darwin’s magic: evolutionary computation in nanoscience, bioinformatics and systems &...

Darwin’s Magic: Evolutionary Computation

in Nanoscience, Bioinformatics, Systems &

Synthetic Biology

Prof. Natalio Krasnogor

Automated Scheduling, Optimisation and Planning Research Group

School of Computer Science, University of Nottingham

www.cs.nott.ac.uk/~nxk

twitter.com/NKrasnogor

of 86IEEE Congress on Evolutionary Computation - New Orleans, USA, 2011

http://www.cs.nott.ac.uk://~nxk

http://www.cs.nott.ac.uk://~nxk

Outline

• Darwin’s Magic and Algorithmic Beauty

• Evolutionary Computation in the Natural Sciences– Self-Assembly and Scanning Probe Microscopy Optimisation– Structural Bioinformatics– Systems Biology & Synthetic Biology

• On Invariants, Decorations and the Future• Conclusions


Outline• Darwin’s Magic and Algorithmic Beauty

• Evolutionary Computation in the Natural Sciences– Self-Assembly and Scanning Probe Microscopy

Optimisation– Structural Bioinformatics– Systems Biology & Synthetic Biology



Darwin’s Magic


Thank you Youtube

Algorithmic Beauty1. Inheritable Instructions Set

2. Limited Resources

3. Imperfect Replication

A Powerful Secondary Effect: Selection

An awe inspiring product:

Evolution by Natural Selection


Evolutionary Computation in the Natural Sciences

• A Research VisionProgrammable algorithmic entry to the vast world of nanoscale physical, chemical & biological systems and processes

Algorithmic and Artificial Living Matter (ALMA)

Com

pute

r Sci

ence Embedded behavior

Information & Algorithms Complexity Robustness Tradeoffs

How does “The Logistics of Small Things” look like?

How (?) do you gain algorithmic entry into


The Spatial Scales Involved


ALMA & The Logistics of Small ThingsHow do you program complex nano/micro scale process :

• through billions of tiny & simple distributed programs/processors?

• when there is no clear distinction between hardware and software?

• when the wetware is not simply a stochastic program:

• when wetware is poorly characterised and is likely to evolve, etc.

function f1(p1,p2,p3,p4){ if (p1<p2) and (rand<0.5)

print p3 else

print p4}

function f1(p1,p2,p3,p4){ if (p1<p2)

RND print p3

RND else

RNDprint p4RND

}


RND print p3

RND else

RNDprint p4RND

}


RND incr p3

RND else

RNDdecr p4RND

} of 86IEEE Congress on Evolutionary Computation - New Orleans, USA, 2011



Optimisation– Structural Bioinformatics– Systems Biology– Synthetic Biology



Molecular Tiles & Programmable Self-Assembly

Algorithmic Self-Assembly of DNA Sierpinski Triangles. P.W.K. Rothemund, N. Papadakis, E. Winfree. PLoS Biology 2:12 (2004)


Tiles System

Supra-structure

Which is the correct input ?

• Finite size square lattice (300x300)

• Fixed T = 4

10 tiles

10 tiles

How can we automatically design a tile system that self-assembles into a target shape?

c1

c4

c3

c0

c0 c1 c2 c3 c4

3 7 2 8 1

7 6 7 0 1

2 7 0 5 3

8 0 5 4 1

1 1 3 1 9

c5 c6 c7 c8 c9

c7

c6

c9

c8

c5

2 3 7 6 5

5 2 3 5 2 4 1 2 3 9

0 4 5 2 3 1 0 6 5 1

6 2 7 7 7 2 6 8 2 9

3 8 3 3 6 3 5 2 7 8

9 7 1 1 5 9 1 9 8 1

5 2 7 3 1

3 5 7 1 6

2 4 2 8 7

5 0 6 3 9

c2

Glue strength matrix M

Evolving tiles for automated self-assembly design. G. Terrazas, M. Gheorghe, G. Kendall, and N. Krasnogor. Proceedings for the 2007 IEEE Congress on Evolutionary Computation, 2007. Best paper award.

Toward minimum size self-assembled counters by P. Moisset de Espanes, A. Goel. Nat Comput (2008) 7:317–334


Tiles with deterministic assembly (Model 1)

Tiles with probabilistic assembly (Model 2)


Phenotype – Fitness Mapping

Minkowski functionals (A, P, X)

A = 12P = 24X = 0

A = 100P = 40X = 1

Evolutionary Design ApproachVariable length individuals

(Genotype) Genotype -Phenotype Mapping

Phenotype

Randomly created Wang tiles

Bitwise mutation

Vs

One-point crossover

Population size = 100, Individuals length = [1,10], Generations = 300, Pcrossover = 0.7, PMutation = 0.1/0.05/0.01


Probabilistic Assembly +No Rotation

Probabilistic Assembly +Rotation

Deterministic Assembly +Rotation

Deterministic Assembly +No Rotation


Two-tile self-assembly Three-tile self-assembly Four-tile self-assembly Five-tile self-assembly

We calculated the equivalence classes of binding pockets defined by “bp1 R bp2 iif NAFE(bp1)=NAFE(bp2)” for the best tile set.

We observed that equivalence classes with NAFE smaller than T are highly likely to participate in the self-assembly process as these are more populous.

More “assembable” binding pockets = Generalised Secondary Structures

How Does Self-Assembly Gets Programmed?


Triangular site lattice

Physical events to capture1. Adsorption: tiles are placed on the substrate at a given

rate2. Diffusion: tiles move or rotate from one position to another

allowing:• Separation from one or more tiles• Motion along a line of tiles• Motion without interaction

3. Diffusion across terraces on the substrate4. Intramolecule strength: energy between two no-

functionalised porphyrins 5. Molecule-substrate strength: energy of a porphyrin to the

substrate6. Rotational strength: molecule-substrate strength for

spinning

Neighbourhood size 6

Neighbourhood size 8

20

0

3

42

DNA Tiles are Too Big!


How Do You Image and Manipulate at This Scale?

D.L. Keeling et al. Phys. Rev. Lett 94, 146104 (2005)

Y. Sugimoto et al., Nature letters 446, 64 (2007).

Hla et al. Phys. Rev. Lett. 85, 2777–2780 (2000)

C60

D. M. Eigler & E. K. Schweizer, Nature 344, 524 - 526 (1990)


Even 3 Variable Problems are Difficult: Optimising a Scanning Probe Microscopy

The tunnel current it is highly dependant on the tip-sample distance, d. This current can be maintained with a feedback loop, G, that actively controls the tip-sample distance.

it exp(−2kd)∝

AScanning tip

Sample surface

X

Z

Y

Axis under direct (piezo) control

G i

V

http://www2.fz-juelich.de/ibn/index.php?index=1021


Understanding the image

J. H. A. Hagelaar et al. PRB 78, 161405R 2008 L.Gross et al. Science 325 1110 (2009)


(Un)Stable and (Un)defined Tip States

Imaging problems, spontaneous tip changes


Two Stage Automation Process

Ex-situIn-situ

Voltage pulsing (deliberate crash) Fine tuning

(changing scan parameters)

Automated probe microscopy via evolutionary optimisation at the atomic scale. R. Woolley, J. Sterling, A. Radocea, N. Krasnogor and P. Moriarty. Applied Physical Letters (to appear)

Cellular GA with Smart Initialisation


Stage 1: Smart Initialisation (coarsely) Conditions the Probe

Streaky Image. Executing cleaning pulse

Cloudy Image. Executing cleaning pulse

Flat Surface. Zooming in to 50nm

Flat Surface. Zooming in to 20nm

Constant Atomic resolution. Zooming in to 4nm

Poor Atomic resolution.Rescanning

Consistent fair atomic resolution. Stage 1 complete. Time elapsed: 1010.1902 (~17mins)

A deterministic approach


Stage 2: Fine adjustment with CGAStarting image

Machine Optimised

GiV

GiV

GiV

GiV

GiV

Cellular GA


Do I really need a cGA? Would a stochastic selection be just as good?

•Standard deviation is from the ‘noise’ of the GA

•RMI average 0.12Insets: 1x1nm2

(a) before cGA, (b) optimised.

•Stochastic selection of parameters, average RMI 0.01


How Does it Compares to an Expert Operator?

Microscopist Ave. RMI

Change in RMI/min

i-SPM 0.20 7.1Human 0.09 2.6


Primary Sequence 3D Structure

Protein Folding & Structure Prediction

Protein Structure Prediction (PSP) aims to predict the 3D structure of a protein based on its primary sequence

(perhaps disregarding the folding process)

Anfinsen’s thermodynamic hypothesis [Anfinsen 1973, Dill and Chan 1997]


Defining and Predicting Useful Features

Contact M. Stout, J. Bacardit, J. Hirst & N. Krasnogor, Bioinformatics 2008

24(7):916-923.

M. Stout, J. Bacardit, J.D. Hirst, R.E Smith, and N. Krasnogor. Prediction of topological contacts in proteins using learning classifier systems. Journal Soft Computing - A Fusion of Foundations, Methodologies and Applications, 13(3):245-258, 2008.


Integrating Multiple Prediction Sources

1. Prediction of Secondary structure (using PSIPRED) Solvent Accessibility Recursive Convex Hull Coordination Number

2. Integration of all these predictions plus other sources of information

3. Final CM prediction (using BioHEL)

Using BioHEL


The BioHEL GBML System• BIOinformatics-oriented Hiearchical Evolutionary Learning –

BioHEL (Bacardit & Krasnogor, 2009)

• BioHEL is a rule-based evolutionary learning system that employs the Iterative Rule Learning (IRL) paradigm– First used in EC in Venturini’s SIA system (Venturini, 1993)– Widely used for both Fuzzy and non-fuzzy evolutionary

learning

J. Bacardit, M. Stout, J.D. Hirst, K. Sastry, X. Llora, and N. Krasnogor. Automated alphabet reduction method with evolutionary algorithms for protein structure prediction. Proceedings of the 2007 Genetic and Evolutionary Computation Conference, ACM Press, 2007.

J. Bacardit, M. Stout, J.D. Hirst, A. Valencia, R.E. Smith, and N. Krasnogor. Automated alphabet reduction for protein datasets. BMC Bioinformatics, 10(6), 2009.

Bronze Medal in the THE 2007 “HUMIES” AWARDS FOR HUMAN-COMPETITIVE RESULTS PRODUCED BY GENETIC AND EVOLUTIONARY COMPUTATION.


How are these features predicted?

Many of these features are due to local interactions of an amino acid and its immediate neighbours We predict them from the closest neighbours

in the chain

Ri

SSi

Ri+1

SSi+1

Ri-1

SSi-1

Ri+2

SSi+2

Ri-2

SSi-2

Ri+3

SSi+3

Ri+4

SSi+4

Ri-3

SSi-3

Ri-4

SSi-4

Ri-5

SSi-5

Ri+5

SSi+5

Ri-1 Ri Ri+1 SSi

Ri Ri+1 Ri+2 SSi+1

Ri+1 Ri+2 Ri+3 SSi+2 of 86IEEE Congress on Evolutionary Computation - New Orleans, USA, 2011

Contact Map dataset

The set of 2811 proteins was randomly halved Moreover, all proteins with more than 350

amino acids were discarded Still, the resulting training set contained more

than 15.2 million instances and 631 attributes Less than 2% of those are actual contacts 36GB of disk space


Samples and ensembles

50 samples of 300K examples are generated from the training set with a ratio of 2:1 non-contacts/contacts

BioHEL is run 25 times for each sample

Prediction is done by a consensus of 1250 rule sets

Confidence of prediction is computed based on the votes distribution in the ensemble.

Whole training process takes about 289 CPU days (~5.5h/rule set)

Training set

x50

x25

Consensus

Predictions

Samples

Rule sets


Critical Assessment of Techniques for Protein Structure Prediction

CASP facts biannual competition

started in 1994 parallel prediction and

experimental verification model assessment by

human experts

9th edition of CASP 150 human groups 140 server groups


Contact Map prediction in CASP 7

Ezkudia et al. Proteins 2009; 77(Suppl 9):196-209

Accuracy for groups that predicted a common subset of targets


Xd results




Remarkable Prediction

L/10 prediction for target T0443-D1

67% accuracy



A larger set of proteins was employed Set of 3262 proteins for training all the 1D

predictors A subset of 2413 proteins used for CM

prediction All proteins with less than 250AA A randomly selected 20% for larger chains

50 Samples of ~660000 instances were generated

The representation remained unchanged 25K CPU hours were employed just to

train the CM ensemble



In terms of performance

These two groups derived contact predictions from 3D models


Energy landscape all-atom force field statistical potential

Search method random walk structure optimisation Folding@home 8.5 peta FLOPS 10 000 CPU days for 10μs of folding[Dill and Chan 1997]

P. Widera, J.M. Garibaldi, J., and N. Krasnogor,. Evolutionary design of the energy function for protein structure prediction, Proceedings of the IEEE Congress on Evolutionary Computation 2009.

P. Widera, J. Garibaldi, and N. Krasnogor. GP challenge: evolving the energy function for protein structure prediction. Journal of Genetic Programming and Evolvable Machines, 11:61-88, 1 2010.

Gold Medal in the THE 2010 “HUMIES” AWARDS FOR HUMAN-COMPETITIVE RESULTS PRODUCED BY GENETIC AND EVOLUTIONARY COMPUTATION

Improving the Energy Function for Full 3D PSP


How to find good quality models?Correlation between energy and distance to the native structure

Requirements energy reflects

distance distance reflects

similarity


How the best of CASP do it?Energy of models vs. distance to a target structure

Similarity measure

Decoys generated by I-TASSER [Wu et al. 2007] Robetta [Rohl et al. 2004]


How the energy function is designed?Weighted sum vs. free combination of terms

Decision support local numerical

approximation

GP input terminals: T1 … T8

functions:add sub mul divsin cos exp log

random ephemerals in range [0,1]

Zhang et al. 2003


Can GP improve over a weighted sum of terms?Nelder-Mead downhill simplex optimisation

Computational cost of experiments 55 proteins, 1000-2000 structures for each 5 different ranking distance measures 20 different configurations of GP parameters total of 150 CPU days

spearman-sigmoid correlation

method d-100 all d-100 all

simplex 0.734 0.638 0.650 0.166

GP 0.835 0.714 *0.740 *0.200


The Cell as an Information Processing Device

LeDuc et al. Towards an in vivo biologically inspired nanofactory. Nature (2007)


Transcription Networks

Gene1 Gene2 Gene3 Genek

Genome

Transcription Factors

Signal2 Signal5Signal1 Signal3 Signal4 Signaln...Environment


Network Motifs: Evolution’s Preferred Circuits

• Biological networks are complex and vast

• Moreover, these patterns are organised in non-trivial/non-random hierarchies

“Patterns that occur in the real network significantly more often than in randomized networks are called network motifs” Shai S. Shen-Orr et al., Network motifs in the transcriptional regulation network of

Escherichia coli. Nature Genetics 31, 64 - 68 (2002)

Radu Dobrin et al., Aggregation of topological motifs in the Escherichia coli transcriptional regulatory network. BMC Bioinformatics. 2004; 5: 10.

Each network motif carries out a specific information-processing function

The C1-FFL is a ‘sign-sensitive delay’ element and a persistence detector.

The I1-FFL is a pulse generator and response accelerator


Evolvable Executable Biology


Nested EA for Model Synthesis

F. Romero-Campero, H.Cao, M. Camara, and N. Krasnogor. Structure and parameter estimation for cell systems biology models. Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2008), pages 331-338. ACM Publisher, 2008. Best Paper Award

H. Cao, F.J. Romero-Campero, S. Heeb, M. Camara, and N. Krasnogor. Evolving cell models for systems and synthetic biology. Systems and Synthetic Biology , 2009


The Fitness Function• Multiple time-series per target

• Different time series have very different profiles, e.g., response time or maxima occur at different times/places

• Transient states (sometimes) as important as steady states

•RMSE will mislead search

•Sometimes the time series is qualitative or microarray data

H. Cao, F.J. Romero-Campero, S. Heeb, M. Camara, and N. Krasnogor. Evolving cell models for systems and synthetic biology. Systems and Synthetic Biology , 2009


Problem Specification


Target


A Signal Translatorfor Pattern Formation

act1Prep2

act2Prep1

rep1Pact1

rep2Pact2

rep3Prep1

rep4Prep2

I2Prep3

I1Prep4

FP2Pact2

FP1Pact1


Uniform Spatial Distribution of Signal Translators for Pattern Formation

E. coli DH5α ∆sdiA/∆lacI (2∆)

pACYC184pBR322


Pattern Formation in synthetic bacterial colonies


pAYCP (1-3)pBR322 (4-6)

2∆ DH5αStarting OD=10

Magnification: 100X


pUC6S (1-6)

2∆ DH5α

Starting OD= 10

Magnification: 40X




Optimisation– Structural Bioinformatics

• Systems Biology & Synthetic Biology



Calculating Pi

Algorithms are Tiny

Factoring: Let n be the number to be factored.

1. Let Δ be a negative integer with Δ = -dn where d is a multiplier and Δ is the negative discriminant of some quadratic form.

2. Take the t first primes , for some .

3. Let fq be a random prime form of GΔ with .

4. Find a generating set X of GΔ

5. Collect a sequence of relations between set X and {fq : q ∈ PΔ} satisfying:

6. Construct an ambiguous form (a, b, c) which is an element f ∈ GΔ of order dividing 2 to obtain a coprime factorization of the largest odd divisor of Δ in which Δ = -4a.c or a(a - 4c) or (b - 2a).(b + 2a)

7. If the ambiguous form provides a factorization of n then stop, otherwise find another ambiguous form until the factorization of n is found. In order to prevent that useless ambiguous forms are generated, build up the 2-Sylow group S2(Δ) of G(Δ).


http://en.wikipedia.org/wiki/Sylow_theorems

They are NOT Algorithms!➡They do not stop, we stop them.➡They are not short pieces of code, but large

systems

What Evolutionary Algorithms are NOT?


What are Evolutionary Algorithms?

N. Krasnogor and J.E. Smith. IEEE Transactions on Evolutionary Computation, 9(5):474- 488, 2005.

Research Paradigms for Problem Solving

T.S. Kuhn. The Structure of Scientific Revolutions, 1962.

Design Patterns and Pattern Languages

C. Alexander, S. Ishikawa, M. Silverstein, M. Jacobson, I. Fiksdahl-King, S. Angel, S.: A Pattern Language - Towns, Buildings, Construction. Oxford University Press (1977)


A Compact “Memetic” Algorithm by Merz (2003)

Invariants and Decorations


A “Memetic” Particles Swarm Optimisation by

Petalas et al (2007)



A “Memetic” Artificial Immune System by Yanga et al (2008)



A “Memetic” Learning Classifier

System by Bacardit & Krasnogor

(2009)



• Many others based on Ant Colony Optimisation, NN, Tabu Search, SA, DE, etc.

• Key Invariants:– Global search mode– Local search mode

• Many Decorations, e.g.:– Crossover/Mutations (EAs based MAs)– Pheromones updates (ACO based MAs)– Clonal selection/Hypermutations (AIS based MAs)– etc



A Pattern Language for Memetic AlgorithmsMemetic Algorithms by N. Krasnogor. Handbook of Natural Computation (chapter) in Natural Computing. Springer Berlin /

Heidelberg, 2009. www.cs.nott.ac.uk/~nxk/publications.html


http://www.cs.nott.ac.uk/~nxk/publications.html

A General Trend: moving away from close-loop optimisation towards open-ended and embodied

optimisation

Effort (e.g. Time, $$$, etc)Programming solving 1 problem – single instances

Programming solving 1 problem – several instances(self) adaptive

Programming Solving a few problem – several classes instances(self) adaptive Self-generating

Programming Solving multiple unrelated problem – several classes instances(self) adaptive Self-generating

Self-Engineering

Reuse

Reuse

Effort (e.g. Time, $$$, etc)



Feedback

Reuse Feedback


The Future of EAsSoftware Nurseries

• Fundamental Change of Temporal Scales Rethink• Software will be “seeded” and grown, very much like

a plant or animal (including humans)• Software will start in an “embryonic” state and

develop when situated on a production environment• What would a software “incubation” machine look

like?• What would a software “nursery” look like?


Cells

Organs

Tissue

Individual

DNA/RNA

Potential To Develop into multiple

different types of cells

Commitment

Specialised Function

Ultimate Solver


http://faculty.washington.edu/kepeter/119/images/sheep_heart4.jpg

Software Cell SC

SC

SC

SC

SCSC

Pluripotential Solver“DNA”

TSP Organ

Euclidean TSP Organ

GraphicalTSP Organ

TSPSolver

SoftwareOrganism

Production EnvironmentInput


Protein Structure PredictionSolver

SoftwareOrganism

Vehicle RoutingSolver

SoftwareOrganism

Graph IsomorphismSolver

SoftwareOrganism

SATSolver

SoftwareOrganism

Bin PackingSolver

SoftwareOrganism

Graph ColoringSolver

SoftwareOrganism

Network InterdictionSolver

SoftwareOrganism

Quadratic AssignmentSolver

SoftwareOrganism

TSPSolver

SoftwareOrganism

An Ecosystem of solvers


As, e.g., Biologists & Physicists have done through an ubiquitous, worldwide spanning informatics

infrastructure, we should be focusing on building an online

worldwide computational problem solving infrastructure




Optimisation– Structural Bioinformatics

• Systems Biology & Synthetic Biology



Conclusions• New types of executable structures• In Nanotechnology

– DNA tiles, DNA origami, etc– Non DNA based tiles– Some have very definite programmable features– Others require the program to be “distributed” and exploit noise and

randomness

• In Synthetic Biology– How to orchestrate activities at multiple temporal-spatial-energetic

scales?– How to cope with noise in the background that execures a program and

in the program itself?!– How to cope for programs that will evolve?


Conclusions• New types of benchmarks• Structural Biology (PSP and GP4PSP)

– Many of these problems can be modelled both as regression or classification problems

– Low/high number of classes– Balanced/unbalanced classes– Adjustable number of attributes– Ideal benchmarks !!

• Scanning Probe Microscopy: – Even a few dimensions are hard– “Chameleons” as it is sampled

• http://www.infobiotic.net/


http://www.infobiotic.net/PSPbenchmarks/

• The emerging trend is moving away from close-loop optimisation towards open-ended and embodied optimisation

• Requires strong links with data mining, ALIFE and, of course, AI (beyond existing trends in constraint satisfaction), search based software engineering (beyond current trends on testing/debugging)

Conclusions

•Requires on-line, computer friendly ontologies of code (e.g the pattern language in the left), self-describing source code, protocols for autonomic code reuse, etc


Missing Components

Missing Components

• Learn From Physics, Chemistry & Biology The Invariants & Patterns, the Decorations are superfluous

• Evolution • Self-Assembly & Self-Organisation• Developmental systems

– Depend on a core genome coding for essential functionality

– Epigenomics canalises development• Hierarchical control systems that modify programs including

susceptibility to horizontal gene (program libraries) transfer• Infrastructure

Conclusions


Acknowledgements

F. Romero-CamperoJ. Twycross

G. Terrazas

J. Bacardit

J. ChaplinP. Widera

A. Ali Shah

J. Blakes

E. Glaab

K. Righetti

M. Franco

D. Sannasy

L. T. Leong

CEC organisers A.E. Smith I. Parmee G. Kendall M. Schoenauer

• M. Camara, S. Heeb, P. Williams• C. Alexander• P. Moriarty, P. Beton• N. Chapness, R. Wooley• M. Holdsworth, G. Basel• Colleagues at ASAP


darwin’s magic: evolutionary computation in nanoscience, bioinformatics and systems &...

Education

ieee congress

natural sciencesselfassembly

rotationdeterministic

rotationprobabilistic

outlinedarwins magic

deterministic assembly

function f1p1

natural selectionpage5