First
International BioInfo'2005 Workshop
Abstracts
Local and "Personalised" Modeling and Knowledge Discovery in
Bioinformatics
Prof. Nikola Kasabov
Director, Knowledge Engineering
and Discovery Research Institute, KEDRI Auckland University of Technology, (http://www.kedri.info/),
nkasabov@aut.ac.nz
Abstract - This presentation introduces first some challenging problems in
Bioinformatics (BI) and then applies methods of Computational Intelligence (CI)
to illustrate possible solutions to them. The main focus of the talk is on how
CI can facilitate discoveries from biological data and the extraction of
knowledge.
Methods of evolving
knowledge-based neural networks and hybrid neuro-evolutionary
systems, characterised by adaptive learning, rule extraction and evolutionary
optimization [1], are highlighted among the other traditional CI methods [2].
CI solutions to BI problems
such as: DNA sequence analysis, microarray gene
expression analysis and profiling, protein structure prediction, gene
regulatory network discovery, medical prognostic systems, modeling
gene-neuronal relationship [3] and others are presented and illustrated.
Fundamental issues in CI
such as: dimensionality reduction and feature extraction, model creation and
model validation, model adaptation, model optimization, knowledge extraction,
inductive versus transductive reasoning, global
versus local models, kernel methods and others are addressed and illustrated on
BI problems.
While inductive modelling is
used to develop a set of local models covering the whole problem space and then
to recall them for a new data vector, transductive modeling is concerned with the creation of a single,
"personalized" model for every new input vector based on some closest vectors
from the existing problem space. Issues of feature selection, dimensionality
reduction, neighbourhood selection, model optimization, model verification,
personalized profiling, and knowledge discovery are discussed and illustrated
on the case study problems with the use of a software environment NeuCom (http://www.theneucom.com/).
A comparative analysis of different CI methods applied to the same problems is
presented in an attempt to identify generic and specific applicability of the
CI methods.
Keywords: Computational Intelligence, Adaptive knowledge-based neural networks,
Bioinformatics, Neuroinformatics, Local modelling, Transductive reasoning, Personalised modelling.
References
[1] N.Kasabov,
Evolving Connectionist Systems: Methods and Applications in Bioinformatics,
Brain Study and Intelligent Machines, Springer Verlag,
2002 (http://www.springer.de/)
[2] N.Kasabov,
Foundations of neural networks, fuzzy systems and knowledge engineering, MIT
Press, 1996 (http://www.mitpress.edu/)
[3] N.Kasabov
and L.Benuskova, Computational Neurogenetics,
Journal of Computational and Theoretical Nanoscience,
vol.1, No.1, American Scientific Publishers, 2004 (http://www.aspbs.com/)
Viewing the Phenomenon of Heterosis as a Network of Interacting Parallel Aggregation
Processes
Elena Tsiporkova1 and Veselka
Boeva2
1Computational Biology Division, Dept. of Plant
Systems Biology
Flanders Institute for
Biotechnology,
2Department of Computer Systems, Technical
University of Plovdiv, veselka_boeva@hotmail.com
Abstract - This contribution develops a mathematical model allowing
interpretation and simulation of the phenomenon of heterosis as a network of interacting parallel aggregation
processes. Heterosis ('hybrid vigour') refers to an improved performance of F1
hybrids with respect to the parents. It has been observed that a cross between
quasi-homozygous parents can in some cases
lead to an offspring (F1) that is better in terms of yield, stress resistance,
speed of development, etc. as compared to the parents. Heterosis is of great
commercial importance since it enables the breeder to generate a product (F1
hybrid seed) with preserved values which in turn, allows the farmer to grow
uniform plants expressing these heterosis features. Besides a commercial
interest there is a more fundamental scientific interest associated with the
biological phenomenon of heterosis performance, as an excellent example of what
complex genetic interactions can lead to. Two main models have been considered
in attempts to explain heterosis [1]: the additive-dominance
model and the epistatic model. The
present work is focused on the former one.
We have initially pursued
expressing the overall heterosis potential in terms of the heterosis potentials
of each of the individual genes controlling the trait of interest. This has
allowed us to gain a better understanding of the biological mechanisms behind
the phenomenon of heterosis. According to the additive-dominance model the net
heterosis potential can be expressed as a weighted mean of the heterosis
potentials of the individual genes weighted with their relative additive
effects respectively. Whenever the alleles are dispersed between the parents
this weighted mean is further rescaled according to their
association-dispersion coefficient.
Next, we have developed a
mathematical formalism that allows to interpret the sub-processes building up
the additive-dominance heterosis as interacting parallel aggregations. The
individual genes controlling the trait of interest are viewed as interacting
agents involved in the process of achieving a trade-off between their
individual contributions to the overall heterosis potential [2]. Each agent is
initially assigned a vector of interacting coefficients, representing the relative
degrees of influence this agent accepts from the other agents. Then the
individual heterosis potentials of the different
agents are combined in parallel with weighted mean aggregations, one for each
agent (i.e. taking into account the degrees of influence of each agent).
Consequently, a new heterosis potential is obtained
for each agent. The above parallel aggregations are repeated again and again
until a consensus between the agents is attained.
Keywords: Heterosis, Additive-dominance model, Bioinformatics, Recursive
aggregation.
References
[1] J.A.Birchler,
D.L.Auger, N.C.Riddle, In
Search of the molecular Basis of Heterosis, The Plant Cell 15 (2003) 2236 - 2239.
[2] E.Tsiporkova,
V.Boeva, Nonparametric Recursive Aggregation Process,
Kybernetika, Journal of the Czech Society for
Cybernetics and Inf. Sciences, 40 1
(2004) 51 - 70.
About a New Method for
Bioimpedance Measurement
Assoc. Prof. Dr.-Eng. Stanislav Dimov
Faculty for Engineering in German language-TU
Abstract - In the article is proposed
a new electrical method for bioimpedance measurement
[1]. The method is based on the indirect measurement of human skin bioimpedance via electrical generator. The electrical power
is applied on the active electrode on the human body and transmitted back via
neutral electrode in the generator. The active part of the bioimpedance
is calculated by means of microprocessor unit according
to the low of Ohm and the capacitive part according to the practical determined
curves of the skin capacitance from the applied voltage. With the proposed
method is obtained a good accuracy by the measurement of bioimpedance.
The method is large applicable in the electro surgery [2].The second topic in
the article is the design of a control program for the microprocessor unit by variable bioimpedance of the human body during
electro surgery intervention. The moment value of the bioimpedance
should be calculated and the output power of the RF generator should be
regulated. Another task, which is considered in the article, is the a priori
determination of the bioimpedance of the human skin
by the different physiological conditions of the patent and different age (for
example children, adults) [3].
Keywords: bioimpedance, electrical measurements, electro surgery,
computer modelling, electricity properties of the tissue
References:
[1] Melab, Lecture in
Biomedical Engineering, Department of Biomedical Engineering, Seoul National
University, Korea, 2004.
[2] Li Wing, Feasibility of
using an Implantable to measure thoroticic
Congestion, Pacing and Clinical Electrophysiology,vol.
28,Isue 5, 2005.
[3] Multi-frequency bioimpedance measurement of children in intensive care,
Medical & Biological Engineering & Computing, 2001.
Dissolution of Bi-component Fibrin Clots with Plasmin: Quantitation of the Modulating Effect of Myosin
Kiril
Tenekedjiev*, Balázs Váradi** and Krasimir Kolev**
* Department of Economics
and Management, Technical University - Varna, Bulgaria (correspondence author), kiril@dilogos.com
** Department of Medical
Biochemistry, Semmelweis University, Budapest, Hungary
Abstract - The effect of myosin on the
fibrin dissolution with plasmin is studied. Turbidimetric data evidence that myosin
retards the fibrin degradation with plasmin. Semi-quantitative evaluation of
the plasmin efficiency defines myosin as a potent modulator of fibrinolysis: at
0.5 molar ratio of myosin to fibrin monomers 8-fold higher plasmin
concentration is necessary to yield the same rate of dissolution as in the pure
fibrin clots. Using a dynamic non-steady state kinetic model and a
multifactorial optimization procedure to the experimental turbidimetric data
gained for various myosin-fibrin ratios and plasmin concentrations, we have
determined the rate constants for the interaction of plasmin with myosin and
fibrin-myosin complexes. The kinetic parameters of our new bi-component model
system suggest that the complex of myosin and fibrin differs markedly from its
separate constituents as a substrate of plasmin. A 50-fold decrease is detected
in the apparent catalytic constant for fibrin in the complex (0.14 s-1
versus 0.003 s-1), whereas the myosin degradation is less affected
(0.91 s-1 versus 0.04 s-1). Thus, in the examined
bi-component system the dissolution of the fibrin matrix with plasmin can
proceed only following removal of the myosin.
Keywords: thrombolysis,
proteolysis, kinetics, heterogeneous phase catalysis
One Way of Protein Structure Representation for Determining
Protein Structure Similarity |
Prof. Stoicho D. Stoichev
1 and Dobrinka Petrova
2
1Department of Computer Systems,
2Department of Computer Systems, Technical University
of Plovdiv, dobi_l5@yahoo.com
Abstract - As a result of many projects number of protein structures,
determined experimentally grows at an accelerated rate. However, it is
impossible to determine all of them by experiments. The requirement of
computational methods for protein structure determination is evident.
In this paper is considered a technique for specifying
super-secondary and tertiary protein structure using an autopsy of a PDB file. This
way of protein structure representation can be used for determining protein
structure similarity and classification. Proposed technique uses rules, which
are created to describe newly extracted structures and to compare them to
already known structures.
Keywords: protein structure representation, protein structure similarity,
classification, super-secondary, tertiary structure
References
[1] Alexandrov,
N.N. and Fischer, D. Analysis of topological and nontopological
structural similarities in the PDB: new examples with old structures.
[2]Rastall,ProteinArchitecture - http://www.food.rdg.ac.uk/online/fs916/lect3/lect3.htm.
[3] http://www.ncbi.nlm.nih.gov/Structure/VAST/vastsearch.html
An Algorithm for Determining Gene Activity Network
Stoicho D. Stoichev 1, Hristina Dinkova 2 and Nikola Kasabov3
1Department of Computer
Systems, Technical University - Sofia, stoi@tu-sofia.bg
2Department of
Computer Systems, Technical University - Sofia, chrissy_p_d@yahoo.com
3Knowledge Engineering
and Discovery Research Institute, KEDRI,
Director, Auckland University of
Technology, (http://www.kedri.info/), nkasabov@aut.ac.nz
Abstract - The correlation between activities of different genes in an organism is of great importance for the science (molecular biology, bioinformatics, medicine ,etc.). We represent these dependences as a directed weighted graph: each vertex represent a gene and the directed edge between genes g2 and g1 has a weight equal to a polynomial (of some degree) giving the partial dependence of the activity of g1 from the activity of g2. We propose an algorithm for determining the coefficients of these polynomials. The input data for the algorithm are values of activity of each gene for several (m) time moments (intervals are hours or days).
The activity
of the gene i is
represented by the formula
,
n is the number of the genes and
polynomials are of degree 4.
The function g(t)
given for several instances g(t1), g(t2), . . . , g(tm)
we approximate by linear segments. The number of the unknown coefficients in
the above formula is 5(n-1) and we need such a number of linear
equations for each gene, i.e. number of values of the activities for the moments
whose number is distributed proportionally to the derivatives of the segments
of the approximation function.
Hybrid Neural Network Applications
Albena Taneva and Michail Petrov*
Technical University Sofia, Branch Plovdiv, Control
Systems Department
*Head
of the Department
Abstract - Many problems still need new solutions in variety
areas. The scientists very often investigate and then attempt to copy the
approaches and methods from the Nature. This paper will present an algorithms
based on human knowledge and computer programming for solving and for improving
the control strategies.
The Takagi-Sugeno fuzzy logic, feed forward neural
network and simplified gradient decent learning are used as basic tools for
the: adaptive control [1] and predictive strategy design [2, 3].
The main focus here is an Adaptive Neuro Fuzzy
Architecture (ANFA) with Takagi Sugeno engine designed as a model and as an
optimizer for different control tasks. The programmed algorithms can be viewed as
examples of the Computational Intelligence submitted to the natural laws. The
common model ANFA has adaptive features and architecture and is based on human
knowledge to the particular task.
Related to the real time implementation is developed
simplified gradient decent algorithm - Recurrent Two steps Gradient Algorithm
(RTGA) for on-line training and updating of the model parameters.
The goal of the whole work was: to evaluate and adapt
the plant model parameters on-line, to optimized the control, and to carried
out real time experiments. The obtained results are showed that the developed
ANFA architecture with RTGA learning are promising tool and can be shifted to
solve the some tasks in Bioinformatics area.
Keywords: Adaptive Neural Network, Sugeno Fuzzy Logic, Local
modeling, Optimization
Reference:
[1] Petrov M., I.Ganchev, and A.Taneva. Fuzzy PID
Control of Nonlinear Plants. First International IEEE Symposium
"INTELLIGENT SYSTEMS", Varna, Bulgaria, September, 2002, pp.
30-35
[2] Taneva A. Predictive controller based on fuzzy
neural model. Journal of the Technical University of Plovdiv "Fundamental
Sciences and Applications", vol.9, 2002, series B
[3] Petrov M., I. Ganchev, A. Taneva. Fuzzy Model Predictive Control of Nonlinear
Processes. Preprints of the International Conference on "Automation
and Informatics'2002", Sofia, Bulgaria, 5-6 November, 2002, pp. 77-80
Learning by
Function Minimization Applied to Breast Cancer Diagnosis
Ludmil Dakovski1 and Zekie Shevked2
1Department of Computer Systems and Technologies,
2Department
of Computer Systems and Technologies, Technical
University of Plovdiv
Abstract - Machine learning proposes
helpful techniques for today's challenge of diagnosis and prediction. Our work
is concerned with learning from examples and its application to medical
diagnosis. We propose representing available training instances as logical
functions and applying a new strategy to minimization. The main goal is to find
a more compact representation of classification function and use it for
prediction of unknown cases. This approach performs well on the problem of
breast cancer diagnosis.
Key words: Learning from Examples, Function Minimization,
Classification, Breast Cancer Diagnosis.