A reminder about this joint meeting of the General Applications
Section of RSS and the British
and Irish Region of the International Biometric Society.
Registration in advance strongly preferred (form at http://
www.rss.org.uk/main.asp?
group=&page=1321&event=434&month=&year=&date=) so we can sort out
catering.
Thanks,
John
Introduction to networks from a statistical point of view
JEAN-JACQUES DAUDIN
This presentation will give an overview of relevant questions and
statistical methods for networks. Many groups are studying the
complex networks that arise in nature and society, but with quite
different points of view and types of publications. However the
intersection between them shows common elements which make up a new
domain of probability, statistics and computer science. Following
some basic notation on networks, the following questions and methods
for handling them will be skimmed over:
- Detection of modules or communities from network topology,
unsupervised classification of vertices;
- Prediction of the colour of a vertex using information about its
neighbours;
- Dynamic models for random graphs such as linear preferential
attachment;
- Static models for random graphs: Erdos-Renyi, exponential random
graphs, latent class models;
- Motifs in networks;
- Sampling questions;
- Future needs for models, precise and computationally efficient
estimates and high quality data.
Inference of Molecular Networks
MICHAEL STUMPF (joint work with TINA TONI & SOPHIE LEBRE)
We will present an overview of three problems related to different
molecular networks that will span the range from system-wide networks
all the way down to detailed dynamical descriptions of molecular
processes in living systems.
(i) We will show how incompleteness of network data affects our
ability to study the properties of protein interaction networks, and
how multi-model inference and model averaging can be employed to
predict properties of the complete network from incomplete and noisy
protein interaction network data.
(ii) We then continue with an overview of how graphical models can be
employed in order to elucidate temporal processes from e.g. gene
expression data.
(iii) Finally, we will discuss recent developments in approximate
Bayesian computation (ABC) applied to the study of dynamical systems
- such as metabolic or signaling processes - in molecular systems
biology. As we will show, ABC can be readily incorporated into a
Bayesian model selection framework.
Reconstructing regulatory networks by nonparametric methods
LORENZ WERNISH Joint work with IOSIFINA POURNARA
Most of the cellular processes are not accessible to direct
measurements.
For example, the level of protein activities can often only be
assessed indirectly by their effects on gene expression or the
modification of
other proteins. The exact form of the functional relationships
describing such interactions are unknown as well. Nevertheless, some
progress can be
made by combining factor analysis or dynamical models with nonparametric
regression methods, which don't impose any form on reconstructed
functions. Interestingly, nonlinear relationships seem to be better
suited
than linear ones for reconstructing causal relationships. On the
other hand, identifiability becomes are major problem and needs to be
balanced
against flexibility. I will report on our experience with simulated and
experimental data in exploring the limits of reconstructing hidden
protein
activities and interaction networks.
Uncovering latent structure in valued graphs: a variational approach
STEPHANE ROBIN
As more and more network structured datasets are available, the
statistical analysis of valued graphs has become more common, that is
where the links have values attached to represent characteristics such
as strength, duration, capacity or flow. Looking for a latent structure
is one of the many strategies adopted to better understand the behaviour
of a network. Several methods already exist for the binary case. We
present a model based strategy to uncover groups of nodes in valued
graphs. This framework can be used for a wide span of parametric random
graphs models. Variational tools allow us to achieve approximate
likelihood estimation of the parameters of these models. We provide a
simulation study showing that our estimation method performs well over a
wide range of situations.
Statistical Modelling of Networks in the Social Sciences
TOM SNIJDERS
Social networks can be defined as the patterns of ties between social
actors. They are usually represented by graphs and digraphs. The crucial
issue for statistical modelling of social network data is how to deal
with the stochastic dependencies between the variables indicating the
existence of different ties. Such dependencies could express, e.g.,
tendencies toward reciprocation in directed networks, and tendencies
toward transitivity. In this presentation attention will be given to
Exponential Random Graph Models, which can be used for modelling single
observations of graphs and digraphs, and to Stochastic Actor-oriented
and Tie-oriented Dynamic Graph Models, which can be used for
longitudinal observations. In the social sciences it is common to
collect longitudinal data in a panel design, which poses special
problems because it is plausible to assume that changes in the network
take place in continuous time, unobserved between observation moments.
The models under consideration do not allow explicit calculations, but
they can be implemented as computer simulation models and analysed using
MCMC methods. Frequentist estimation methods will be discussed, based on
stochastic approximation procedures. Some illustrative examples will be
presented and open problems mentioned.
Sexual networks and the evolution of sexually-transmitted
AZRA GHANI
Identification of areas of the sexual network in which infection
persists is important in the control of STIs (sexually-transmitted
infections). However the epidemiological data required for this task
are frequently difficult to obtain. In a study in London we tested the
feasibility of using molecular typing methods to identify clusters of
related isolates. To aid interpretation of the clusters observed I will
present details and results from a network model of gonorrhoea
transmission to investigate the patterns of strain structure expected
under a variety of network structures. The dynamic network model for
gonorrhoea transmission was adapted to incorporate sequence types (ST)
defined by their allelic profile at two loci. All ST were assumed to be
equally fit with new ST arising through recombination and mutation. The
talk will explore the extent of strain diversity that can be expected in
the absence of significant mutation, how strain structure might relate
to the underlying properties of the sexual network, to what extent we
can understand the underlying sexual network from the strain structure
observed in a population and under which scenarios we can expect a high
concordance in ST in sexual partnerships or short transmission chains.
Meeting Contact: Chris Glasbey (BIR-IBS) & John Whittaker (Gas-RSS)
Organising Group(s): General Applications Section/IBS
The charges for this event are:
£25 for RSS Fellows and IBS members
£17.50 for student RSS Fellows and IBS student members
£20 for CStats/GradStats
£35 for non-members
John Whittaker
Professor of Genetic Epidemiology and Statistics
Department of Epidemiology and Population Health
London School of Hygiene & Tropical Medicine
Keppel Street
London WC1E 7HT
[log in to unmask]
Phone: +44 (0)20 7927 2025
Fax : +44 (0)20 7580 6897
|