not invented here.

Notes on bioinformatics methods, software and discussions.

[fiamh-blog]

BTG 2014 Session 7: Bioinformatic analysis of cancer genomes

The final session of the meeting, chaired by Nuria Lopez-Bigas, starts with the Bioinformatics Challenge results from Mike Schatz.

Pan-cancer network analysis of combinations of somatic mutations

Ben Raphael (Brown University) follows Mike with a slightly revised talk on mutational heterogeneity. Recap of driver mutations, intra-tumor heterogeneity, long tail discussion and inter-tumor heterogeneity. Announces a “whirlwind tour” of methods developed in his lab (uhoh).

Can we infer tumor composition from single, mixed tumor sample? Number of SNV- and CNA based methods (shows dozen of references from ABSOLUTE to TITAN to PyClone and more, 2014 the year where this exploded). Copy number aberations used to be based on SNP arrays, shift to NGS. Detect subtle shift of red depth, combine signals across multiple aberrations and subpopulations (probabilistic algorithm THetA, recommends reading the paper for optimization tricks, works for ~7x WGS data). Example from ~200x BRC samples with 3 subpopulatins and how THetA automatically deconvolutes them; another example from whome-exome data (compares favorably to ABSOLUTE).

Wanted to go further to study tumor evolution process using variant allele frequencies. Noisy for single mutations at lower coverage, but can be clustered across genome (dirichlet process mixture models) and used to create the original tree. Can’t measure leaves without single cell data, but can convert allele frequency histogram to create the evolutionary tree using binary tree partitions (assumption: only one clonal expansion event at a given time). Simulated data examples look reasonable (but only around 50% precision at 8+ populations).

Next compare mutations across different genomes. Ranked list of significantly mutated genes — how to handle the long tail? Genes don’t act on their own, univariate analysis not ideal. Analyze gene groups instead and use topology of gene interactions: aim to discover pathways given enough samples, but can’t do it without any prior information. HotNet and Dendrix work on that intersection to find significantly mutated subnetworks. Consider gene score and interaction scores at the same time (HotNet, HotNet2 using heat diffusion methods to spread significance scores across network, with HotNet2 using directionality information). Applied to TCGA pan-cancer project to find consensus subnetworks, e.g., cohesin complex, condensin complex, SWI/SNF identified. Mutual exclusively another story they have been working with. What about mutually exlusive gene sets (Dendrix and now multi-dendrix for multiple pathways, and Dendrix++ with probabilistic scoring).

Also started looking into some visualizing approaches with d3.js just out of necessity. Overall aim is to combine different data types, get better at stratifying samples.

Ginkgo: Visual analytics for single cell CNV analysis

Somehow I missed the announcement of Tyler Garvin’s talk (from Mike Schatz’ lab) . Quick summary of single cell sequencing, CNVs in cancer. Underlying concept are CNV profiles for each sample/cell, allowing you to cluster patients or individual tumors and explore the resulting tree. Data is quite noisy due to whole genome amplification, wet lab protocols, sequencing biases. Correct for outliers, provides lots of QC metrics for each sample/cell, clustered and visualized. [Followed by demo of the ginkgo software using data from Navin et al (2011). Framework looks good, but again hard to live-blog.]

Validated with five public data sets, easily reproduced results in ~30 minutes. Compares GC biases across different data sets and data dispersion, and the need to correct for these. Couple of recommendations: use DOP-PCR for WGA, sequence healthy diplied cells, plan 25% cells to be filtered out and plan accordingly, save your FACS data (get estimate of ploidy from staining), start with 500kb bins, 100 reads/bin (1.5 million reads per cell) and do control for gender.

Mapping intra-tumor heterogeneity using multidimensional single cell data

I’m scared of trying to cover [Dana Pe’er]’s talk (Columbia University) given the amount of information she tends to pack into a single presentation. Talk on work with the Silva lab (Felix Sanchez-Garcia). Unclear what went wrong for a primary tumour (dissecting the plane crash: what went wrong, what do the broken parts do), vs functional approach (analyze individual features). Integrative approach (Helios) to combine the best of both approaches. Gives examples of oncogene addiction, Helios scores neighbouring genes to come up with best candidates. Training sets too small and biased, using an unsupervised model with frequency of alterations as key. Quick detour to copy number modeling and how to integrate these and other features (EM feature learning to identify important features and associated drivers).

Shows how it pinpoints BCL2, 6th based on CNA but oncogene addiction raises it to top of peak. Looked at 17 most frequently amplified regions and test them one by one. In vitro validations tends to score known candidate genes highest, cloned picked genes for genes without known targets in six replicates to see if they drive tumor development. 10/12 genes validated, overall 93% accuracy.

Switch themes to intra-tumor heterogeneity. Mass cytometry as a game changer looking at millions of single cells, label biomarkers with metal isotypes, 45 dimensions simultanously measured in millions of individual cells (see viSNE paper for details on data analysis of healthy immune system, detecting rare cell populations in ALL and more). Now to the unpublished stuff (with Nolan lab at Stanford). Ceullar profiling of an AML cohort vs healthy bone marrow. 28 samples, 19 perturbations, 16 million single cells. Must clustering approaches fail for this kind of data. Don’t know number of clusters, size assumption, unclear distribution. Graph-based clustering to tackle the toplogy, each cell a node connected to most similar cells (similar to already published work). Their ‘PhenoGraph’ seems to be doing better than competing methods (tested on healthy data). AML graph landscape is mixed between patients. Each patient resides in multiple areas of the map, each region contains multiple patients. Some structure to the aberrant surface expression of AML, visible when color-labelled by surface marker expression: developmental gradient apparent.

Try to find ‘metaclusters’ shared across patients, iterate on graph to find clusters of clusters. Each patient has its own combination of (repeated) metaclusters. Samples with similar genetics have similar meta-cluster profiles, i.e., genetics is driving some of the big structure but not the difference between the subpopulations. These seem to map to intracellular signaling structure — response to drugs, growth factors and other perturbations. Matrix of 16,000,000 x 31 cells reduced to tiny fraction to learn surface-signalling axis is disrupted in AML, some CD34- populations signal like stem cells (surface, signalling link uncoupled). Signalling data can be used to classify healthy cells successfully. Try to find AML subpopulations signalling like healthy cells [and I got lost at this point.] Possibly new definition of ‘primitive’ cell, % greatly differs between samples and does not agree with CD34+ fraction currently used to describe them. Map to gene signature, then used gene signature to test in much larger data set, linked to survival.

[I need a break now…]

Structural variation analysis of tumor genomes

Peter Park (Harvard Medical School) with the last ‘full’ talk of the session (another disclaimer: like Nils Peter is a collaborator). Quick summary of SV events in cancer (refers to Nils who made some of the graphics of different events). First part of the talk on published work regarding complex SV events, e.g., duplication with deletion. Need to look at all events combined to understand the actual event. Tons of manual curation to find unusual / novel cases. Found SV in TERT promoters (part of the TCGA work), asked whether they can infer mechanism based on sequence homoogy at breakpoints (fork stalling, template switching, non-homologous end-joining, etc., see work from Kidd et all, Cell 2010 for HapMap). Try to do this in short-read cancer data. Nice summary of somatic SV types and mechanism spectrum (by frequency) from 140 samples.

Characterized GBM, reconstructed complex re-arrangements in detail. About 20% of somatic deletions are complex formed by replication errors. Many focal deletions caused by FoSTeS/MMBIR, both DNA DS-breaks and replication erros drive somatic complex events. Amazing reconstruction graph involving 45 events that might explain a single EGFR event, trying to proof that this is the only way to get there, but no automated way to generate these yet.

Switch to transposable element analysis pipeline (see paper), checking for somatic mosaicism in brain based on data from Chris Walsh. Is there somatic variation in different neurons, is there retrotransposition? Genome mosaicism driven by retrotransposition might reshape genetic circuity. Whole genome amplification, sequenced L1-sequencing of 300 single neurons to detect L1 insertion sites. Observed ~.6 insertions / cell, much lower than previous publications. Sequenced 16 cells at 40x (MDA) compared to populatin of 100 neurons and bulk data. Compares MDA vs MALBAC, biases. Retrotransposon detection workflow in the paper, show L1 insertion in same places across muliple neurons (expect to be the case if they have the same progenitor, validated). Think close to 80% sensitivity in single cell compared to bulk. Found additional insertion missed by L1-seq (longer than DNA-fragments captured by L1). See another recent paper on impact in cancer genomes.

Other bits and pieces: nozzle shown, used by the Broad for Firehose (and other NGS reports). Shoutout to Refinery, a repository framework on top of Galaxy (work with Shannan Ho Sui in our group). Facetted browsing for modENCODE data. Followed by list of ~10 other tools (Meerkat, BIC-Seq, StratomeX again, etc.).

Using de novo assembly to exhaustively catalog tumor mutations

A selected talk from David Jaffe (Broad Institute) on mutations only seen by de novo assembly. Start with cheap data, 0,5ug of DNA, PCR-free library, paired end read 50x, 250bp from breast cancer cell line, but also tested on primary tissue. DISCOVAR de novo is the algorithm, tumor and normal assembled together, mutation caller still work in progress and unpublished. Walks through the graph assembly algorithm. Bubbles represent alternative paths including somatic mutations. Look for tumor only edges, reads from just from the tumor not the normal sample. Compare against MuTect, Strelka, Taiga. [Live demo having minor difficulties due to having to reboot his Mac and of course getting logged out of all his sessions]. DISCOVAR finds ‘funky’ nversions not found by standard methods — reads can’t be aligned. Also found a ‘crazy quilt’ rearrangement (love the names). Re-arrangement supported by 140 reads in tumor, trace in assembly which goes through 13 segments all across the genome, each 100-200 bp.

A general mutation finder that sees invisible somatic mutations, more than 500 of these found in HCC1143. Want to test drive on interesting samples, volunteers welcome.

I did have to miss the last talk of the meeting from Hossein Khiabanian on Moduli spaces of phylogenetic trees describing tumor evolutionary patterns unfortunately.

Thanks to the organizers and chairs for a terrific meeting. Just the right size for networking and a good mix of talks!

[fiamh-blog]

BTG 2014 Session 6: Data visualization and big data challenges, part 2

Split this session into two as there are a lot of talks.

Visual Exploration of Genomic Data

Helga Thorvaldsdottir from the Broad Institute resumes the session after the coffee break, tackling big data challenges in the context of visualization. Genomic data: anything specified as genomic coordinates and a reference genome. Exploration: interactively browsing through representation of the data with the purpose of finding ‘stuff’; re-inforces a number of key concepts from Cydney’s talk. Examples via IGV and how non-reference bases are highlighted in different colors, de-emphasis of low quality calls, highlight on interesting events in the summary plots. Dynamic re-sorting of reads by base calls, drill down to get numeric summary information.

Switch to structural visualization using paired-end sequencing data; cartoon visualization of structures such as inversions and what this means for reads being in the ‘wrong’ orientation; flagged and colored in IGV to help with breakpoint visualization, can also flag inter-chromosomal re-arrangements.

Next switch to use multiple samples: glioblastoma from TCGA, showing CNV data from ~200 patients in a whole genome view and an easy to spot global pattern. Overlaid with somatic mutations works well due to sparseness of the data again making it easy to identify recurrent events. Can still zoom in to individual genes and events to identify what might be affected by focal events. Need an intermediate high/low level view to understand how events relate, done by splitting window into panels showing three views side-by-side. Add metadata from samples such as clinical parameters and use these to sort views [would be nice if one could test for association on the fly, actually]. Split panes again, this time horizontally, by subtype and sort by genomic features. Last example switches to RNA-Seq with similar requirements. Can change what is displayed for example by switching to an ‘exome view’.

Guided visual exploration of patient stratifications in cancer genomics

Nils Gehlenborg from Harvard Medical School is next [Disclaimer: Nils is a collaborator], talking about StratomeX and other visualization approaches. Replaced the intro of this talk with a joke which I’ll try not to capture — not like it’s going to work without the visualization. Followed by a quote (chain as strong as the weakest link), with human being the weakest link in the data generation — data analysis — human interaction trifecta. But we need knowledge and cognition which happen on the human side, IBM’s Watson notwithstanding. Not enough tools to remove/augment the weakest link yet.

Visualization an external cognitive aid to make sense of (numerical) data, e.g., how often do two functions intersect. Case study on subtype detection in cancer to improve diagnosis and treatment, all leading to precision medicine. Again uses TCGA mRNA, microRNA, DNA methylation, CNV, mutations, protein expression and clinical data. Can stratify on any single data type via clustering or more sophisticated methods, different number of clusters for each data type. Coins the ‘Stratome’ (along with a slide of The ‘ome Depot’) specifically for Jonathan Eisen’s benefit and a link to StratomeX (and the concepts it is built around). Compare data sets across the stratifications with a hands-one demo of the glioblastoma data. [Again difficut to blog without the slides/images readily available, but in a nutshell links different stratrification columns by data types with ‘bands’ to show how the patients are distributed, intersections of data sets.]

Key question is what makes the data sets or observations different? How to compare data sets within each stratification and across? Example of age boxplot vs mutation status, add survival information (Kaplan-meier plots), mixing stratification columns with dependent columns that summarize information obtained from stratified column. Can add ‘omics data as heatmaps to check for matching clusters in expression data. Adds in pathway maps to annotate expression clusters with function.

How to find relevant stratifications and pathways de novo if you can have thousands of different stratifications? From knowledge driven exploration to a (hopefully unbiased) data driven exploration? “Guided exploration” let the users chose clusters and ask the computer about any other sets of patients in any other set of stratifications that overlaps the chosen cluster. E.g., is there a mutation that overlaps this mRNA cluster? A mutation that is mutually exclusive to the one I picked? Rank results and visualize using multi-attribute ranking approaches (originally developed to compare University rankings). Nils now runs through a quick demo of Query Wizard to find stratifications that overlap with expression clusters, pulls out miRNA clusters (at least one cluster overlaps strongly with an mRNA cluster) based on overlap of samples in that cluster. Check for survival overlap with log-rank test, etc. Enhances the ability of the human weak link to interrogate the data.

Part of the Caleydo visualization framework which also has tons of other tools. Available as binaries and on GitHub. Planning to extend as ‘Domino’ in two dimensions to be released in the late fall.

[Sidenote: awesome slides as always from Nils.]

Allele-specific quantification of structural variations in cancer genomes

Sessions closes with two selected short talks, first one by Jian Ma from the University of Illinois at Urbana-Champaign. Quick intro to SV in cancer genomes. WGS data processed with Delly, Meerkat, Crest and Breakdancer, struggly with aneuploidy. Some tools such as ABSOLUTE can handle allele-specific copy number, but do not handle SV; room for new methods in-between. SVs can further modify the aneuploid genome into a mixture of genomic segments with a rather extensive range of CNVs. How do SVs interact with alle-specific CNVs?

Quick summary of the method using probabilistic graphical models, constructing the cancer genome graph from genomic segments, SNP-linkage and LD information, corrects for mappability, GC content to adjust for intrinsic read coverage differences. Simulations look decent across a range of coverages. Case study of ASCN and SVs in MCF-7; 33% of regions have copy number > 1, 68% of regions with imbalanced copy number, almost 300 SV likely happened after whole chromosome duplication.

[And I got distracted during the final slides… not quite sure this is available.]

Clustering tumors based on patterns in alternative splicing

Final talk of the session is another short talk by Elizabeth Purdom (University of California Berkeley). Introduction to different transcript isoform patterns between patient; bias of isoform usage between different patient subgroups, question is whether these patterns are common and shared between different genes?

First approach, cluster patients on isoform instead of gene expression result in roughly similar clusters, but unclear whether differences are real or just deviations of clustering. Transform by clustering on differences between isoforms, or rather relative usage (in the case of several isoforms): what proportion of an isoform is contributing to the gene expression in a given patient, then cluster by relative usage (which requires special clustering approach as we are now clustering vectors of different lengths). Applied to AML test set identifies set of genes with strong shared patterns of isoform switching (ANKHD1, BAG6, KDM5A, others).

[fiamh-blog]

BTG 2014 Session 6: Data visualization and big data challenges

If anyone wondered what happened to the IBM Keynote or the SevenBridges talk (both which seem to have been terrific): conference calls. Sigh. Back to BTG and the second to last session, to be chaired by Jan Aerts. Mike Schatz gave his by now traditional bioinformatics challenge, then we are off to the talks.

Re-inserting human interaction into cancer genome interpretation

Cydney Nielsen from the University of British Columbia starts with background information on augmenting human reasoning with intelligent systems and data visualization. Can be enhanced by interactivity which helps to better understand data and structure; required to make sense of overwhelming text data sets (nicely displayed by a VCF text dump). Example of human genome assembly which required sophisticated visualization frameworks to be able to finish the genome, identify and close gap, detect problems: computer automation followed by human expert curation, quite often iteratively. Can result in better automization, e.g. tools such as AutoFinish. Can also direct automatic algorithms to specific regions (‘focusing’ of computer resources).

Second example: cancer genomics data integration with patient metadata. View SNPs, CNVs, re-arrangements and Co in the context of the clinical information. Tried using a number of the tools out there to solve analysis problems and came up with a list of requirements which they could not find anywhere.

Flexible integration of views. Integration of different data types (here from MutationSeq, Titan — SNPs and CNVs), want to explore mutation in CNV context and expression information

Compare data filters for single data set. Impact of SNP probability thrshold on coverage and allele ratio distribution

Explore views of different data types. Be able to see different data types (from one or muliple samples) and explore connections

Built a component-based systems (data, views) still in prototype stage [wish I could link to screenshots somehow as Cydney walks us through the interface; looks clean and uncluttered]. Different viewer, filtering options, interactive exploration, dynamic linking to additional information from viewers. [The video is also hard to blog.. sigh :) ]. Hints at API to generate new viewers.

Last feature they were looking for: scalability. Most address perceptual and interactive scalability (how to enable dynamic querying and rendering of millions of data points in real time). Tackled with ‘search’ rather than a traditional database approach. Document-driven approach, query performance a function of the number of query matches, not the data size. Tends to work well for load-once-read-many workflows. Documents are records, e.g., a mutation document with different fields (sample ID, position, change, ..). Prototype has 800 million events, optimization in process. Still standardizing loaders and adding APIs. Perceptual scalability is the problem of having more data points than pixels: how to present meaningful summaries. Prevent overplotting, for example.

Goal of the project is to facilitate scientific discovery, means to support data exploration.

Genomic media and clinical cancer medicine

Eliezer Van Allen (Dana-Farber Cancer Institute) starts with a review of precision cancer medicine, prospective clinical interpretation (matching the patient to the right drug), try to understand why patients develop resistance based on genomic information, retroactively understand responders/non-responders and improve the prospective options in the next round.

Paradigm shift in clinical cancer medicine from non-specific chemotherapy to testing for tumor-specific targets that can be drugged along with targeted therapy. Hard to do in the clinic due to practical and logistical challenges (profiling, interpretation, communication and decision making). [Woot. Day 3 and first show of NHGRI sequencing cost slide, that took a while.] Augments cost slide with data points per patient and how those changed during his MD career: in the beginning only had to keep 10 data points in memory (imaging, clinical parameters). In 2010 hotspot mutations in selected genes get add, but number of parameters increases to 100, none of which was taught in med school. Now easily 10-100k data points that could be evaluated. Needs to be summarized for clinicians. Survey on how confident clinicians are with genomics data, majority in the ‘somewhat confident’ category — but that might very well be overconfident based on anectodes Eliezer shares. [Which are on the scary side.]

How to make data approachable? Better data visualization, add to clinical record to make it usable in real time. PHIAL (Precision Heuristics for Interpreting the Alteration Landscape) as a development to tackle this [named after Garraway's Tolkien suggestion]. How do you condense an exome worth of data into a single page at most? Presents a ‘gel’ like visualization with clinically important variants at top, linked to pathways; makes it easier to distinguish highly actionable from less important mutations. Fulfills the ‘at a glance’ requirement, but most clinicians miss the genomic expertise (as do most patients) (the “Bob Mayer” problem - you can’t send clinicians a plain VCF and expect them to figure things out). But how do you get information about genomics into clinical record systems built in 1991 (as used by Brigham)? PDFs state of the art [ouch…], ‘huge step in the right direction’ but not exactly interactive (refers to previous talk). Medical and legal issues with non-static reports, but we lack best practices to make cutoff decisions. Works with Nils on CGEC (cancer genome report) to create web-based reports. Includes summary, but also pointers to raw data, IGV screenshots, etc. — helps overcome clinician’s trust issues. How to port some of that into EMR is an entirely different story, need to encourage health record vendors to work with the community to allow integration with genomic data, include visualization that might not be common for electronic health records.

Field in its infancy: needs standardization, best practices, prospective tests in the clinic. Encourages collaboration.

[fiamh-blog]

BTG 2014 Session 5: Beyond the cancer genome

Chaired by Mike Schatz who also announces the now traditional computational challenge.

Understanding intrinsic and extrinsic control of intra-tumor phenotype switching

Talk by Rosalie Sears (Oregon) presenting on a collaboratie project with the Gray lab. Again an intro to tumor heterogeneity along different axes of differentiation (stem/differential, luminal/basal, mesenchymal/epithelial), but probably a combination of all three axes. The goal is to understand and manage heterogeneity.

Useful public data on cell lines and patients. Problem that we are not getting full response to single drugs even in cell lines, reflected by heterogenous lineage marker expression in basal like breast cancer cell lines. For this project taking an imaging approach to add to the ‘omics world, provides insights into resistance mechanism. Luminal, basal, mesenchymal markers again expressed differently across cells in different cell lines. Workflow expands heterogeneous cell lines, screen 120 kinase inhibitor compounds, measure proliferation and apoptosis, marker expression. Found that targeted therapeutics induce dose dependent alterations of lineage marke expression, and that classes of inhibtors differentially affect lineage marker expression. The phenotypic shift is accompanied by an (expected) pathway change. These phenotypic shifts can happen within 12 hours (depending on inhibitors), and cells recover (switch back to heterogenous state) after removal of inhibitor.

From cell-intrinsic to extrensic model work. Array screening to interrogate thousands of protein interactions in one assay to test different microenvironments of different ECMs/cytokines, readout by fluorescent assay. The environment clearly influences cellular heterogeneity, can steer them towards a certain phenotype. Impacts sensitivity to different drugs. Further explored with Novogen MMX 3D tissue bioprinter to create cellular tissues that can be treated like regular tissue (‘bioprinted breast tumors’). Can last for weeks and can be used for therapeutic testing, contrast with heterogenous tripple negative cell lines.

(Some neat technology in this talk, difficult to confer all the information while liveblogging).

The human microbiome – cancer link

From Sarah Highlander (J. Craig Venter Institute), picking up the thread from the previous talk on other extrinsic confounding factors — the microbiome. Short intro to the second genoe / human microbiome and their diversity across body sites, impact on digestion, immunity development, pathogen interaction. Shifts in population can be indicators for health status changes. Profiled through 16s rRNA sequencing and whole genome shotgun sequencing. Shoutout to Human Longevity Inc. Recap of human microbiome project, sampling of body sites, different phylogenetic composition but conserved metabolic representation (paper), explains the concept of dysbiosis (microbial imbalance).

Number of cancer links, with Helicobacter pylori the only established one. Virulence fators promote inflammation, DNA damage, epithelial cell transformation. Association projects in progress include progession of esophagael cancer, colon cancer and microbial inflammation in pancreatic cancer. Some of the hints come from animal models; germ free rats have fewer colorectal tumors than controls; antibiotic treatment can reduce tumor formation in mice, but rodent models not ideal predictors for what happens in humans. Proposed mechanism of cancer promotion by the microbiome is dysbiosis/inflammation, loss of barrier function and engagement of TOLL like receptors through microbial ligands or general genotoxicity, none have been shown yet in humans (see paper for details).

Liquid biopsies as non-invasive tool to monitor tumor genomes

The drawback of having a conference in the city you work in is that you get dragged in to meetings despite being ‘out of office’. Missed the talk by Michael Speicher, unfortunately.

Genetic similarity between cancers and comorbid Mendelian diseases identifies candidate driver genes

The first seleced short talk in this session is given by Rachel Melamed (Columbia University) with an intro to familial mendelian cancer syndromes (which are very rare). Looking at mendelian diseases with a risk of common cancer (e.g., Rubinstein-Taybi syndrome). Are the same functions altered in comorbid cancer? Sift through electronic health records to find rare cases in 110 million cases. Check comorbidity of detected cancer (from TCGA) and their recurrently mutated genes against OMIM germline diseases and check for genetic similarity using a number of in-house metrics. Is there significant overlap for paired diseases? or PPI modules, pathways? Co-expression of cancer/mendelian genes?

1/3 of comorbity statistically associated with genetic similarity. Cancer is more likely to come with a mendelian disease that alters similar genes. Melanoma example and its link to albinism with candidate genes around MITF, MC1R. Pan-cancer mendelian associations distinguishes about 25% of mendelian diseases with no cancer, some with only 1 (albinism - melanoma, for example), but some linked to a large number of different cancers (linked to core cancer progresses, e.g., apoptosis defects in mendelian diseases). Novel source of information to understand cancer processes.

Bacteria-human somatic cell lateral gene transfer is enriched in cancer samples

Second and final selected short talk from Julie Dunning Hotopp (University of Maryland Baltimore). Leads with spiderman as the typical example of lateral gene transfer. Kudos! Not talking about inheritance but what happens in single cells as a somatic event; pieces of DNA can move into eukaryotic cells. If this cell — for whatever reason — is transformed and expands clonally you should be able to detect insertion (transfer) event. Some evidence found in blood samples. Take PCR artifacts, other issues into account. Calibrate pipeline in HeLa vs HPV, 30k paired end reads with both reads mapping to HPV. Applied to subset of TCGA data; bacterial integration enriched in tumor samples, but these are not standard normal/tumor pairs (not perfectly matched). Particularly present in LAML, STAD cancers. In STAD (stomach) integration of pseudomonas in four different genes (all data available). In summary, not just viruses but also bacteria place their genome data into host cells. Worth checking bacterial component of tumors being screened already.

Thanks data providers as noone would give her money to chase this on her own.

[fiamh-blog]

BTG 2014 Session 4: Translation and clinical application

Another short session just before lunch.

Drugging the cancer interactome: rational target selection for drug discovery

Presentation from Bissan Al-Lazikani (The Institute of Cancer Research, London, UK) starts with a quick intro of CRUK’s efforts in drug discovery and hurdles in drug discovery (target selection, biological validation, tracability for drugging, clinical efficacy). It’s an exciting time given the availability of large scale data for cancer gene identification (see previous talks), but every study identifies new/non-overlapping lists of cancer drivers resulting in a rather large target space. Try to assess targets through all stages of discovery, i.e., balancing risks (technical and biological); find targets with solid initial rational, also make sure it’s likely to be druggable, non-redundant, etc. Precedence, 3D-structure evaluation, chemistry assessment, network position as a multifaceted target assessment workflow, all depend on large contextual database (available as canSAR).

Quick survey of different steps, e.g., cavitiy assessment in 3D structures and how well ligands could bind (including decision trees from positive/negative examples), identifying viable sites (such as novel allosteric sites). Global approach to look at multiple structures, every protein family member to build evidence profiles (see recent paper for a good summary showing 123 out of 489 putative targets with direct evidence for chemical tractability, including 46 with no prior work — some of which have now been successfully targeted such as BRG1 and IDH1; BRG1 is a tumor suppressor with two druggable domains).

Switch to case study on Rhabdomyosarcoma, a rare childhood cancer without a targeted therapy (hence neglected), limited to surgery and radiotherapy. Collated data from ~100 patients: gene expression, survival, came up with list of potential target list, scored interaction network and identified best chemical tractability scores to enrich for biologically relevant targets that are hopefully druggable. 41 targets with 44% hit rate which is very high indeed; drug screens in cell lines and xenografts now starting. Interesting network approach: the usual hairball but colored by druggabiity, pinpointing cancer drug targets, the targets of non-cancer drugs, and predicted druggable targets currently not covered. Find communities that are not well covered by existing drugs (the social network approach); cancer drug targets tend to occupy distinct subgraphs within the network. Try to find drug combinations hitting complementary areas of the network to reduce resistance development.

Genomic medicine – transforming cancer research and care

Lynda Chin (MD Anderson Cancer Center) is next; she switched talking slots with Michael Speicher who now talks in the next session. We should have a fairly complete catalogue of relevant genes within the next year or two; still away from complete functional validation, much less an understanding of the genome. Even further away when looking at epigenomic alterations. Quick sample case of a chromatin-mark differences between responding and non-responding lesions from a patient drove the decision for a global chromatin mark survey. ChIP-Seq of 36 different marks, but the challenge then is to define chromatin states for different phenotypes (normal, tumor, metastasis), and this is more difficult when only looking at one or two individual states on their own.

Started with isogenic melanocyte cell model systems (HMEL-BRAF, PMEL-BRAF) with different states (non-tumorigenic, tumorigenic, metastatic) to explore the function of the epigenome integratively: test for 35 marks, run RNA-Seq, total of ~4 billion reads. Usual genome segmentation approach via ChromHMM, condensing all data into 18 states (highly acetylated promoter, transcribed enhancer, poised TSS, etc.). Verified by lining up state predictions with known, annotated genomic regions. Check for changes in state transition between different phenotypic conditions, e.g., NT vs T across all promoters characterized by a loss of acetylation marks, H3K4 methylation marks. Link to acetylation change to RNA-Seq expression changes show no significant correlation (which is a bit surprising): deactylation changes are unidirecton, but gene expression changes are bidirectional. Deactylated and downregulated genes are enriched in cancer pathways. Response likely not dictated by a single chromatin signal.

Use the same model system to generate loss of function and gain of function libraries (475 shRNA contructs of 95 epigenetic enzymes), inject in non-tumorigenic cells to identify which can cause a change towards tumorigenicity. Knockdown of just two histone acetyl transferases and four H3K4 modifiers induce the tumorigenic phenotype. Check whether inhibition of these targets recapitulate the epigenetic landscape? MLL2 and Taf1 knockdown seems to recapitulate the acetylation changes seen in tumorigenic cells.

Ideal experiment is to go in with a tumor sample and do all marks. Only been able to do this with 8 TCGA samples so far giving them a complete epigenome and genome map. Still sequencing, at 6 billion reads not reached saturation yet. Other interesting questions can be asked.. what is the higher order of DNA structure? How does the epigenomic change map to those mapped in the NIH Roadmap program (adult vs fetal vs ES cells).

Good final pitch on complexity of cancer. We just look at one mutation rather than the whole tumor system, the microenvironment, genetic background and social/environmental aspects. Need to integrate the research in context of clinical care — the Apollo project.

Lessons from the study of extraordinary responders

Final full talk of the session by David Solit](http://www.mskcc.org/research/lab/david-solit) (Memorial Sloan-Kettering Cancer Center) around speeding up drug discovery. Genotype to Phenotype (G2P) to define targets. Started working on BRAF / melanoma in 2002, but the identification takes a large cohort. How to work with existing data? Test retrospectively, first applied to the response to everolimus in a blader cancer patient which did not work (2 out of 45 patients responded). One patient still alive after 5 years, question to ask then is: why did this patient responded extremely well to mTORC1 inhibition? Could not figure out with targeted sequencing, but WGS showed 17,000 somatic mutations (!), 140 NS coding mutations (paper here). Perfect patient for the drug due to TSC1/2 mutations.

Since gone on and repeated this in other unusual cases (ovarian cancer patient on AZD6244, mutation missed initial due to low cellularity, complexity of the MAP2K1 mutation; other rare and quite extraordinaty cases) [and I’ll skip the details as most are in press / published and I am way out of my depth here]. In brief, use WGS outlier analysis to salvage drugs with low response rates; profile patients with TCGA pipeline and see if those cases can be solved.

Can this be done prospectively instead? Usually based on genotyping patients with a typical disease and allocate patients to clinical trials. Problem is drugs spread across multiple companies, negotiations take ages, drugs often not best in class but just what is available. Ethical problem of randomizing patients when you see response rates in other trials. Finally, number of patients is low (states that 200 patients is ambituous) making it difficult to assess low frequency subtypes.

Alternative approach: basket study. Different cancer patients in same trial, what percentage of lung / breast / ovarian cancer responds to a given drug (see summary)? Challenge of getting multiple disease teams to work together, and identifying patients remains a challenge. IMPACT approach at MSKCC to tackle this. Primary goal to identify molecular drier in every patient beginning with 12,000 patients/year (!). Beyond Foundation Medicine testing to ensure tumor/normal testing. Fun summary of discrepancies between Foundation, their studies, Myriad profiles depending on how germline mutations are handled (Myriad does not report mutations >1% frequency in population).

Bioinformatic analyses approaches for personalized oncogenomics

Katayoon Kasaian (British Columbia Cancer Agency Genome Sciences Centre, Canada) closes the session with a selected short talk. Leads with overview of British Columbia and the patient network. Personalized Oncogenomics (POG) program since 2012 to augment therapeutic decision making in incurable malignancies. Genome/transcriptome analysis, WGS of blood, same for archival samples, sugest therapeutics based on identified cancer drivers and aberrant pathways. Track clinical data and response.

123 patients enrolled in first two years of the program, 81 finished. ~40 days from biopsy to final report on average (goal to shorten this). Lessons learned include that genomic approaches (whole genome) are superior to standard / current clinical tests (Katayoon provides a few examples that I’ll skip here, mostly revolving around mutations missed in other assays such as FiSH). Found WTS necessary and complementary to WGS, particularly for pathway aberration analysis. Useful to correlate copy number change and expression. Case of colorectal cancer patients with KRAS G12V mutant, but WTS found no expression of KRAS (neither wildtype nor mutated) which would have changed the treatment plan. Also found high heterogeneity between tumors from different sites and over time, at times resulting in different treatment options against different tumor locations (rather than one global treatment).

Survey found POG to be informative for 87% of cases, treatment offered to 55% (but since these are end stage patients about 20% die before they get to this point). Future directions for POG 2.0: approval for 5000 cases over 5 years, increase speed and accuracy of prediction, xenograft validation, etc.

[fiamh-blog]

BTG 2014 Session 3: Integrated systems and network approaches

Chaired by Jan Korbel. I did skip the first (sponsored) talk from Bio-Rad laboratories but made it for the first regular talk of the session. Christopher Plass couldn’t make it, unfortunately, but the talk title remained similar.

Epigenetic reprogramming in glioblastoma

Mario Suva (MGH) stepped in for Christopher.

Dissecting cellular state in glioblastoma: the most frequent brain tumor, peak incidence betwee 45 and 70 years, poor prognosis. Assumption of different cell populations, solid malignancy possibly following the cancer stem cell model with genetic and epigenetic changes (though disagreement on surface markers used to identify them). Cocktail targets used to induce pluripotency overlap with targets deregulating cancers. Describes an in vitro GBM stem cell model using neurospheres (matched pairs of stem cells, differentiated cells from the same patient, tested by their efficieny to induce tumors after injection into mice). Using H3K27ac as comprehensive readout, marking active regulatory elements in TPC (tumor propagating cells). Loci known to differentiate cells become active in differentiating cells.

Can use this readout to classify regulatory sites (active in stem cells, active in differentiating cells), check for genomic sequence enriched in regulators to predict transcription factors involved. Combine with RNA-Seq to overlap with differentially expressed TFs, came up with 19 candidates (validated by western, then functional test to see if TF expression confers phenotype). Introduce individual TFs in differentiated line, for some (POU3F2 for example) by themselves result in weak sphere formation. Complemented best factors with second TF, repeat with 3rd TF (time consuming, in many cases waited for a year for mice with injected cells to develop phenotype). None were developing tumors. Added helix-loop-helix factor which did the trick, cells induce tumours potently. [Might have missed what factors were used precisely, assume confidential still]. Tested epigenetic landscpae, TF mix re-actives the elements they predicted — not just phenotypic reversal but epigenetic reversal. Tried reverse approach knocking TFs down with shRNA, impairs self-renewal ability.

Moved back to human GBM to see how well the model transfers. Staining for TFs indicate subset of GBMs co-express all four TFs, mostly cells with cell surface markers for the stem cell phenotype. Used ChIP-Seq to map TF binding sites in genome, confirm preferential binding to stem cell program elements. Cross-talk between genetic mutations, epigenetic state that confer key properties in GBM.

Subsequently applied single-cell RNA-Seq to cells in primary glioblastoma, again because tumor tends to heterogeneous (at least four subtypes defined as well). Tried to distinguish infiltrating glia cells from cancer cells by single cell chromosome copy number variation as variant calling on single cells tricky. Averaged gene expression of 300 gene blocks based on location as markers, 420 cells with gain/loss events typical for GBM, 10 without (removed from study). Checked pattern of expression of surface receptors, turned out to be heterogeneous (PDGFRA, EGFR in particular subdivide in two classes). Re-arrangements of EGFR within the same tumour observed. Can arrange cells in a gradient by expression of stem cell signatures, but no discrete classes — a continuum. Core TF network found previously correlates with degree of stemness readout. Used in combination to classify into high/low stemness to identify other candidate regulators. [Tour de force to integrate different aspects of GBM biology, indeed.]

Interrogating regulatory networks to discover master regulators of tumor initiation, progression, and drug resistance

Talk from Andrea Califano (Columbia University) who starts on the genome (rather than beyond the genome). Number of statistical associations of genotypes to phenotype rapdily increasing, some fully penetrant (and often with no idea of their mechanism). Huge statistical problem to find useful links due to combinatorial problem (more genotype combinations than single cells on the planet), and without mechanistic clues unclear how the genotype causes phenotype. We lack the power to identify epistatic interactions.

Segue to model-based genomics. Treat the cell as a detector to understand how cells integrate different signals to build mechanistic understanding of genotypes/epigenotypes. Optimal to elucidate epistatic, synergistic interactions but the model needs to be accurate and reasonably comprehensive (see paper). Build models based on different tools to figure out transcriptional interactions, post-translational interactions, transcriptional interactions, identify master-regulators to get to a mechanism of action. Ran models for large number of different tumors with good success [I am too blind to read the tiny print on most of Andrea’s slides, unfortunately, so notes will be sparse here.] Process is to find some sort of ‘omics signature that distinguishes a particular phenotype transition (tumor/normal, resistant/non-resistant), then find targets of signature to unravel mechanism. Cancer progression driven by a relatively small number of master regulators, themselves dysregulated by a large number of upstream genetic and epigenetic changes. These can be unraveled, finding oncogene induced dependencies. Requires two states of the cell, find genes differentially expressed / methylated / anything, find most likely upstream regulators of singature, find master regulators that necessary and/or sufficient to cause the observed signature (and by extension the phenotypic change); usually a set or ranked list of candidate regulators. Can reverse / model the impact of changing these regulators to see if they replicate the signature.

Walks through a few examples next, e.g., reversal of glucocorticoid resistance in T-ALL or identify mesenchymal subtype in glioma. Walks through potential application for personalized medicine, particularly with regards to relapse/resistance. If master regulators implement bottlenecks then drugging these will reduce resistance as any upstream mutation (where they see most changes) will not be able to bypass the ‘drugged’ bottleneck. Clinical trial in breast cancer now ongoing to test that in practice targeting STAT3; modified algorithm for single patients based on single cell analysis to identify likely master regulators for disease progression. Targeting just some of the essential/central master regulators results in the collapse of the whole case/control signature. Find drugs to inverst master regulator signature (SYNGEN algorithm now in press). Applied to neuroendocrine tumours of the GI tract, ~1000 fresh frozen samples, generated regulatory model (ARCNe, RNA-Seq and Co) to identify likely master regulators with strong impat of 50 top hits (conserved in cell lines, xenografts). Validated with drugs targeted nodes/bottlenecks. Tested drugs for signature reversion by RNA-Seq, top hit again now in clinical trial.

Uncovering master regulators of oncogenic transformation by network analysis of the LINCS library of transcriptional signatures of cellular perturbations

Mario Medvedovic from the University of Cincinnati gives the first short selected talk of the session, introducing the LINCS Project and figuring out regulatory networks and functional networks from transcriptional profiles. LINCS generates perturbation signatures of given nodes resulting in a transcriptional signature / profile which can be used in reverse: given an expression profile what nodes are likely changed? A cube of different cell types x perturbations x assays: cancer cell lines, primary normal/diseased cells, differentiated iPSCs, etc.; drug/chemical perturbations, genetic changes, microenvironment changes, etc.; readouts are transcriptomic, proteomic, phenotypic assays, phosphoproteomic, etc.

LINCS L1000 Data (CMap III by the Golub lab) now consists of 77 cellular contexts, 20k reagents, 22k genetic perturbations. More than 1,3 million gene expression signatures generated. Use to integrate activity scores of regulatory networks using kernel methods to ‘summarize’ the relationships between nodes (genes) connected in PPI networks, determining the level of interaction between neighbouring nodes. Use results to predict mutated regulators in TCGA luminal A breast cancer data: define mutational signatures, use network to rank importance of mutated gene for the gene (works better than based on expression based networks alone). Transcriptionally regulated neighbour of a driver gene are informative.

Genomic, epigenomic and transcriptional analyses in a Tet-Myc driven mouse model of liver cancer­

Last talk given by Valerio Bianchi](http://www.iit.it/en/people/valerio-bianchi.html). Starts with an intro to the Myc gene and how they generated a liver-specific Myc overexpression system in mice under the control of Tet. Can induce tumor formation, reverse with Dox treatment. Number of controls (normal liver, different Myc switch-off times) while tracing RNA-Seq and ChIP-Seq data for Myc and various histone marks. Seems that Myc overexpression results in ‘invasion’ of already active promoters and enhancers. Myc-induced tumors show expression pattern similar to human hepatoblastomas, indicating the mouse model is a valid test environment.

[fiamh-blog]

BTG 2014 Session 2: Biology from cancer genomics

The second session is kicked off by Marc Gersten talking about Human genome analysis, or more specificially on large scale non-coding annotatio and its application ot cancer genomics. Highlights personal genomics as a gateway into biology rather than the other way around. People normally not interested in biology will have their genome sequenced, get mutation information and will want to understand what this means. Requires the ability to look at their genome into the context of everyone’s genome. Focus on non-coding information, includes functional genomics (ChIP-Seq, ncRNA, etc) and conservation information. Range from peaks to basic RNA information, promoters, enhances with functional information such as segmentation (i.e., different levels). 1st level linear annotation, 2nd level network annotation, combine to prioritize variants.

Starts with experimental signal, detect peaks at different window sizes (quick detour to multi-mapping correction used by MUSIC algorithm, apologizes for acronym overlap, in press in Genome Biology), enhancer annotation (linkage of changes of histone marks with expression measurements across a large number of samples/tissues), conservation analysis (e.g., depletion of common variants across region, less in conserved region, measured by an increase in rare variants). Large number of genomes allows subcategoriation, looking at loss-vs-gain of function motifs, tissue-specificity (e.g., stronger selection in ubiquitously expressed / utilized motifs). Found almost 700 categories, subset of non-coding elements with the same footprint as an exome, smaller subet with the footprint of coding regions (ultra-sensitive regions). Highly enriched in disease-causing regions.

Secondary level (network annotation) combines proximal and distal sites into a network (network of half a million edges, 25k strong ones); associate enhancers with target genes. [Note: Marc talks way too fast for me to capture all the discussed papers, list here though). As found in other studies more connected network hubs have less variation, i.e., greater constraints. Can contrast essential genes with LoF-tolerant genes and check for different position in network, can use positional information from the network for prioritization.

Put this together for a workflow for prioritizing cancer mutations (this is working towards describing their recently published FunSeq2). Majority of somatic mutations are non-coding, the usual where-is-waldo problem. Do they hit sensitive regions, to they break motifs, do they fit in network hubs, do they recurr across patients or cancers? Walks through workflow. Important to turn into a useful piece of software (available code and web) which is tricky given the vast amount of data context being used: needs to separate variant prioritization when users upload data from context generation. Another scoring system that comes up with a score for a single position in the genome (similar to CADD, fitCons, extends to include recurrence information, highlights additional features (cancer genes, repair genes, differentially expressed genes, etc.). Validated the workflow by comparing core score of recurrent vs non-recurrent scores, scores of germline pathogenic variats vs controls, and applied to invidual tumor genomes. Seems to work better than CADD (which has known problems for non-coding scores). [Love his acknowledgements referring to 1000 Genomes group to the 1000 authors group].

Marc’s talk is followed by Nazneen Rahman on Delivering large-scale clinical testing of cancer predisposition genes. [Another very fast speaker…] How to get genetics/genomics out to cancer care? Focus on germline variants / cancer predisposition genes with clinical utility. ~100 known variants across 40 cancers with different contributions. Negative test also has value for stratifying patients, overall high clinical utility (diagnosis, follow-up, tailored therapies, cancer prevention in relatives, impact on health care cost): strong arguments for testing more genes in more people.

Three components: sequencing is only one part, followed by analysis and interpretation / clinical reporting. Done here via TruSight Cancer (collab with Illumina) testing for 100 cancer predisposition genes. Need high throughput that are at least equivalent to existing tests (50x coverage, every mutation class to be detected, high precision/recall, <8 weeks turnaround); likely to change to exome/whole genome, but for now panel based. Optimized from 12 samples per TruSight to 48 samples, HiSeq 2500, 576 samples/week/instrument. Pipeline takes about 6 days, 60ng DNA required, low failure rate. Analysis needs to be fast, short hands on time to minimize cost. “No bioinformatician necessary” [this scares me when followed by a pipeline slide with lots of .sh suffixes. They use Stampede/Platypus for analysis, 96 core cluster]. Validate rare variants by Sanger/PCR. Interpretation also automated as much as possible. Variants of uncertain significance tricky. Biggest problem is no clear triage process. 95% of variant classification can probably be automated (decision trees). Inputs are case variant data, control variant data, in silico prediction, functional assay information. Gene level data include mechanisms of pathogenicity, variability index, structure, gene-phenotype relationship information. Key novelty are the output classes, every variant has to go into a clinical output class (even if as simple as relevant vs non-relevant).

Usually 4-5 point scale (which is more fine-grained than the clinical management scale of 2). All variants innocent until proven guilty. Iterative approach as more information becomes available. Example BCRA interpretations, 80 positive reports, 511 negative, only 4 require custom reporting. Pre-generated baseline interpretation by checking all possible mutations in target genes, augmented with live link to Google Scholar searches. 2 days to extract DNA, 6 days sequencing, half day analysis, 3 days for interpretation and confirmation, 1 day for the report, total of 2 weeks turnaround.

Can do routine testing but needs to get into the clinic. Challenge of medical genetics (testing people with disease) and predictive genetics (testing healthy invidivuals). Going for a ‘mixed model’ via medical testing in cancer patients, BRCA testing in ovarian cancer. Relatively common but not too common for pilot, history of poor triage. Genetics appointment only offered to patients with identified mutations; 119 women tested, 20 with mutations, 14 with impact on treatment. Every single patient offered the test accepted it, much more commonly accepted now. Extending to breat cancer patients, making workflows available to other sites. Requires integration of multiple different disciplines and a willingness for change.

The third talk is given by Nicholas Navin on Clonal evolution in breast cancer revealed by single nucleus genome sequencing. From differences in individual patients to differences between individual cancer cells. Cites the TCGA 2012 Nature paper looking at inter-patient heterogeneity in breast cancer and how they typically only share one or two non-synonymous mutations, but have 30-70 different ones. Now check for intra-tumor heterogeneity (which is bad news for clinical diagnostics as the tissue samples might not be representitive), but is great for tracing mutations throughout the history of the tumor to distinguish early/late mutations in tumor progession. [Lovely slide from the Navin paper showing different models of tumor evolution.] Diversity difficult to measure in bulk tissue as the signal is averaged, de-convolution can be tricky — hence applied single cell genome sequencing. to breast cancer patients. Questions include range of clonal diversity, early/late mutations, gradual mutations or bursts, does mutation rate change?

Sufficient coverage for large CNVs (50kb-ish) via NUC-Seq [skipping details on the library prep, kit information, etc. as it’s all published, shoutout to Jay Shendure’s group]. Coverage at ~50-60x looks good, mean coverage breadth of exome (at least one read) at >90%, achieving overall around 88% SNVs and 54% indels in single cell exomes. Technical error rates at around 10% (allelic dropout rate), false positive rate about 1 error in million bases (but random, easy to filter). Applied to ER+ breast cancer patient, work with Ken Chen and Franziska Michor on the informatics and evolution models. Deep sequencing of 2N and 3N tumour cell populations; whole-genome sequencing of 4 single tumor cells (6N) in addition to standard whole genome sequencing of the sample at 50x. About 12 non-synonymous mutatons.

Single-cell CNV analysis shows highly similar profile across all tumour cells. ~50 single cell exome sequencing (and 12 normal cells without any mutations identified in the cancer cells at all), see paper. Number of mutations in present in almost every single cell (also detected in population sequencing), but plenty of mutations only present in 2-3 cells despite being very significant for tumor growth. Validated with an orthogonal approach (single molecule ultra deep targeted sequencing using molecular identifiers). Sequenced bulk tumor tissue to find variants present in as little as one in 10000 DNA molecules, sequenced to 100,000X, Validation rate for clonal frequencies at 95%, mutations present in 2+ single cells validated at 90%. Singletons validated at 20% (previously filtered out, but some of them still seem to be real).

Tested in triple-negative breast cancer patient without metastatic lesions, no treatment before sample collection. This one with ~360 somatic mutations in bulk sequencing, 15 in driver cancer genes. Single exome on 16 tumor, 16 control cells indicate diversity in the population, 160 additional mutations not found in bulk sequencing — these form three distinct subpopulations. Common variations again validated at 99%, subclonal at ~65%. Based on SIFT/Polyphen indicate ~30 mutations at below 1% with likely functional impact. CNVs show evolution occurred in punctuated bursts whereas point mutations were gradually acquired over time. Can trace mutations through evolution of tumor, links to punctuated equilibria (concept from evolutionary biology).

The main talks are followed by another round of short talks, led by Weng Khong Lim on Exome sequencing identifies highly recurrent MED12 somatic mutations in breast fibroadenoma. Most common benign tumor occurring in women, millions of new cases/year. Found during routine breast examinations, can sometimes be challenging to distinguish between malignant carcinomas. Aim to develop biomarker to distinguish benign/malignant cases. Do not bear common mutations seen in breast cancer including P53. Ran whole-exome seq over 8 matched tumor/normal pairs, only found MED12 exon2 alteration as recurrent. Targeted sequencing in 90 cases showed 60% of patients with this mutation. [Nice use of lollipop graphs to highlight mutation pattern]. Laser capture microdissection shows that mutation is exclusively present in stromal cells (n=4); mutation profile shown before in uterine leiomyoma, specific to benign, hormone-dependent tumors might indicate a role in dysregulation of estrogen signaling. Preliminary expression analysis on MED12 mutants have estrogene-response genes with differential expression patterns.

Had to run at this point and ended up missing both the second short talk from Sorana Morrissy (on Drastic genomic divergence of recurrent medulloblastoma invalidates targeted therapies discovered at diagnosis) and the final talk of the session from Frederick Alt on Convergent sense/antisense transcription at intragenic super-enhancers targets AID-initiated genomic instability which is unfortunate since we collaborate with members of his lab. Hoping that the Twitter coverage makes up for it.

[fiamh-blog]

BTG 2014 Session 1: Cancer genomics and epigenomics

The 2014 Beyond the Genome meeting has a simple tweet policy: talks are assumed to be sharable unless the speaker says otherwise, hashtag of the conference is #btgcg14. As always, all errors are mine.

First talk by Gad Getz on Cancer genomics and evolution and completing the catalog of cancer genes who also multitasks as session chair. Quick summary of cancer mutations, driver events (which increase the fitness when it occurs), clonality, passenger events and Co. Cancer genes defined as genes that harbor driver events. Recap of the MutSig algorithm to assess mutation frequencies, adjusting for gene length; results in a long tail distribution of cancer genes. Studies with higher power continue to find more genes at lower frequencies, vast majority at <20% in patients.

Look at larger and larger studies and the gene list increases, including hundred of olfactory receptors which indicates false positives. Problem is algorithm assumes constant background mutation background (all patients, all regions of the genome with same rate). Mutation rate across cancers varies up to three orders of magnitude; mutation rate also varies with expression rate due to transcription-coupled repair . Most important factor is location in the genome; rate correlates with DNA replication time. The ‘fishy’ genes mentioned above tend to have low expresison rate and late replication time. New version MutSivCV takes these factors into account.

Use for pan-cancer study testing ~4700 tumor/normal pairs from 21 different tumours checking for genes mutated in more than 2% of patients. Combine evidence to ind cancer genes, e.g., for clustered mutations in conserved regions of a protein; yields 224 cancer genes. Combined all tumor types together identifies 114 significant genes (which may not have reached significance in the individual tumor tests) — combined list of 254 cancer genes. Study found ‘all’ previously known cancer genes, but also identified 33 novel genes at an FDR of 10%.

Tried to assess completeness of the gene list via downsampling; indicates with more data and more tumour types the list increases steadily. Stratifying by frequency it seems the >20% frequency genes are mostly discovered, but at frequency of less than 5% the list is still rising rapdily. Clearly more genes to find. Power calculations indicate (ballpark figure) 50 tumour types with 2000 samples each (ongoig effort to study 15,000 samples). This is just for point mutations in coding regions. More work to do for non-coding regions, amplifications, epigenetic changes, etc. Great introductory talk setting the scene for the rest of the meeting.

Had to skip the second talk from Jan Korbel on Origins and consequences of structural rearrangements due to the inevtibable conference calls but made it back for the third talk of the session by Nuria Lopez-Bigas on Therapeutic landscape of cancer drivers. Nuria gives a quick proof of concept using Vemurafenib for BRAF. ~7000 tumor samples from 28 tumors (ICGC and TCGA), what is the opportunity to identify therapeutic opportunities to target cancer drivers. Iterative approach to identify drivers, find drugs targeting drivers, assign targeted drugs to patients based on driver event (not sure how this then links back to the original cohort).

First step to find driver events identified via somatic mutations, CNVs, fusion gene information, links back to Gad’s talk and described approaches to find positive selection across tumours re-sequenced genomes (MutSigCV, MuSiC-SMG, also measure functional impact of mutations with OncodriveFM and check for clustered mutations. These feed into the Intogen mutations pipeline (pipeline available through the web interface or for local installs).

Idea is to use complementary signals to obtain a comprehensive list of driver genes. Ended up with ~460 candidates (larger number than in Gad’s talk, but also additional tumor types sampled), again with a long tail distribution of mutation frequencies. Other findings are also similar, for example pooled analysis identified additional drivers, but misses drivers specific to a single tumor type. Needs both analysis. Next step is to distinguish loss of function from activation, key to drug development (activators easier to dampen, loss of function requires more thought). Estimate ~90% precision, 83 unclassified, 207 LoF, 170 Act. Further subset to focus on drivers dominating the clonal landscape, finds 73 major drivers. Includes CNA, fusion gene drivers to come up with list of ~480 driver events. Checking back on the original data it seems that ~90% of tumours have at least one of the recognized driver events, 68% have multiple events (median of 3). Differs a lot between different tumour types: most bladder samples have more than 1 (median 7).

Second step to identify and classify drugs targeting drivers (small molecule libraries, text search, etc). Only 96 targeted, most drugs through directed targeting. Classified remaining non-targeted drivers as druggable or not. Generated database with drug target interaction, dependency on specific mutation, recommended tumor types, repurposing suggestions to end up with ‘in silico drug prescriptions’ for 4068 patients as the third step. Estimate that 65% of patients could benefit from drugs already in clinical trials (again with the expected differences across various tumor types). Questions revolve around confidence of driver events, finding additional drivers, what to correct for (ignoring SNPs in CNV regions, cellularity adjustment, etc.).

Fourth full talk is from Michael Taylor on Epigenome alterations drive lethal posterior fossa ependymomas of infancy. Third most common childhood brain tumor, incurable in ~50 of the cases, Occur in almost any given part of the nervous system, treatment by surgery and radiotherapy, no chemo works. Morphological not distinguishable between site, but the transcriptomes differ quite a bit between sites. Extend of resection best indicator for good outcome, but with very negative impact on kids due to potential nerve/brain damage. Tumour transcriptome mimics cells of origin. Two distinct two groups (PF-A and PF-B) based on microarray profiles, replicated in Toronto and Heidelberg. PF-B much better survival rate than -A group, replace the surgery resection prognosis, PF-A vs -B tends to dominate prognosis estimates. Deep exome-seq on PF-A samples along with limited WGS. Zero recurrent SNVs in the samples they studied, lowest mutation rate among all tumours studied (PF-B have recurrent copy number events, in contrast). Cites Holmes on what to do when you eliminated impossible, so started checking for epigenetic changes. Found high rate of promoter CpG methylation along with gene silencing.

Checked methylated/silenced genes in two cohorts followed by pathway analysis. No convergence in PF-B, for PF-A PRC2-related genes, normally used by undifferentiated ES cells to prevent differentiation seem to be enriched. Can be reversed with demethylation agents. Identified good overlap of tri-methylation between PF-As and ES cells, no overlap with PF-B. Seems to be re-capitulatng embryonic program. No recurrent CNVs, indels, SVs, somatic mutations found so far. Lots of potential drugs that could replace highly damaging radiotherapy (e.g., 5-Aza-2’-Deoxycytidine), other drugs selective for EZH2 look promising as well (GSK EZH2 inhibitor). Moving through phase 1 clinical trial.

In the meantime revisiting prognosis estimates. Created PF cohort (with PF-A patients younger than PF-B, PF-B patients doing better than PF-A as expected). Look just at PF-A based on genomic tests it turns out that amount of resection does not matter. No difference between aggressive surgery that might put kids in wheelchairs vs less aggressive surgery. Likewise the effect of tumour grade on prognosis goes away once accounted for the two different tumor types. Huge impact on patient care, radiation plans and subsequently child development. Thinks these are not cancers, but ES-related neuron development disorder (balance of too many vs not enough neurons disturbed).

The session is conclucted by a two selected short talks, the first one from Ingegerd Elvers on Tumor/normal exome sequencing in dogs identifies known and novel lymphoma genes. Breeding resulted in landscape useful for studying somatic mutations in cancer. Live in the same environment as we do, but are more homogenic within breeds (and different across). Treatments dogs receive are similar to human treatment options, wel diagnosed. Picked golden retrievers (13% lymphoma over lifetime), boxers (20%), cocker spaniel (7%) to study disease across 3 unrelated breeds. MuTect/MuSIC to identify variants and significantly mutated genes, 10-50 events founds depending on breed and T/B-cell lymphona source. Little overlap between B-, T-cell lymphoma, usually specific for subtype. In addition, b-cell lymphomas between breeds are similar (share half of the significantly mutated genes), T-cell are not, indicating a genetic background role. A subset of the B-cell genes shared with human B-cell lymphoma or other cancers, reinforcing the concept of dog cancer studies as a useful model system. Expanding to additional breeds now and checking for germ-line predispositions.

The second short talk is given by Chai Bandlamudi on Pan-cancer gene fusion discovery across 7,470 primary tumor transcriptomes using MOJO. Quick intro to typical fusion events, frequently hallmarks of cancer with a minimal number of secondary hits. Interesting since they often can present novel therapeutic targets. Went through COSMIC and identified ~200 events in more than 3 samples based on FiSH and more recently sequencing. As always very different frequency across tumor types, enriched in blood and sarcoma, depleted in solid types. Not a case of event frequency but detection bias. Extended analysis to ~8000 tumor transcriptomes from TCGA, detected by discordant reads (using MOJO, Minimal Overlap Junction Optimizer, to be released soon). Ran it in a ‘high sensitivity mode’ with subsequent false positive filtering. Most fusions found between genes less than 1MB apart, fair amount still inter-chromosomal. Number of fusion events differs, highest in ovarian cancers, least in colorectal cancers. Quick comparison to TCGA indicates a number of noval fusion events they are now following up on.

[fiamh-blog]

Genome in a Bottle 2014: Figures of merit working group notes

Love the RTG approach, but general consensus is that we need an Open Source version of the variant comparison framework, and under a suitable license. Nils Homer to explore and make recommendations.

Will require some sort of default settings (given the target audience), at least for phase 1. Balance between getting a useful framework out sooner rather than later, and allowing lots of tweaking. Comparison between variant calls currently unsolved problem, lots of implementations out there (Complete Genomics, GeT-RM, bcbio.variation, RTG). Different truth sets, different regions of interest, different comparisons metrics (variant call vs getting the genotype right, boundaries of SVs, etc.).

Definition of the comparison metrics might have to wait for phase 2. Getting something out early to the community early makes sense, though it might mean we have to throw away the codebase at some point if the underlying framework do not meet the community needs (i.e., limited lifespan).

Long discussion about how to best handle truth sets and discrepancies. How to encourage users to submit evidence that the truth set is wrong in place? Wouldn’t they rather just re-run the analysis until the data converges better, even though that might be biased? Part of this will be submission of raw data as much as possible and iterative re-processing of all data to refine the truth set continuously.

Can we afford to ignore complex variants a little bit longer? Genome centers, tools developers want the more systematic comparisons along with APIs; the clinicians might be okay with just looking for rare variants for a bit. We also need community feedback on how to actually evaluate some of the more complex variants and structural changes.

Command line version obviously required (offline) for groups that cannot upload data, ability to define own truth set (not as a standard, but as a gVCF/BED file combination to re-run). Do we need some sort of meta-analysis (variance between VCFs)? Not required for the first round.

What variant classes are supported? Different layers, region by region, genotype by genotype, summarized top level. Not even getting started on phasing. Software needs to include all variants present and defined in the GiaB truth set; big CNVs/SVs off the table right now.

[fiamh-blog]

Genome in a Bottle 2014: Usage and release of reference materials

Several versions of RM released throughout the pilot have been used already:

Part of the FDA approval for Illumina sequencer

Variant caller and workflow assessment at HSPH

Tracing variant call changes over the development of a variant caller (FreeBayes, work with Marth lab)

Personalis categorizing exome regions, titrate coverage to compare different exome capturing methods, target sizes, etc; error rates outside of high-confidence regions can go up by an order of magnitude (poster (PDF))

Clinical pipeline validation at Mt Sinai (with internal validation), tuning of workflow to find better cutoffs (VQSR and Co)

Qiagen used material to identify PCR artifcacts, compare PCR enriched amplicon sequencing to whole exome sequencing, tuned workflows

Claritas Genomics to chase false negatives in assay development (Nils Homer)

Call for pedigree calls as another very valuable resource; better annotation for the non-high confidence calls to be able to have context on why they failed to make the high confident call set. Algorithm callibration (for example for tissue frequency) another use case (NA12878 as a spike in, ContEST and Co as modified downstream tools).

Releasing RM and future plans

Quick update towards establishing call sets and DNA as NIST (Certified) Reference Materials. Needs to go through various assessments (homogenous, stability, etc.). Difficult to do — needs lots of sequencing samples to reach statistical significance.

Confidence that the sample tested in a year is the same as the one tested this year. Protect against changes to the ‘physical form’ of the DNA (size distribution stable even under accelerated aging tests, see notes from day 1).

Might have to sequence batches side by side from Coriell each time new samples are released, trace changes over time. Again, need lots of sequencing to detect minor differences with confidence. Could be a repository for non-NIST lots; capture data from labs that run NA12878 routinely anyway. Automated process to upload BAM, run standard workflow, capture variant changes over time. [Could even be used to detect workflow specific signatures…]. Interest / support for this from DNAnexus.

[fiamh-blog]

Genome in a Bottle 2014: Real Time Genomics and Platinum Genomes Pedigree Analyses

(Presented by Francisco De La Vega)

Venn diagram hell required ground truth to sort out. All kinds of components — experimal validation, orthogonal data, references from different platforms all with their own biases and issues. Idea: use mendelian segregration as ground truth.

Given sufficient offspring this becomes feasible. Illumina/CG sequenced complete CEPH/Utah Pedigree 1463 and made the data available. 2x100bp, 50x, three generations. Trace variant segegration across 11 offspring, can be done for SNPs, InDels, more complex variants. How to distinguish between false positives and ‘real’ mismatches?

Use phasing approach to identify crossovers phase contiguity extension, connect haplotype islands, check calls vs haplotype framework. Requires filtering of problematic regions, algorithm minimizes number of required crossovers to explain haplotype. Found ~1-2 events per chromosome, total of ~650 crossovers. Can then test for phase consistency of variants: discordant variants that meet the phase restrictions are likely to be true (others likely false positives).

Can filter based on the likelihod that a set of genotypes is phase consistent by chance alone. See good improvement of using trio pedigree over raw NA12878 truth set (in ROC curves). In GiaB “confident” regions good agreement, improvements in the remaining regions. See GiaB note for pointers to software and Co.

Requires collecting data for whole family. Can merge multiple platforms for a phase-consistent truth set. Still needs to deal with CNVs, SVs, systematic errors.

[fiamh-blog]

Genome in a Bottle 2014: Working group reports

WG1 (Reference material): NCI-based fosmid data for somatic variant calls; additional cancer-specific mutations from Affymetrix. How to make use of plasmid samples with undefined quality characteristics? How to vet these, compare different types of material for suitability? What other methods / sets are needed? Deep sequencing of a cell line? Cell lines will change over time. ACGT collection might fit the required needs (pending the usual consent issues).

Come up with a study design and conduct experiments required to answer these questions (volunteers welcome).

WG2 (Characterization): Degraded DNA material, synthetic mutations, maybe mixed in with trio data, test sequencing depth required to detect synthetic variants. Start making use of long reads for phasing purposes in the trio (PacBio, Moleculo).

WG3 (Bioinformatics): From data to integration. Reviewed FTP structure, needs layer of representation for pedigrees etc., tools folder starting to link to GitHub and Co for better re-usability, can also include binary tools that otherwise could not be re-distributed. Accession approach (submitter, format, version information) for BAMs, VCFs and Co required, including everything needed for subsequent submission to SRA/ENA.

Consortium data model in flux, Global Alliance for Genetics and Health might be a good partner to gain recgonition. Would also provide minimal metadata framework. Will require format conversion at least for a while so data remains accessible.

New genome build a major discussion point. Value hg20 over hg19 at this point not entirely clear; but if consortium wants to drive / lead developments switch is required. Haplotype-aware aligners not widely used just yet; right now at least need minimal solution (i.e., ignoring the alternates). Extra layer handle alternates to decide which ones fit the data best. Final layer would be de novo assembly; back to the question what the truth set is. Pipeline can be imperfect but the answer set is perfectly concordant with reference set.

What are the frameworks for pedigree-based analysis, how to keep them reproducible? How to handle problematic regions — compile internal region lists, consolidate with NIST regions? How to use low coverage survey sequencing projects?

WG4 (Visualization / Figures of Merrit): How to allow users to evaluate their own data in the context of NIST, other reference data? Target audience for now prioritized to regulators (FDA), clinical labs, and to some extend industry for R&D in platform development. Requires web interface for labs without bioinformatics support; uploaded files still need some sort of standardization (API for bioinformaticians who want to go beyond the UI).

Requires high-level report (performance metrics), with the ability to drill down into the whys of discordance. Data, scores, matrices need to be made available for certification purposes. Punt file formats to WG3 (most likely not VCF). Key point is that context is needed: what evidence for what variant, how were variants filtered.

Reference set not should be changeable, but mechanism to provide feedback along with evidence needed. [See previous blog entry for more details on this session.]

[fiamh-blog]

Genome in a Bottle 2014 workshop in Stanford: Visualization group notes

Previously led by Justin Johnson (through EdgeBio and the XPRIZE validation framework). Now led by Deanna Church (Personalis.. still feels weird writing that) and Melissa Landrum (NCBI); converging with Lisa Kalman’s GeT-RM project.

Goal: user-friendly interfaces to reference genome data. Idenrtify target audience, define user interface requirements (truth sets to use, input/output, visualization, integration).

Existing: GCAT, GeT-RM with consensus SNP tracks. Started to collect metadata, gather tracks (region categorization), coordinate different efforts.

Quick intro to (ten year old) GeT-RM program from Lisa. Interaction with clinical labs, require quality control materials. Need predates NGS: develop, validate, maintain — not just for individual genes but now ~20k genes. Coordinate with NIST, Coriell, etc. Also selected 1000G/HapMap samples, collected variant call data (VCF/BAM where available) from ~30 clinical labs (gene/exome/WGS); metadata hard to capture. [Still surprised there was no pushback to get participating labs to send unified VCF rather than tons of different formats]. Visualize results in GeT-RM browser to make results available and accessible to participants. Plan is to add more tools to the browser, e.g. ,ability to chose different truth sets, ability to upload one call sets, reporting functionality. Target audiene: clinical labs (with limited bioinformatics expertise). Maybe an API [fingers crossed].

NIST might set up a competition for visualization approaches; how to abstract from browser tracks to higher level feedback.

Questions / notes from the discussion:

difficult to understand where discrepancies between data sets are coming from? Filtered by QC, filtered because of validation, threw it out due to non-pathogenecity (which are not reported at all, a problem for validation studies). Might need this level of granularity.

Also need a stronger focus on the difficult variants. 40bp indels, tricky CNVs and Co; the non-standard variants that have a high likelihood of being pathogenic.

Specifying validation tests almost an impossible task (technical, political reasons). Tool that follows some of the better workflows, is reproducible, allow it to be tweaked. I.e., rather than specification paper find ways to make the environment useful to the scientific community.

Should we use existing resource to slice/dice subsets? What is platform specific, what do you need pick up under any condition, etc. Goes back to #1; if I find something present in my VCF but not in the reference set where else was it found? I.e., platform or workflow specific variant call or something entirely different?

Overall problem of having to focus on some aspects to move working group forward. Focus on regulators, clinical decision makers?

Invitae (Steve Lincoln's) example of clinical validation workflow. For each sample previous data, run through own lab and workflow, obtain HGVS string, list concordance, discordant pairs, add reproducibility data (agreement with own data run on a different day). Calculate sensitivity & specificity (not ROC curves, not FDR). Also, Analytic Range: given the increasing size of samples and sites not everything can be validated. State ‘in regions like this we did find’ rather than global statements (genomic regions, class of variants, etc.). One table for each of the subsets (concordant, discordant, reproducibility arm). No good standard exists for that. [Back from the days of single gene comparisons.]. Discussion on what to use as range for GiaB. RefSeq CDS, other simple sets?

Could be useful to have an intersect in frameworks such as GeT-RM: pick truth set, pick analytical range (rather than merge the two into individual truth sets).

Focus on clinical end user, what is needed for accreditation (summarized results, but also the full data matrix). Other users might be interested in ways to slice and dice the variant information.

Need to have the ability to capture community information. When users go to the trouble of re-sequencing reference material to solve difficult regions that information should be accessible… somehow. [Wiki? Browser tracks? VCFs?]

[Footnote: Invitae to release complete, open-source HGVS parser; there’s also one from Counsyl..)

[fiamh-blog]

Genome in a Bottle 2014 workshop in Stanford: Day 1 morning session

Couple of notes taken on the fly; all errors mine.

Intro from Marc Salit

NIST initiative in the Stanford area (ABMS, Advances in Biological/Medical Measurement Science Program). Hub of genomics innovation despite Boston claims (hah!). Life, Agilent initial partners. Short summary of GiaB history, from idea to kickoff meeting to recent developments.

This meeting about initial reference materials and suite of tools to evaluate. “Yellow light” tools to indicate all looks well, carry on with caution. “Red light” the priceless signal that is needed. Technology and platform comparisons important second priority. Trying to stay away from regulatory involvement.

NA12878 still the initial plan, but consenting limited. Extension to PGP trios for a more compete set. Eight trios, focus on children; varying biogeographic ancestry. Open consent, state of the art. Details on Marc’s slides.

Usual working groups, with additional ‘meta’ topics. Includes move beyond the easy regions of the genome, incorporating new technologies, and selecting future genomes. Larger families for more power in confident variant calling, phasing? What approach to take for somatic variant calls? How to version reference materials?

Justin Zook providing more context

Whole genome RMs vs current validation methods: * Sanger confirmation, High depth NGS confirmation, Arrays, Mendelian inheritance, simulated data, Ti/Tv: all useful, complementary. Goal: approximately 0 false positive/negative calls in confident regions, don’t just take the intersection, avoid biases towards platforms or workflows.

Integration approach

Integration currently includes 14 datasets from 5 different platforms, keeps growing. Complex workflow to derive integrated call set, check characteristics of data sets at discordant sites to come up with reference agreement along with confidence score (excluding difficult regions with known duplications, long repeats, structural variation areas). Paper now accepted in Nature Biotechnology.

Work with GCAT to compare different workflows, algorithms. Can already spot platform systematic errors, regions different aligners are struggling, negative variants dominated by complex variant representation problems (225k highly confident variants are within 10bp of another variant). FP/FN enriched for complex variants; RealTimeGenomics’ vcfeval tries to deconvolute these.

Structrural variation

Moving towards structural variation with a similar approach. Combine different methods (depth of coverage, paired end mapping, assembly based, etc.). Strong disagreement (less than 10% in the intersection of methods). Gathering metrics (soft clipping, mapping quality, # SNP calls, coverage) summarized by new software package, will be used to asses performance

The todos

Todo: need to tackle ‘hard’ regions of the genome. Move towards pedigree calls as a different way to get high confidence calls. How do we integrate these calls with existing variant calls, and should we looking for additional large, consented families to assess?

Pilot reference material (NA12878) being tested at various sites (at NIST, Illumina, CG, Garvan Institute, NCI, INOVA); up to 300x, arrays, exome/WGS, etc. Illumina/SOLiD/IonTorrent include. Assessment now in progress, including additional QC — are there differences between first, last vial of a batch (vial, library, day, flow cell, lane, sampling, etc.)? How stable is the same material (freeze/thaw cycles, vortexting, etc.)? Check size distribution with PFGE on top of that. Size degradation likely to be the biggest confounding factor.

Availability

On NIST FTP site, Amazon S3. Official release of reference material in the next couple of months.

[fiamh-blog]

Exploring the Cancer Methylome

_[Peter Laird](http://pibbs.usc.edu/faculty/profile/?fid=241), University of Southern California, USA_ 70% of CpGs methylated. In a cancer cell a number of changes. Widespread (rather than global) hypomethylation and focal CGI hypermethylation in cancer, frequently in promoter regions. What is the relationship between those two discordant events? 125 colorectal adenocarcinomas genomes profiled for methylation, 29 adjacent normals. Clusters nicely, but are those associated with clinical features? Align with BRAF mutation status, for one (but mutations do not induce the methylation status). Another Distinct CpG-island methylator phenotype, CIMP+, micro-satellite unstable subtypes also overlap (if incompletely). Can be used to generate epigenetic subtypes of colorectal cancer, applied to TCGA with the same outcome. Synergy between cancer genetics an epigenetics. ES-Cell polycomb repressor complex targets are prone to abnormal DNA methylation in cancer (polycomb keeps master regulators of differentiation in poised state). Large number of promoters acquire methylation in cancers compared to matched normals, enrichment of polycomb targets among those cancer-associated methylated targets. Same genes also acquire increased methylation with age. Polycomb crosstalk likely leads to cumulative stochastic methylation. Loss of polycomb and replacement with more permanent silencing method -- i.e., methylation -- means target no longer able to differentiate, gets stuck in stem cell state without full differentiation capability. Great target to accumulate additional mutations over time. Potentially not an active process in cancer, but a hallmark of the event that lead to the cancer development. Not a competitive advantage to develop a cancer, but a passenger event indicating that the progenitor cell could no longer differentiate. Would explain DN methylation of ~50% of cancer-specific methylated genes, consistent with stem cell-like behaviour of cancer cells, explains observation of epigenetic field effects adjacent to tumors: a differentiation block. Focal hypermethylation and long-range hypomethylation: WG bisulfite sequencing of primary tumors and normal tissues. Shows sample CpG site in normal/cancer, striking difference of signal. Zooming out to ~20kb windows, nice heatmap/scatterplot of methylation signal show partially methylated domains. Shows erosion of methylation pattern in window from ES cells to differentiated colon to cancer. Hypomethylation _not_ uniform distributed, clear windows. Epigenetic unstable regions close to the nucleous (late replicating regions, lamin attachment regions). Comparison across cancer types: holds up across multiple cancers studied, with individual differences that should be interesting to tease apart. Comparison of 2200 TCGA cancers 409 normals reveal cancer-specific profiles in unsupervised clustering and pairwise correlation analysis of all cancer types using all sites.

[fiamh-blog]

Analysis of somatic retrotransposition in human cancers

_[Peter Park](http://compbio.med.harvard.edu/), Harvard Medical School, USA_ 45% of human genome derived from transposable elements (TEs), able to replicate (copy/paste) across the genome via an RNA intermediate. Previous studies of TEs mostly in germline, >7000 insertions in 185 samples from the 1000G set. Implicated in single gene diseases with ~100 insertions reported so far (L1s, ALUs, others), e.g., neurofibromatosis type 1 contains 18 retrotransposon insertions. Some events found in cancer, for example L1 in MYC -- but discovered in low-throughput studies. Studied 43 cancer/matched normal genomes (150 billion reads) from TCGA (GBM, ovarian, colorectal, prostate, myeloma). Thousands of novel germline insertions, 194 _high confidence somatic insertions_ identified (majority L1s). Detecting events challenging. Numerous, often identical TE instances. Find cluster of read pairs where one end maps uniquely, other end maps to TE consensus sequence. Use of clipped reads (partially aligned reads) key. Added custom assembly of repeat elements to the genome to be able to find repeat families at once. Tool called [Tea](http://compbio.med.harvard.edu/Tea/) (transposable element analyzer); insertions validated by Sanger. All somatic L1/Alu insertions in cancers of epithelial cell origin, _none_ in blood or brain cancers (_Ouch: postdoc started with those tissues by chance._). 64 out of 194 insertions located in genes, including tumor suppressors (UTRs and introns). Somatic insertions tend to occur in genes commonly mutated in cancer, disrupt expression levels of target genes, biased towards regions of cancer-specific DNA hypomethylation (all statistically significant). Unclear where and how often these happen Quick shout out to Galaxy and Peter's Refinery system, a data repository connected to the Galaxy backend for data analysis.

[fiamh-blog]

MutaScope: a high sensitivity variant caller for amplicon sequencing

_[Shawn Yost](http://bioinformatics.ucsd.edu/students/11/), University of California San Diego, USA_ Targeted therapy in cancer hinges on having more actionable genes: only 47 genes actionable (approved or in clinical trial). Designed ~2000 amplicons (150kb, 1000X coverage) to cover these. PCR amplicon and MiSeq allow for quick turnaround (UDT-Seq, Ultra-Deep Targeted Sequencing). Amplicon has fixed directionality, start/stop position, differs from whole exome sequencing. All mutations will have the same position in a read, making it difficult to identify false positives by usual metrics. High depth of coverage also unusual, but needed due to sample heterogeneity. Applied to ~50 samples with wide range of percent of invasive tumour cells. MutaScope approach: BWA alignment, refine alignment, calculate experimental error rate using a germline sample, variant detection and classification. Read refinement: each read assigned to the amplicon from which it was sequenced, allows to use RG information to improve variant calls. Unclip soft-clipped bases to allow variant calls at the end of reads. Error rates: based on germ line sample to correct for error rates dependent on position of read, CG content of reference. Use two models of tumor heterogeneity (spiked in tumor samples at 1/5/20% as a control or two mixed germ line samples). Tested against Samtools, Varscan, GATK. MutaScope works better for this kind of sequencing. Applied successfully to clinical samples, studying prevalence of somatic and germline variants. Designed specifically for PCR amplicon, modular workflow from FASTQ to PCR. Additional information in VCF such as detection p-value, classification p-value and mutant allele read group bias.