However, little is known about its function or the consequence of its mutation. The frequent finding of truncating mutations has led to the widespread assumption that these represent loss-of-function variants and, accordingly, that SGK1 must act as a tumor suppressor. In this study, instead, the most common SGK1 mutations led to production of aberrantly spliced messenger RNA neoisoforms in which translation is initiated from downstream methionines. The resulting N-terminal truncated protein isoforms showed increased expression related to the exclusion of an N-terminal degradation domain. However, they retained a functional kinase domain, the overexpression of which rendered cells resistant to AKT inhibition, in part because of increased phosphorylation of GSK3B. These findings challenge the prevailing assumption that SGK1 is a tumor-suppressor gene in DLBCL and provide the impetus to explore further the pharmacological inhibition of SGK1 as a therapeutic strategy for DLBCL.

Transcription-dependent cohesin repositioning rewires chromatin loops in cellular senescence.

Senescence is a state of stable proliferative arrest, generally accompanied by the senescence-associated secretory phenotype, which modulates tissue homeostasis. Enhancer-promoter interactions, facilitated by chromatin loops, play a key role in gene regulation but their relevance in senescence remains elusive. Here, we use Hi-C to show that oncogenic RAS-induced senescence in human diploid fibroblasts is accompanied by extensive enhancer-promoter rewiring, which is closely connected with dynamic cohesin binding to the genome. We find de novo cohesin peaks often at the 3' end of a subset of active genes. RAS-induced de novo cohesin peaks are transcription-dependent and enriched for senescence-associated genes, exemplified by IL1B, where de novo cohesin binding is involved in new loop formation. Similar IL1B induction with de novo cohesin appearance and new loop formation are observed in terminally differentiated macrophages, but not TNFα-treated cells. These results suggest that RAS-induced senescence represents a cell fate determination-like process characterised by a unique gene expression profile and 3D genome folding signature, mediated in part through cohesin redistribution on chromatin.

A KLF6-driven transcriptional network links lipid homeostasis and tumour growth in renal carcinoma.

Transcriptional networks are critical for the establishment of tissue-specific cellular states in health and disease, including cancer. Yet, the transcriptional circuits that control carcinogenesis remain poorly understood. Here we report that Kruppel like factor 6 (KLF6), a transcription factor of the zinc finger family, regulates lipid homeostasis in clear cell renal cell carcinoma (ccRCC). We show that KLF6 supports the expression of lipid metabolism genes and promotes the expression of PDGFB, which activates mTOR signalling and the downstream lipid metabolism regulators SREBF1 and SREBF2. KLF6 expression is driven by a robust super enhancer that integrates signals from multiple pathways, including the ccRCC-initiating VHL-HIF2A pathway. These results suggest an underlying mechanism for high mTOR activity in ccRCC cells. More generally, the link between super enhancer-driven transcriptional networks and essential metabolic pathways may provide clues to the mechanisms that maintain the stability of cell identity-defining transcriptional programmes in cancer.

NF-κB-Dependent Lymphoid Enhancer Co-option Promotes Renal Carcinoma Metastasis.

Metastases, the spread of cancer cells to distant organs, cause the majority of cancer-related deaths. Few metastasis-specific driver mutations have been identified, suggesting aberrant gene regulation as a source of metastatic traits. However, how metastatic gene expression programs arise is poorly understood. Here, using human-derived metastasis models of renal cancer, we identify transcriptional enhancers that promote metastatic carcinoma progression. Specific enhancers and enhancer clusters are activated in metastatic cancer cell populations, and the associated gene expression patterns are predictive of poor patient outcome in clinical samples. We find that the renal cancer metastasis-associated enhancer complement consists of multiple coactivated tissue-specific enhancer modules. Specifically, we identify and functionally characterize a coregulatory enhancer cluster, activated by the renal cancer driver HIF2A and an NF-κB-driven lymphoid element, as a mediator of metastasis in vivo We conclude that oncogenic pathways can acquire metastatic phenotypes through cross-lineage co-option of physiologic epigenetic enhancer states.Significance: Renal cancer is associated with significant mortality due to metastasis. We show that in metastatic renal cancer, functionally important metastasis genes are activated via co-option of gene regulatory enhancer modules from distant developmental lineages, thus providing clues to the origins of metastatic cancer. Cancer Discov; 8(7); 850-65. ©2018 AACR.This article is highlighted in the In This Issue feature, p. 781.

NOTCH-mediated non-cell autonomous regulation of chromatin structure during senescence.

Senescent cells interact with the surrounding microenvironment achieving diverse functional outcomes. We have recently identified that NOTCH1 can drive 'lateral induction' of a unique senescence phenotype in adjacent cells by specifically upregulating the NOTCH ligand JAG1. Here we show that NOTCH signalling can modulate chromatin structure autonomously and non-autonomously. In addition to senescence-associated heterochromatic foci (SAHF), oncogenic RAS-induced senescent (RIS) cells exhibit a massive increase in chromatin accessibility. NOTCH signalling suppresses SAHF and increased chromatin accessibility in this context. Strikingly, NOTCH-induced senescent cells, or cancer cells with high JAG1 expression, drive similar chromatin architectural changes in adjacent cells through cell-cell contact. Mechanistically, we show that NOTCH signalling represses the chromatin architectural protein HMGA1, an association found in multiple human cancers. Thus, HMGA1 is involved not only in SAHFs but also in RIS-driven chromatin accessibility. In conclusion, this study identifies that the JAG1-NOTCH-HMGA1 axis mediates the juxtacrine regulation of chromatin architecture.

Phenotype specific analyses reveal distinct regulatory mechanism for chronically activated p53.

The downstream functions of the DNA binding tumor suppressor p53 vary depending on the cellular context, and persistent p53 activation has recently been implicated in tumor suppression and senescence. However, genome-wide information about p53-target gene regulation has been derived mostly from acute genotoxic conditions. Using ChIP-seq and expression data, we have found distinct p53 binding profiles between acutely activated (through DNA damage) and chronically activated (in senescent or pro-apoptotic conditions) p53. Compared to the classical 'acute' p53 binding profile, 'chronic' p53 peaks were closely associated with CpG-islands. Furthermore, the chronic CpG-island binding of p53 conferred distinct expression patterns between senescent and pro-apoptotic conditions. Using the p53 targets seen in the chronic conditions together with external high-throughput datasets, we have built p53 networks that revealed extensive self-regulatory 'p53 hubs' where p53 and many p53 targets can physically interact with each other. Integrating these results with public clinical datasets identified the cancer-associated lipogenic enzyme, SCD, which we found to be directly repressed by p53 through the CpG-island promoter, providing a mechanistic link between p53 and the 'lipogenic phenotype', a hallmark of cancer. Our data reveal distinct phenotype associations of chronic p53 targets that underlie specific gene regulatory mechanisms.

Soluble IFN receptor potentiates in vivo type I IFN signaling and exacerbates TLR4-mediated septic shock.

Circulating levels of a soluble type I IFNR are elevated in diseases, such as chronic inflammation, infections, and cancer, but whether it functions as an antagonist, agonist, or transporter is unknown. In this study, we elucidate the in vivo importance of the soluble type I IFNAR, soluble (s)IFNAR2a, which is generated by alternative splicing of the Ifnar2 gene. A transgenic mouse model was established to mimic the 10-15-fold elevated expression of sIFNAR2a observed in some human diseases. We generated transgenic mouse lines, designated SolOX, in which the transgene mRNA and protein-expression patterns mirrored the expression patterns of the endogenous gene. SolOX were demonstrated to be more susceptible to LPS-mediated septic shock, a disease model in which type I IFN plays a crucial role. This effect was independent of "classical" proinflammatory cytokines, such as TNF-α and IL-6, whose levels were unchanged. Because the increased levels of sIFNAR2a did not affect the kinetics of the increased interferonemia, this soluble receptor does not potentiate its ligand signaling by improving IFN pharmacokinetics. Mechanistically, increased levels of sIFNAR2a are likely to facilitate IFN signaling, as demonstrated in spleen cells overexpressing sIFNAR2a, which displayed quicker, higher, and more sustained activation of STAT1 and STAT3. Thus, the soluble IFNR is an important agonist of endogenous IFN actions in pathophysiological processes and also is likely to modulate the therapeutic efficacy of clinically administered IFNs.

Silencing of Irf7 pathways in breast cancer cells promotes bone metastasis through immune escape.

Breast cancer metastasis is a key determinant of long-term patient survival. By comparing the transcriptomes of primary and metastatic tumor cells in a mouse model of spontaneous bone metastasis, we found that a substantial number of genes suppressed in bone metastases are targets of the interferon regulatory factor Irf7. Restoration of Irf7 in tumor cells or administration of interferon led to reduced bone metastases and prolonged survival time. In mice deficient in the interferon (IFN) receptor or in natural killer (NK) and CD8(+) T cell responses, metastasis was accelerated, indicating that Irf7-driven suppression of metastasis was reliant on IFN signaling to host immune cells. We confirmed the clinical relevance of these findings in over 800 patients in which high expression of Irf7-regulated genes in primary tumors was associated with prolonged bone metastasis-free survival. This gene signature may identify patients that could benefit from IFN-based therapies. Thus, we have identified an innate immune pathway intrinsic to breast cancer cells, the suppression of which restricts immunosurveillance to enable metastasis.

Independence of repressive histone marks and chromatin compaction during senescent heterochromatic layer formation.

The expansion of repressive epigenetic marks has been implicated in heterochromatin formation during embryonic development, but the general applicability of this mechanism is unclear. Here we show that nuclear rearrangement of repressive histone marks H3K9me3 and H3K27me3 into nonoverlapping structural layers characterizes senescence-associated heterochromatic foci (SAHF) formation in human fibroblasts. However, the global landscape of these repressive marks remains unchanged upon SAHF formation, suggesting that in somatic cells, heterochromatin can be formed through the spatial repositioning of pre-existing repressively marked histones. This model is reinforced by the correlation of presenescent replication timing with both the subsequent layered structure of SAHFs and the global landscape of the repressive marks, allowing us to integrate microscopic and genomic information. Furthermore, modulation of SAHF structure does not affect the occupancy of these repressive marks, nor vice versa. These experiments reveal that high-order heterochromatin formation and epigenetic remodeling of the genome can be discrete events.

Spatial coupling of mTOR and autophagy augments secretory phenotypes.

Protein synthesis and autophagic degradation are regulated in an opposite manner by mammalian target of rapamycin (mTOR), whereas under certain conditions it would be beneficial if they occurred in unison to handle rapid protein turnover. We observed a distinct cellular compartment at the trans side of the Golgi apparatus, the TOR-autophagy spatial coupling compartment (TASCC), where (auto)lysosomes and mTOR accumulated during Ras-induced senescence. mTOR recruitment to the TASCC was amino acid- and Rag guanosine triphosphatase-dependent, and disruption of mTOR localization to the TASCC suppressed interleukin-6/8 synthesis. TASCC formation was observed during macrophage differentiation and in glomerular podocytes; both displayed increased protein secretion. The spatial coupling of cells' catabolic and anabolic machinery could augment their respective functions and facilitate the mass synthesis of secretory proteins.

A re-annotation pipeline for Illumina BeadArrays: improving the interpretation of gene expression data.

Illumina BeadArrays are among the most popular and reliable platforms for gene expression profiling. However, little external scrutiny has been given to the design, selection and annotation of BeadArray probes, which is a fundamental issue in data quality and interpretation. Here we present a pipeline for the complete genomic and transcriptomic re-annotation of Illumina probe sequences, also applicable to other platforms, with its output available through a Web interface and incorporated into Bioconductor packages. We have identified several problems with the design of individual probes and we show the benefits of probe re-annotation on the analysis of BeadArray gene expression data sets. We discuss the importance of aspects such as probe coverage of individual transcripts, alternative messenger RNA splicing, single-nucleotide polymorphisms, repeat sequences, RNA degradation biases and probes targeting genomic regions with no known transcription. We conclude that many of the Illumina probes have unreliable original annotation and that our re-annotation allows analyses to focus on the good quality probes, which form the majority, and also to expand the scope of biological information that can be extracted.

INTERFEROME: the database of interferon regulated genes.

INTERFEROME is an open access database of types I, II and III Interferon regulated genes (http://www.interferome.org) collected from analysing expression data sets of cells treated with IFNs. This database of interferon regulated genes integrates information from high-throughput experiments with annotation, ontology, orthologue sequences from 37 species, tissue expression patterns and gene regulatory information to enable a detailed investigation of the molecular mechanisms underlying IFN biology. INTERFEROME fulfils a need in infection, immunity, development and cancer research by providing computational tools to assist in identifying interferon signatures in gene lists generated by high-throughput expression technologies, and their potential molecular and biological consequences.