1st year Graduate Presentations TT2010
Schedule of Talks: Lecture Room (1 South Parks Road)
1:30 - 1:45 Robert Gaunt
Title: Rates of Convergence of Generalised Hyperbolic Approximations via Stein's Method
Abstract: Stein's Method is a powerful technique which allows the bounding of errors in asymptotic approximations and hence obtain the rate of convergence in important settings. In this talk we look at how Stein's Method can be used to obtain approximation results involving members of the family of Generalised Hyperbolic distributions. These approximation results will have applications in computational biology and finance.
Questions and discussions
1:50 - 2:05 Cornelius Probst
Title: A survey on changepoint models
Abstract: A changepoint in a discrete time series indicates a (typically abrupt) change in the underlying data generating process. Occurrences of such changepoints abound in areas as diverse as industrial quality control, finance, and seismic monitoring; and it is apparent that a fast and reliable detection mechanism is required.
Theoretical underpinning for such mechanisms has been given within a manifold of statistical models. We will present a survey on these proposals, and show how certain particle methods allow for realistic modeling and efficient analysis in a comprehensive class of model
Questions and discussion
2:10 - 2:25 Ewan Maddock
Title: "Simulating A Coalescent Dual Process In The Moran Model"
Questions and discussion
2:30 - 2:45 Tze Leung Choy
Title: Distance metric learning with L1 penalty
Abstract: Nearest neighbour classification has poor performance for high-dimensional datasets. The performance of nearest neighbour classification can often be improved if we learn a distance metric. We will outline several distance metric learning algorithms. Examples will be given to illustrate how learning a metric can improve classification performance and how we can perform dimensionality reduction with such algorithm. We will also discuss how L1 penalty helps to introduce sparseness and prevent overfitting the metric.
Questions and discussion
2:50-3:05 Eleni Frangou
Title: Allowing for gene interactions in Genome Wide Association Studies
Abstract: Genetic association studies play an important role in the identification of genes affecting various diseases. The common tool for identifying several disease associated polymorphisms is single SNP association studies. Interest is now focused on the detection of effects produced by gene-gene interactions or interactions of genes with environmental factors. Models with interactions have higher power and can identify causal variants that potentially could be missed if the study was based on single locus association.
During this presentation I will discuss an existing method for detecting gene interactions based on MCMC sampling.
The limitations of this method lead us to the development of a more powerful method that will be able to accommodate them. A brief overview of the new model will be given.
Questions and discussion
3:10 - 3:25 Therese Graversen
Title: A Gamma model for analysis of mixed traces of DNA
Abstract: In Cowell et al. (2007) a model for DNA mixtures containing DNA from more than one person was presented, assuming Gamma-distributed peak weights. We present the Bayesian Network described by Cowell et al., and in this connection a few prerequisites for Bayesian Networks are provided. The model contains a set of unknown parameters - the total amount of DNA and the proportion of DNA from the main contributor - and we explain how the model can be fitted to a dataset by using maximum likelihood estimation as well as MCMC. As an example of a relevant application, we demonstrate how the fitted model can be used to identify possible DNA profiles for the contributors.
Questions and discussions
3:30 - 4:00 Tea
4:00 - 4:15 Denise Xifara
Title: "Using Hidden Markov Models for Phasing Genotypes"
Abstract: "Reliable samples of haplotypes are particularly useful in the field of association studies and population genetics on the whole. However, experimental haplotype determination is costly and current statistical approaches, based mainly on MCMC, make a compromise between computational performance and phasing accuracy. We propose a novel approach towards phasing by exploiting the embedded coancestry within a sample of genomic data. We employ the imperfect mosaic idea and set up a Hidden Markov Model in order to choose appropriate pairs of individuals from our sample to represent the 'ancestors' of a particular genotype. We then phase using the Viterbi algorithm and appropriate choices of suboptimal paths. For a recombination rate of 10 and a sample of 50 individuals and 50 SNPs we achieve an accuracy of ~97%. ".
Questions and discussion
4:20 - 4:35 Faisal Khan
Title: Identifying Mitotic MAP proteins in Drosophila: a network biology-directed approach
Abstract: Protein-protein interaction networks are becoming increasing popular guides to help biologists identify and target uncharacterised proteins that might possess a putative function in different biological processes, like mitosis (cell division). In this inter-disciplinary project, we aim to produce an extended network of mitotic microtubule-associated proteins, or MAPs, that would help us characterize previously unknown proteins in this process. The network is based on an experimental dataset of MAP proteins in Drosophila and is extended by adding homologues and interologs from similar datasets in other organisms. This extended network will be 'layered' with different biological data that, we hope, will help us in selecting the best candidates for subsequent functional validation and detailed characterization using experimental methods in cell and molecular biology and biochemistry.
Questions and discussion
4:40 - 4:55 Valentina Iotchkova
Title: "A Method for Joint Analysis of Quantitative Trait Data in Genetic Association Studies"
Abstract: Increasingly people have started collecting more and more phenotypic data. But so far most Genome-Wide Association Studies have used a single phenotype. However, if a single gene influences multiple phenotypic traits then we might be able to infer more information by performing joint analysis. So here we propose a new Bayesian Model Averaging Method for joint analysis of multiple continuous phenotypes.
Questions and discussion
5:00 - 5:15 Yuan Yuan Liu
Title: 'Bayesian Classification and Regression Tree'.
Abstract: 'The classical classification and regression tree (CART) models use deterministic methods and greedy searches to split thus construct a tree. In contrast, another possible way to find CART models is via Bayesian approach, which consists of prior specification and stochastic search. The basic idea is to have the prior induce a posterior distribution that will guide the stochastic search toward more promising CART models. As the search proceeds, such models can then be selected with a variety of criteria."
Questions and discussion
5:20 - 5:35 Hongyu Jiang
Title: Stochastic Models in Mathematical Genetics".
Abstract : The presentation is concerned with the classic Wright-Fisher model and Moran model in population genetics, and more recent developments - the diffusion approximation model and Kingman's coalescent model. The derivations of some quantities of genetic interest from these models will be shown. The way to accommodate mutation and selection will also be introduced.
Questions and discussion
5:40 - 5:55 David Reshef
Title: "Nonlinear variable selection in high-dimensional graphs"
Questions and discussion
Joshua Lospinoso and Paulina Preciado-Lopez have submitted their presentations in pdf format as they will not be presenting
Paulina Preciado:
Title: Proximity Matters: Exploring the Distance Dependency of Adolescent Friendships
Abstract: Sociologists and social network analysts concur that when it comes to friendships, proximity matters. Yet little is known about how and under which conditions the likelihood of friendship falls off with distance.
This paper studies, from a descriptive perspective, how the probabilities of friendship existence, creation and maintenance depend on the geographic distance at which individuals live using logistic generalised additive models on network data of a three-wave annual study of 336 adolescents iving in a small, geographically isolated Swedish town. The analyses are performed for several groups of individuals that are similar or dissimilar in characteristics relevant for homophily. It is found that the dependency of the likelihood of friendship on distance is a smooth and simple function of the logarithm of distance and that the effect is stronger when the individuals have fewer other settings in common in which they can meet and interact. The results of this study are intended to provide information about how to transform distance for use as a dyadic covariate in parametric actor-based models for network dynamics. Paper [pdf]
Josh Lospinoso
Title: Assessing and Accounting for Time Heterogeneity in Stochastic Actor Oriented Models
Abstract: Paper [pdf]
Printer Friendly Version 