Identification of Hybrid Time-varying Parameter systems with Particle Filtering and Expectation MaximizationSubmitted by lsr on Fri, 07/12/2013 - 15:00.
Abstract:One limiting assumption of many mathematical models for dynamic systems is that the parameters of the
system do not change during the observation period, which however does
not necessary hold in many cases.
This is typical for biological and medical systems, where we observe a high intra-individual variability in the model parameters. Hybrid time-varying parameter framework is able to capture the changes of parameters that may represent the change of state of the individual, for example
in HIV infected patients, changes of conditions in regulatory metabolic networks or diauxic bacterial growth on mixed sugar medium.
Thus, in these scenarios, a subset or even all the parameters have to be treated as time-varying in
order to capture the dynamics of the system.
An offline (batch) algorithm that combines particle filtering and the expectation maximization is
introduced for the identification of such systems.
The efficiency of the proposed method is illustrated through simulated and real-world examples.
My program of research focuses on the integrated use of neurocomputational models and empirical techniques – most notably brain imaging – to investigate the neural bases of learning and cognition in healthy subjects and the disruption of these processes in psychiatric disorders. This talk will present three examples in which this integrated approach has been crucial to arrive at novel insights that would be out of reach for more classical approaches. In the first example, I will use a standard reinforcement-learning model to show that learning to avoid negative outcomes depends on an internally generated reward signal. The firing of dopamine neurons in the brain represents reward signals, so this model made new predictions concerning patterns of dopamine release during avoidance learning that we have confirmed experimentally. In the second example, we have used model-based functional magnetic resonance imaging (fMRI) – a technique that fits a computational model to behavior and fMRI data – to identify the neural substrates of habit learning in humans. We found that learning a habit depends on the engagement of a specific brain region; in fact, we were able to distinguish participants who learned the habit from those who did not based on whether or not they engaged this region. In the third and final example, using a neurocomputational model in which we manipulated neurotransmitter levels, we found that reduced levels of a specific neurotransmitter produce the behavioral deficits that characterize attention-deficit/hyperactivity disorder (a disorder characterized by inattention, hyperactivity, and impulsivity). The model made several new predictions concerning abnormal patterns of brain connectivity in patients with this disorder, which we have confirmed using fMRI. These examples illustrate the power of using neurocomputational models in tight integration with empirical techniques to advance our understanding of the neural bases of high-level cognitive processes in healthy subjects and to elucidate how these processes go awry in psychiatric disorders.
Amyotrophic Lateral Sclerosis is a devastating neurodegenerative
disease characterized by a usually fast progression of
muscular denervation, generally leading to death in a few
years from onset. In this context, any significant improvement
of the patient's life expectancy and quality is of major
relevance. Several studies have been made to address problems
such as ALS diagnosis, and more recently, prognosis.
However, these analysis have been mostly restricted to classical
statistical approaches used to find the most associated
features to a given outcome of interest. In this work we explore
an innovative approach to the analysis of clinical data characterized
by multivariate time series. We use a distance measure
between patients as a reflection of their relationship, to
build a patients network, which in turn can be studied from
a modularity point of view, in order to search for communities,
or groups of similar patients. The preliminary results
show that it is possible to extract relevant information from
such groups, each presenting a particular behavior for some
of the features (patient characteristics) under analysis.
About SING (http://sing.ei.uvigo.es/) @ University of Vigo
The Next Generation Computer Systems Group (SING, Sistemas Informáticos de
Nueva Generación) brings together a reduced number of researches with the
aim of developing intelligent models and deploying them in real
environments. The expertise of the members comes from different areas
related with previous research in developing symbolic, connexionistic and
hybrid AI systems, solving security problems, administration of networks,
e-commerce, VoIP, implementation of web applications and developing systems
working with documental data bases. The projects carried out by the SING
group always follow a practical point of view, but taking into consideration
the formal aspects needed in any research work. Indeed, most interesting
techniques employed in previous works cope with the utilization of
case-based reasoning, artificial neural networks, fuzzy logic, rough sets,
intelligent agents and multi-agent systems, etc.
In metabolic engineering or synthetic biology robust models with high predictive power are required. Constraints-based modelling methods such as metabolic flux analysis (MFA), flux balance analysis (FBA), elementary flux modes (EFMs) or extreme pathways (EP) have been widely used. The success of these methods is however conditioned by the many times insufficient mechanistic knowledge base. In this study, we built upon a previously developed hybrid constraints-based modelling method to develop E. coli models with improved predictive power. In particular, we apply a projection to latent pathways (PLP) method that merges together mechanistic and statistical constraints. It may be considered as a middle-out modelling approach that combines reliable knowledge and reverse engineering to extract unknown mechanisms from “omics” data sets. The method is applied to predict the central carbon fluxes of several E. coli strains (both wild-type and single gene KO mutants). We show that the central carbon fluxes of several single gene KO E. coli mutants could be predicted with high accuracy from the combined information of gene deletion and environmental conditions.
It has recently become possible to record the EEG simultaneously with fMRI, providing whole-brain maps of the hemodynamic (fMRI) correlates of electrophysiological (EEG) activity. The combination of the EEG high temporal resolution with the fMRI high spatial resolution offers a unique opportunity for studying the spatio-temporal dynamics of brain activity noninvasively. Here, I will focus on the application of EEG-fMRI to the study of spontaneous epileptic activity in patients undergoing pre-surgical evaluation. The EEG is used to identify electrical discharges associated with epileptic seizures, as well as interictal epileptiform spikes, and the fMRI correlates of such EEG features can then be used to localize the brain networks involved in the epileptic activity. I will first introduce the basic principles and main challenges of the EEG-fMRI technique. I will then present results regarding the investigation of the link between EEG and fMRI signals, as well as the functional brain connectivity underlying seizure propagation. Appropriate biophysically-inspired models are employed in order to extract the relevant information from the EEG-fMRI data. It is shown that, in this way, important insights may be gained into the spatio-temporal dynamics of epileptic activity.
I will give an introduction to hybrid modeling methods for bioprocess and biochemical networks modeling. Hybrid methods combine parameter-free modeling with statistical modeling tools. They enable to blend mechanistic knowledge and statistical relationships into models with improved performance and broader scope. Examples of such techniques are hybrid bioreactor modeling for optimisation and control, hybrid metabolic flux analysis for modeling formation of complex recombinant products, metabolic pathway analysis constrained by statistical relationships with “omic” data sets, and reverse envirome-guided metabolic reconstruction without the knowledge of kinetic parameters.
This tutorial aims at (i) giving an overview on theoretical fundaments of hybrid modeling for systems biology and (ii) to provide an introduction to the software tool HYBMOD, a MATLAB toolbox for systems biology hybrid modeling. The presented theoretical methods will be exemplified by examining simulation and experimental case studies.
Computational modelling of protein interactions (docking) is an important endeavour because protein complexes are difficult to determine by experimental methods alone. Nevertheless, computational prediction of protein interactions is no trivial task either and there is much to be done to improve the reliability of protein docking methods. Protein coevolution traces that are identifiable from the analysis of multiple sequence alignments can help predict protein interacting contacts and can be used to constrain the search space of constrained docking algorithms, such as BiGGER. Although several methods for detecting it have been developed, only a few were extended to the study of protein-protein interactions and not much is known about their ability in identifying contact points. We addressed this issue, developing an integrated system, Pycoevol, that is a set of open source tools for automating the identification of contact points from inter-protein coevolution.
Systems biology provides new approaches for in silico metabolic engineering and drug development through the application of analysis, simulation and optimization methods for metabolic models. In silico modeling of cellular metabolism is divided between genome-scale stoichiometric models and small-scale kinetic models. While the former are analyzed using optimal assumptions predicting intracellular microbial fluxes and growth rates, the later are used for dynamic behaviour simulations. However, there is currently a separation between these two modeling approaches. In this talk, I will discuss my work related to the challenges to build novel computational approaches for complex large metabolic networks of biological systems and to fill the gap between kinetic and genome-scale stoichiometric models.
Subcellular location is an important property of proteins, carefully regulated
by the cells. To determine subcellular location on a proteome-wide scale,
fluorescent image data is most commonly used and a classification system is
employed for analysis. These systems assign each protein to one of a small set
of predefined location classes (typically the major organelles).
This is a limited representation of the underlying biology as proteins are
often in multiple organelles. I will present techniques that go beyond the
case of single location assignment. These techniques were applied on a large
collection of images of fluorescently tagged mouse proteins, which included
several proteins for which no location assignment had been previously reported
in the literature.