Towards Evolutionary Network Reconstruction Tools for Systems - - PowerPoint PPT Presentation

▶
towards evolutionary network reconstruction tools for
SMART_READER_LITE
LIVE PREVIEW

Towards Evolutionary Network Reconstruction Tools for Systems - - PowerPoint PPT Presentation

Nov 24, 2022 466 likes •674 views

Introduction Artificial Network Evolution Results Case Study Conclusions Towards Evolutionary Network Reconstruction Tools for Systems Biology T. Lenser T. Hinze B. Ibrahim P . Dittrich {thlenser,hinze,ibrahim,dittrich}@cs.uni-jena.de

slide-1
SLIDE 1

Introduction Artificial Network Evolution Results Case Study Conclusions

Towards Evolutionary Network Reconstruction Tools for Systems Biology

  • T. Lenser
  • T. Hinze
  • B. Ibrahim

P . Dittrich

{thlenser,hinze,ibrahim,dittrich}@cs.uni-jena.de

Bio Systems Analysis Group Friedrich Schiller University Jena www.minet.uni-jena.de/csb 5th European Conference on Evolutionary Computation, Machine Learning and Data Mining in Bioinformatics

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-2
SLIDE 2

Introduction Artificial Network Evolution Results Case Study Conclusions

Outline

Towards Evolutionary Network Reconstruction Tools in Systems Biology

Introduction

Motivation, Cell Signalling, ESIGNET

Artificial Network Evolution

Two-Level Evolutionary Algorithm Operators, Parameterisation, Fitting Selection and Fitness Evaluation

Results

Evolving Arithmetic Functions log,

3

√ Effect of Duplication Operator

Case Study: Spindle Checkpoint

Biological Background Modelling and Evol. Optimisation

Conclusions

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-3
SLIDE 3

Introduction Artificial Network Evolution Results Case Study Conclusions

Motivation

  • Systems Biology deals with interplay of

biological components rather than components themselves.

  • Reconstructing nonlinear networks from

(incomplete) data is a necessary but difficult task. = ⇒ Evolutionary computing is well suited to this!

  • Furthermore, bio-inspired algorithms provide a flexible,

fault-tolerant, reliable computing paradigm. = ⇒ Evolutionary computing can support design of such algorithms.

  • Help in understanding emergence of biological complexity.

= ⇒ Evolution becomes observable.

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich gene expression data visualised by microarray (TU Dresden, BIOTEC)

slide-4
SLIDE 4

Introduction Artificial Network Evolution Results Case Study Conclusions

Motivation

  • Systems Biology deals with interplay of

biological components rather than components themselves.

  • Reconstructing nonlinear networks from

(incomplete) data is a necessary but difficult task. = ⇒ Evolutionary computing is well suited to this!

  • Furthermore, bio-inspired algorithms provide a flexible,

fault-tolerant, reliable computing paradigm. = ⇒ Evolutionary computing can support design of such algorithms.

  • Help in understanding emergence of biological complexity.

= ⇒ Evolution becomes observable.

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich gene expression data visualised by microarray (TU Dresden, BIOTEC)

slide-5
SLIDE 5

Introduction Artificial Network Evolution Results Case Study Conclusions

Motivation

  • Systems Biology deals with interplay of

biological components rather than components themselves.

  • Reconstructing nonlinear networks from

(incomplete) data is a necessary but difficult task. = ⇒ Evolutionary computing is well suited to this!

  • Furthermore, bio-inspired algorithms provide a flexible,

fault-tolerant, reliable computing paradigm. = ⇒ Evolutionary computing can support design of such algorithms.

  • Help in understanding emergence of biological complexity.

= ⇒ Evolution becomes observable.

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich gene expression data visualised by microarray (TU Dresden, BIOTEC)

slide-6
SLIDE 6

Introduction Artificial Network Evolution Results Case Study Conclusions

Biological Principles of Cell Signalling

Information Processing in Living Cells

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

genomic dna gene expression

cell membrane

phospholipid bilayer

cytosol

transformation, amplification via pathways signal transduction, cell response ADP ATP phosphorylation activation by protein kinases activation cascade GDP GTP

external signal

endocrine (dist.) paracrine (near) autocrine (same cell) ligands hormones, factors, ... inner membrane

receptors

enzyme−linked ion−channel G−protein−linked

nucleus

slide-7
SLIDE 7

Introduction Artificial Network Evolution Results Case Study Conclusions

ESIGNET – Research Project

Evolving Cell Signalling Networks (CSNs) in silico

European interdisciplinary research project

  • University of Birmingham (Computer Science)
  • TU Eindhoven (Biomedical Engineering)
  • Dublin City University (ALife Lab)
  • University of Jena (Bio Systems Analysis)

Objectives

  • Study the computational properties of CSNs
  • Developing new ways to model and predict real CSNs
  • Gain new theoretical perspectives on real CSNs

Computing Facilities

  • Cluster of 33 workstations

(two Dual Core AMD OpteronTM 270 processors, Rocks Linux)

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-8
SLIDE 8

Introduction Artificial Network Evolution Results Case Study Conclusions

ESIGNET – Research Project

Evolving Cell Signalling Networks (CSNs) in silico

European interdisciplinary research project

  • University of Birmingham (Computer Science)
  • TU Eindhoven (Biomedical Engineering)
  • Dublin City University (ALife Lab)
  • University of Jena (Bio Systems Analysis)

Objectives

  • Study the computational properties of CSNs
  • Developing new ways to model and predict real CSNs
  • Gain new theoretical perspectives on real CSNs

Computing Facilities

  • Cluster of 33 workstations

(two Dual Core AMD OpteronTM 270 processors, Rocks Linux)

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-9
SLIDE 9

Introduction Artificial Network Evolution Results Case Study Conclusions

ESIGNET – Research Project

Evolving Cell Signalling Networks (CSNs) in silico

European interdisciplinary research project

  • University of Birmingham (Computer Science)
  • TU Eindhoven (Biomedical Engineering)
  • Dublin City University (ALife Lab)
  • University of Jena (Bio Systems Analysis)

Objectives

  • Study the computational properties of CSNs
  • Developing new ways to model and predict real CSNs
  • Gain new theoretical perspectives on real CSNs

Computing Facilities

  • Cluster of 33 workstations

(two Dual Core AMD OpteronTM 270 processors, Rocks Linux)

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-10
SLIDE 10

Introduction Artificial Network Evolution Results Case Study Conclusions

Artificial Network Evolution

Introductory Example

Task: addition of two positive real numbers

R0

  • utput1

X1 input1 R1 X2 R2 input2 R0 X1

  • utput1

input2 R2 R1 input1

snapshots of artificial network evolution

  • R0, R1, R2 identify reactions
  • input1, input2, output1:

distinguished species

  • X1, X2: auxiliary species
  • Stepwise modification of network

structure and kinetic parameters

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-11
SLIDE 11

Introduction Artificial Network Evolution Results Case Study Conclusions

Two-Level Evolutionary Algorithm

Artificial Network Evolution in Detail

  • Separation of structural evolution from parameter fitting
  • Idea: parameters can adapt to mutated network structure

R R

R

i i

i

n n

n

p p

p

u u

u

t t

t
  • u

u

u

t t

t

p p

p

u u

u

t t

t

R R

R

1 1

1

R

R

2

2

r r

r

e e

e

s s

s
  • u

u

u

r r

r

c c

c

e e

e

X

X

4

4

Mutation of Network Structure

mutate parameters create offspring & selection parameter sets

Parameter Fitting & Fitness Evaluation Selection & Offspring Creation Population

  • Upper level: network structure, analogue to graph-GP
  • Lower level: parameter fitting using standard Evolution

Strategy = ⇒ All networks handled as SBML models

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-12
SLIDE 12

Introduction Artificial Network Evolution Results Case Study Conclusions

Operators and Parameterisation

Artificial Network Evolution in Detail

EA used here employs eight different mutations Operators for structural evolution

  • Addition/deletion of a species
  • Addition/deletion of a reaction
  • Connection/removal of an existing

species to/from a reaction

  • Duplication of a species with all its

reactions (discussed in detail later) Operator for parameter evolution

  • Mutation of a randomly selected

kinetic parameter by addition

  • f a Gaussian variable

Network size can be limited.

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

addition of a species deletion of a species addition of a reaction deletion of a reaction disconnection of species connection of species species duplication

slide-13
SLIDE 13

Introduction Artificial Network Evolution Results Case Study Conclusions

Parameter Fitting by an Evolution Strategy

Artificial Network Evolution in Detail

  • Small population size

= ⇒ due to high computational costs of fitness evaluation

  • Non-overlapping generations (comma-selection)

= ⇒ supports self-adaptation

  • Self-adaptation of strategy parameters

= ⇒ balancing between exploration of search space and fine-tuning

  • Parameter settings copied from parent to offspring

networks = ⇒ incremental parameter fitting

  • Initial parameters uniformly distributed between given

minimal and maximal values = ⇒ no extra bias introduced

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-14
SLIDE 14

Introduction Artificial Network Evolution Results Case Study Conclusions

Selection and Fitness Evaluation

Artificial Network Evolution in Detail

Selection on structural level

  • Overlapping generations (plus-selection), elitistic

= ⇒ good solutions cannot be lost

  • Fixed population size (10 . . . 100)

= ⇒ due to computational costs Fitness evaluation

  • Numerical integration of reaction

network using ODE solver (SOSlib)

  • Fitness measure given by weighted

squared distance to target time course (output species)

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

addition of two positive real numbers (introductory example) fitness development (best, average, worst)

slide-15
SLIDE 15

Introduction Artificial Network Evolution Results Case Study Conclusions

Result: Logarithm Network

Evolving Arithmetic Functions

  • We compare three settings:
  • two-level EA
  • one-level EA (simultaneous structural and parameter

evolution) for many generations

  • one-level EA with a larger population
  • Setup such that all approaches use same number of

fitness evaluations (normalisation)

  • Two-level approach clearly superiour for this task
  • Both other approaches converge prematurely

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-16
SLIDE 16

Introduction Artificial Network Evolution Results Case Study Conclusions

Result: Third Root Network

Evolving Arithmetic Functions

  • Same three settings as in logarithm example
  • Depicted on logarithmic scale
  • Result also confirms advantage of two-level EA
  • Differences are not as pronounced as in logarithm example

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-17
SLIDE 17

Introduction Artificial Network Evolution Results Case Study Conclusions

Result: Effect of Duplication Operator

Duplication of a species with all its reactions

species duplication

  • Search for “soft" mutation operators
  • Inspired by gene duplication in living organisms
  • Adapted to evolution of arithmetic functions

2 4 6 8 x 10

5

0.1 0.2 0.3 0.4 0.5 0.6 0.7 Duplication and Addition Fitness evaluations Average fitness 2 4 6 8 x 10

5

0.1 0.2 0.3 0.4 0.5 0.6 0.7 Only Addition Fitness evaluations Average fitness 2 4 6 8 x 10

5

0.1 0.2 0.3 0.4 0.5 0.6 0.7 Only Duplication Fitness evaluations Average fitness

  • In log example, duplication doesn’t improve or worsen observed results
  • However, we still regard duplication as potentially promising

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-18
SLIDE 18

Introduction Artificial Network Evolution Results Case Study Conclusions

Case Study: Human Spindle Assembly Checkpoint

Biological Background

  • Sequence of events starting from
  • ne cell leading to two daughter cells
  • Focus on checkpoint mechanism in

mitosis (cell division)

  • Spindle Assembly

Checkpoint (SAC)

  • SAC prevents cell cycle progression until all chromosomes are

attached to mitotic spindle

  • Defects lead to cell death, aneuploidy, aging, and cancer

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-19
SLIDE 19

Introduction Artificial Network Evolution Results Case Study Conclusions

Case Study: Human Spindle Assembly Checkpoint

Modelling and Evolutionary Optimisation

  • 17 species, 11 reactions
  • Compartments represent

chromosomes

  • Structural evolution adds two

(unrealistic) reactions:

BubR1Z → Mad1∗

X + BubR1∗ Y

BubR1∗

X + Cdc20Y

→ Mad2X + Cdc20Y

1 2 3 4 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Phases Steady state level of APC:Cdc20

target: ———low——— –high–

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich

slide-20
SLIDE 20

Introduction Artificial Network Evolution Results Case Study Conclusions

Conclusions and Outlook

  • Two level approach enhances performance of EA for

biological network reconstruction

  • Duplication operator interesting and promising approach in

general, although first results for the logarithm network show no convincing effect yet

  • So far, we mostly tested evolution of networks for

arithmetic functions

  • Evolutionary method can improve and predict realistic

complex networks exemplified here by human Spindle Assembly Checkpoint

  • Further studies will target additional parameter settings as

well as application to evolution of computing devices

  • Interested in our software? thlenser@minet.uni-jena.de

Towards Evolutionary Network Reconstruction Tools

  • T. Lenser, T. Hinze, B. Ibrahim, P

. Dittrich