#recomb

Joe Nadeau: There be dragons

A talk around the fringes of what is known. Current model of complex traits -- many genetic variants, small additive and independent effects. Take two inbred strains of mice and swap one chromosome. New strain is inbred and homozygous except for this one combination. Partition genome into independent, non-overlapping segments. Can make statements about individuals rather than across a section of the population. By controlling noise (or standardizing it) should be able to find weaker effects. Took seven years, identified plenty of effects, but missed part of the story in the initial publication.

Fractal genomics

Mix of genetics, environment/life style choices (lovely example of David statue before and after he came to the US). Obesity has genetical component, collectively only explain 10% of the heritability (whereas heritable factors are expected to contribute up to 50% of the intra-individual differences). Test gene-diet interactions in inbred mice (A/J vs B6), B6 obese only with a HFHS diet. Check in 18 obesity resistant 4 susceptible strains: many chromosomes confer resistance to obesity, but resolution limited to chromosome level. Zoom in on chr6 using congenic strains (staggered, overlapping snippets of the chromosome from another strain), Lots of segments with strong, reproducible effects, but effects are context dependent. Split down to smaller and smaller segments -- effect size begins to decrease from 50% for segments with an average of about 4 genes, but is still very large. Contradictory to the expectation, many genes with large effect sizes.

Pervasive epistasis

Obrq1 in 628 congenic strain, 2% of the AJ genome in a B6 background, high risk diet, still lean. None of the genes show high risk on their own, but depend on the genomic context. QTLs in an intercross differ from those identified in congenic analysis. Unit of treatment is an individual, but how you map observations from population studies to a single patient is unclear. For 40 of the 41 traits pervasive non-additive effects.

Ceilings and floors

No examples of transgression (no strain more extreme than the parent strains). Are the parents bistable strains, are there genetic biases? What is controlling the range of phenotypic variation? Nine strains independently reduce cholesterol, 9 different solutions. A/J has all 9 QTLs, but is no more extreme than any strain with just one.

Modifier genes

Variants in one gene that modulate the phenotype of variants in other genes. Almost no cases where mutations are independent of genetic background, almost all genes are modifier genes. Do they identify network targets for modulating phenotypic outcome? Switching problem around, focus on the genotype of the healthy model. A/J eats 40% more than B6, sleeps more, moves less.. but is lean. What protects it from disease?

Family genome sequencing

Project at ISB, change to whole genome sequencing of families. Error rate under control when sequencing multiple generations of the same family allows the estimation of spontaneous mutations, create complete recombination maps. 150 genomes done, another 650 this year. Find modifiers of Huntington's disease. Check for Apoe4 mutations of Alzheimer, why are 20% free of Alzheimer's Disease (again, switching around: what protects the healthy).

Mendel and Lamarck

Identified in various mouse models that the genotype of a current mouse result from genetic exposures in previous generations, i.e., an epigenetic mode of inheritance. B6 mice where daughters differ in genetics only in the Y chromosome of the fathers. Minimized social influence, shared environment for siblings. Males with alternative Y differ as expected. 23% of the traits differ for genetically identical daughters.

QTL for trans-generational effects indicate persistence of effects lasting up to at least F4 (when going through males, two generations through females reverses effect). Again, epigenetic effect depends on chromosome context. Map of the chromosomes are needed to map those interactions -- interacting genes in different generations. Either environment or genetics induce stress, induce epigenetic changes (heritable through the germlline, not just a response of the soma). Pass on a pre-adapted gene expression profile to the offspring.

Cell death avoidance, uncontrolled cell division, eventually resistance to treatment. Causes include infectious diseases. mutations (predisposing inherited vs acquired through mutagens, environment). Complete decoding of genome, transcriptome and epigenome required to understand cancer biology. (Disclaimer: might also have to include the tumour microenvironment)

Cancer genome evolution: gradual accumulation of changes until the internal control systems are overwhelmed. Think of cancer as a heterogenous group even within the same tumour (including minority variants already present which give rise to resistant subtypes). Decoding genomes helps unravel the heterogeneity and supports treatment choices, Gives an overview of Smith Genome Center projects. All tied to new developments in algorithms and methods.

Focus for this talk on recent mutation discovery in B cell lymphomas. Cancer in the lymphatic cells of the immune system, present as an enlargement of the lymph nodes, extra-nodal sites include skin, brain, bowel, bone. 55% of all blood cancers, 5% of all cancers, 43 types organized into 4 large groups (Hodgkin (4 subtypes), mature T and NK neoplasms (13), immunodeficiency associated, mature B cell neoplasms (14) -- focus here on follicular and diffuse large B-cell lymphoma).

Cell type of origin for B cell malignancies seem to be diverse (different cell types in the germinal center), fundamentally different regulatory programs of development. Non-hodgkin lymphoma with a 4% compound increase in diagnosis in North America, mortality also increases (unlike most other cancers). Subtypes with different gene mutation profile in terms of frequencies (follicular vs diffuse germ cell vs diffuse activated; also differ in response to treatment). 90% of follicular lymphoma (FL) with translocation between chr14/18 (BCL2 under control of an Ig promoter), not seen in ABC type at all.

Genome and exome data for a number of FLs and Diffuse Large B-cell Lymphoma (DLBCL), about 100 RNA-Seq libraries. Mutation discovery based on normal/tumour genome comparison to find somatic, tumour-specific changes yield 143 candidate genes (validated), 400 more in the pipeline. Intersect with RNA-seq, final set of 231 genes (change in the genome and at least two more hits in the RNA-seq data). No matched controls for RNA-seq, so those are filtered against the standard known mutation databases, but still a lot of 'private' (germ line) variation present, or include RNA-editing changes. Identify hot spots, mutations that hit the same amino acid but different parts of the codon, yields a small set of genes enriched in chromatin regulation. Focus on two particular identified genes with high mutation frequency. Check for genes with non-random distribution of changes (ie, having a selection signature), both for mutations (potentially with gain of function) and truncation (loss of function).

EZH2, methylates H3K27 to turn off gene expression as part of the PRC2 complex. Only present in FL, DLBCL, not in other subtypes. Pattern of mutation recur, but certain residues are never affected (which is unexpected unless the cancers require a certain function). Assay system showed no activity for mutated versions on naked histone tails, but does show activity on histones already methylated. Mutations always present in heterozygous forms, yielding complexes with WT and mutated EZH2. Abundance of H3K27 (tri-methylated) strongly increased in heterozygous cases. WT with high affinity for unmethylated histones, mutants with the opposite preference. Potentially complications in regards to therapy -- what happens when you knock down EZH2 in other cancers? Inactivating mutations also associated with myeloid disorders.

MEF2B: transcriptional co-activator/repressor by recruiting HAT/HDACs, altering histone acetylaton. Mutations clustered around N-terminal part of the gene, again restricted to germinal center subtypes (not in ABC subtype). About half the mutations affect only three residues.

MLL2: adds methyl groups to lysine 4 on histone 3 (H3K4). Different pattern to EZH2 suggesting loss of function of MLL2. Mutated in about 90% of the FL samples, potentially linked to the translocation event. 90% of mutations are truncating, often in both alleles (unclear what the advantages are to the tumour)

Tien-ho Lin: Learning Cellular Sorting Pathways Using Protein Interactions and Sequence Motifs

Subcellular localization determined by protein sorting pathways. Image-based technologies allow to follow transport processes. Question changes from where a protein goes to how the process works. Determined by a carrier recognizing a motif. Path through the cell leaves traces of carrier and motif co-occurrence. ER, Golgi, others each with different carrier and motif sets.

Use a HMM simulating a protein transport between unobserved (hidden) intermediate states, emit signal peptide features or protein interaction, try to find the most likely transportation path through the compartments. DAG structure enables more than one path to the target. Static observation of a dynamic process is a challenge. HMM built with a structural search process.

[Still not quite sure what the input / training data for the model is -- state co-occurrence in the proteins?]

Seems to work well on recovering structure compared to SVM in both simulated and yeast data set (predicting final destination). Ongoing research with data from the Human Protein Atlas, extend model to handle additional complexity of alternative splicing, different cell lines, etc.

Tobias Petri: Experiment Specific Expression Patterns

Claim: predictable genes are less interesting, want to score 'unexpectedness' of a gene expression. If we can predict a gene's expression X with a model based on A/B/C remove it; if it does the opposite investigate. Workflow: start with diverse data set of diseases (cancer, infection, ...), build for each gene a model of regulation with a prediction quality given this background set. Take a new experiment, predict and compare the quality scores, check for outliers.

Models trained with patterns of fold changes within an experiment (between conditions). Train SVM for pattern detection yields model M(X) that predicts gene expression given all other gene expression changes. Check for correlation of predicted and observed expression; background distribution allows to qualify how well a gene can be predicted in general. Prediction quality for a gene based on leave one out analysis (leave out one experiment, that is).

Yosef Prat: Recovering Key Biological Constituents through Sparse Representation of Gene Expression

Geometric interpretation of gene expression: angle between gene expression vectors (correlation, clustering, SVD with rotation, etc.). Leverage expression as vector of a large dimension. Expression as a linear combination (gene expression vector and vector of correlations). Find the smallest group of profiles that can reconstruct the gene expression (a sparse representation of the overall expression profile).

Adapted work from compressed sensing and sparse signal recovery. For each gene find a small set (minimizing support) of other genes that allow the reconstruction. Tested in yeast (6000 genes, 170 experiments), learn coefficients explaining a gene given all other genes. reconstruction is very robust. [Would be interesting to see a comparison to NMF/Metagenes]

Support sets belong to dense PPI (at least a subset), those have functional enrichment. Try to use SPARCLE algorithm as a function predictor using machine learning. Features include the geometry, correlations, support size, coefficient values, gene number in intersection of pairs, but no external features. Predict label for each pair of genes.

Test case using STRING, BioGrid, Interact, ... for 200k interactions. Compare predictions to networks, significant higher prediction accuracy compared to just using gene correlation. Functional annotation: classify whether gene pair shares GO (SLIM) terms or not, again good prediction rates. Also improves annotation rates in poorly annotated organisms.

Leonid Chindelevitch: Causal reasoning on biological networks: Interpreting transcriptional changes

Help with the analysis of large data sets to extract biological insights. Focus on causation, not association (but keeps track of direct and inverse causation). Lipitor decreases HMG-CoA which... Grahs built from knowledgebases (Ingenuity, Genstruct), extract causal relationships and build graph with 40k nodes, 400k edges. Main contribution from protein-mRNA and ppi relationships. Hypothesis generation using ternary expression daya )over/under-expressed or not affected). If Hypothesis is plausible it correctly explains most of the observed expression changes. 40k hypothesis that need to be ranked. Assume shortest path from upstream regulator to downstream target provides sign of regulation. Score using 3x3 contingency table (due to ternary state descriptors).

Tested with a perturbation data set (oncogenic pathway signatures in human cancers). Perturbed genes tend to be among the top predicted ranks. Second test set on cardiac hypertrophy, manual compilation of predictions into a biological network / modules.

Sinan Erten: Disease Gene Prioritization Based on Topological Similarity in Protein-Protein Interaction Networks

Functional predictions, co-expression all useful for prioritization of candidate genes in complex diseases. Set of known disease genes (seed set) with level of association to disease of interest. Candidate genes and PPI network (with weighted edged), outcome is a likelihood of association for each candidate gene with disease.

For each candidate interaction partners are mapped; interaction partners checked for association with disease, expand subnetwork and repeat. Quick overview of scoring methods (count, information flow / network propagation). Try to capture additional information using topological similarity [difficult to convey without the subnetwork figures]. Example showing driver genes of colorectal cancer, not directly interacting or in proximity, but location relative to growth-factor receptors seems relevant.

Proximity captured using random walks with restarts (rather than shortest distance), similarity measured as correlation of distances. Calculate average of vectors of disease-associated genes results in a vector for comparison with the candidate genes. Data sets are all OMIM diseases, HPRD (?); most diseases only related to small number of genes. Leave-one out cross-validation by removing one seed gene and treating it as a candidate gene. Method outperforms existing information flow methods [wonder whether there is a sound way of combining the network analysis methods into one prediction]. Better performance potentially due to decreased vulnerability to network degree effects.

Each method identifies somewhat different set of candidate (test) genes; their method seems to be particularly good at recovering poorly connected proteins which may indicate lack of functional annotation / information.

Dana Silverbush: Optimally Orienting Physical Networks

Inferring directionality in physical interaction networks, Build mixed graph from directed and undirected graph, try to infer source-target pair directions. [Had to skip the talk]

Ron Shamir: Understanding gene sequence variation in the context of transcription regulation in yeast

Linking polymorphims to gene expression, overview of yeast data sets (essentially eQTL analysis). Linkage profile between gene expression of each gene and each position in the genome, can result in very large intervals with multiple genes within them. Combine with information from regulatory proteins to 'clean up' the signal to more specifically identify causal genes. Defined ReL modules: groups of co-regulated and co-linked genes (assume common regulatory mechanism). Input data is yeast linkage matrix (6000 genes x 3000 markers), 300 expression profiles (reg signatures) from the literature for regulatory protein knockouts.

Compare signature of a given TF with the linkage of a given marker, is there a distinct distribution of marker-associated genes (elevated or reduced compared to reference). Yields a matrix of 300 TFs x 3000 markers. Find modules within the matrix that show 'high signal' (genetic markers in consecutive order), yields ReL modules through bi-clustering [details in the paper]

Explore linkage intervals in identified clusters to identify putative causal regulators for each module. 9/13 identified modules already established, seven validated experimentally. Provides details on two additional novel modules including respiratory module genes and their link to mitochondrial respiratory complexes. Overlap, but not identical, work in different parts of the cellular machinery. Identified key regulatory protein, Swi3, shared by both modules.