arXiv:nlin/0006051v1 [nlin.AO] 30 Jun 2000 sta re ilapa ohsotnosyaddue and spontaneously both appear will order answer that correct is As the debate selection? selectionist/neutralist “spon- the inherent in appear any order of does irrespective or taneously”, the the environment, to of its due consequence system of a a effect in order place biolog- take is the that changes emer- than put: adaptive rather the Simply physical on world. the ical inspiration in natural order for the of draws pre- gence in which been order has [1] a paradigm of sented and another source recently century More principle a world. the nearly for natural as alone of half concept stood selec- Darwinian has of the selection roles mutation competing and the tion over debate the Modulo Introduction 1 FETO UAINADRCMIAINO THE ON RECOMBINATION AND MUTATION OF EFFECT oesadas ysm uhmr compli- more simulations. much two-locus numerical some cated and by also one and simple models asso- some results with analytic ciated explicit, are by points above supported The the more brittleness. of a of some problems to circumventing leads evolution potentially it robust it Additionally, that phe- shown the self- is orthogenesis. to or and of order system in nomenon the increase in an organization to gener- leads it that ically shown is can it Specifically, inves- of operators are tigated. breaking” consequences such symmetry The “induced that this symmetry. shown this val- is break fitness it equal have to ues, phenotype corresponding same genotypes “symme- the many a geno- i.e. exhibits the try”, landscape when fitness particular, is typic In map recombi- genotype-phenotype and considered. the than mutation on other as nation, operators such genetic selection, of effect The Abstract EOYEPEOYEMAP GENOTYPE-PHENOTYPE NM icioEtro,APsa 70-543 A.Postal Exterior, Circuito UNAM, NP nttt eCeca Nucleares, Ciencias de Instituto NNCP, -al [email protected] e-mail: ´ xc ..04510 M´exico D.F. .R Stephens R. C. oemc oenntiilmdl n nscin5I 5 section conclusions. in in some and found make results models will some non-trivial discuss more briefly much models. will some two-locus I and 4 in one section simple breaking In some symmetry of give context induced will the I of 3 examples section In sym- analytic order, breaking. of symmetry concepts and the introduce metry will I 2 phe- section In the about. come circumstances may what orthogenesis under of and nomenon we particular, how In see understood. will be can framework self-organize this quantitative may which and map within how- qualitative this 7], a how show provide [6, to and instance new is for map it empha- see ever genotype-phenotype to evolution, the new Darwinian not of in course pheno- importance of same the is the size It to value. correspond fitness typic genotypes in many-to-one, many inherent is it that i.e. when is map to genotype-phenotype referred the breaking”. has here symmetry also “symmetry” “induced that — The systems physics explaining in biological for origin in its alternative order of third origin a op- the discussing present- other by and a put selection. onto vis ing a to recombination, vis attempt footing and democratic mutation will more as I such paper erators, short this In time at population the ee mdfiigagntcoeao ob n op- any be to operator genetic order. a of eration defining sources am potential I not as Here, discounted or gen- have whether been selection of erally than other issue evolution. operators neutral genetic the from Thus, raised benefit have can evolution 5] oth- adaptive though 4, drift, statement [3, genetic no of ers for value makes adaptive [2], guise any theory traditional about Neutral its The in instance, recombination effects. and “disordering” mutation and as selection agent ordering view an to as been question. has vexed tendency more the much Traditionally, a is predom- other one the systems over which inates for However, selection. to H uhthat such P ( t t . )= 1) + HP ( t ,where ), P ( t is ) 2 Order, Symmetry and Symmetry fore f(Ci)= f(Cj ), one sees that P (Ci,t)/P (Cj ,t)= Breaking constant, ∀t. We can in fact take this to be the defin- ing characteristic of the symmetry: that for Cg ⊂ G I will not go into detail about a precise definition of where φ(Cg ) = Cq, Cq being a given phenotype, “order”. For the purposes of this paper its most salient P (Ci,t)/P (Cj ,t) = constant, ∀t, and ∀ Ci, Cj ∈ Cg. characteristic is the following: that for a dynamical How may this symmetry be broken? In a finite gene system with state space G of dimension DG for late pool the symmetry will be broken spontaneously by times the system occupies a subspace U ⊂ G of di- stochastic effects. This can be understood in several mension DU ≪ DG. Thus, the more ordered a system ways, e.g. via the theory of branching processes [10] is the smaller the subspace into which it dynamically or using Kimura’s difusion approximation [2]. To lend evolves. a term from physics, such “spontaneous symmetry Intuitively, it is clear that selection will induce order breaking” lies at the heart of Kauffman’s ideas about in this sense. For example, in the presence of pure se- the origin of order. Thus, even in the absence of selec- lection an entire population will eventually order itself tion a system can dynamically evolve to a smaller sub- around the optimum present in the initial population. space, i.e. spontaneous symmetry breaking can lead to The dynamical attractor in this case is typically of di- an increase in order. mension zero. In the presence of mutation, such as in I will now turn to another form of symmetry breaking the Eigen model [8], the quasi-species represents the by considering the effect of the other genetic operators dynamical attractor. i.e. if one starts with a disor- besides selection defining dered random state then the effect of selection is to arrive at a more ordered state — the quasi-species. P (Ci,t +1)= As is well known, however, for a large class of fitness H({f(Cj )}, {pk}, {P (Cj,t)},t)P (Ci,t) (2) landscapes there exists a critical mutation rate above which there is no dynamical reduction onto a smaller where H is an operator that depends on the fitness dimension attractor, i.e. selection has its limits. landscape, {f(Cj )}, the probabilities, {pk}, to imple- ment the various genetic operators and on the pop- However, we must first ask what does selection mean? ulation composition {P (C ,t)}. I assume that one Selection can be most precisely thought of in terms of j can write H ≡ Hs + Ho, where Hs is the part of fitness and the corresponding notion of a fitness land- + the evolution operator associated with pure selection scape [9]. Fitness, fQ :−→ R , is most naturally de- and Ho contains the effect of the other genetic op- fined on the space of phenotypes, Q. In conjunction erators. The landscape symmetry will thus be pre- with the genotype-phenotype map, φ : G −→ Q, where served by the action of the other genetic operators if G is the space of genotypes, one may define an induced HoP (Ci,t)= HoP (Cj ,t) ∀t, and ∀Ci, Cj ∈ Cg. If this fitness function on the space of genotypes, fG = fQ ◦φ. condition is not satisfied we will say that the symmetry As the genotype-phenotype map is more often than has been broken by the action of the other genetic op- not non-injective (many-to-one) the function fG will erators; instead of a spontaneous symmetry breaking be degenerate, many genotypes corresponding to the there is an “induced” symmetry breaking. same fitness value. Thus, fitness defines an equivalence relation on G, many genotypes being equivalent selec- As a quantitative measure of this symmetry breaking tively. A simple example of this would be the stan- we will use the concept of “effective” fitness, defined dard synonym “symmetry” of the genetic code. I will via [11, 12] therefore refer to the equivalence of a set of genotypes feff (C ,t) under the action of selection (i.e. they’re all equally i P (Ci,t +1)= ¯ P (Ci,t) (3) fit) as a symmetry. Obviously, by definition, selection f(t) preserves this symmetry. One can see this explicitly, One may think of the effective fitness as represent- assuming proportional selection as a concrete exam- ing the effect of all genetic operators in a single “se- ple, from the evolution equation for the probability of lection” factor. Hence, if only pure selection was al- finding a genotype Ci lowed feff (Ci,t) would represent the fitness value at time t required to increase or decrease P (C ,t) by f(Ci) i P (Ci,t +1)= P (Ci,t) (1) the same amount as an evolution involving all the f¯(t) genetic operators and with selective fitness f(Ci). If where f¯(t) is the average population fitness. Consider- feff (Ci,t) > f(Ci,t) then the effect of the genetic op- ing the same equation for a genotype Cj , where Ci and erators other than selection is to enhance the number Cj both correspond to the same phenotype and there- present of genotype Ci relative to the number found in the absence of those operators. The converse is true of random mutations. This is the phenomenon of or- when feff (Ci,t)
0.005 the genetic operators of mutation and recombination themselves break the symmetry. The effective fitness 0 measures the strength of this induced symmetry break-
M(l) ing is. -0.005 As a specific example, consider a time dependent land-
-0.01 scape defined on 6-bit chromosomes that code the in- tegers between 0 and 63, where the initial landscape -0.015 has a global optimum situated at 10 and 11 and a lo- cal optimum at 40 and 41. However, after 60% of the -0.02 0 5 10 15 20 25 30 population reaches the global optimum the landscape generation is suddenly changed to a new landscape wherein the original global optimum is now only a local optimum. Figure 1: Graph of M(l) versus t in unitation model The original local optimum at 40 and 41 remains the with pc = 1. same but with a higher fitness value than the new local optimum at 10 and 11 and furthermore a new global optimum appears at 63. I will denote this landscape the “jumper” landscape. Figure 2 shows the results of 4 Numerical Examples an experiment where the mutation and crossover prob- abilities were coded in the chromosomes, either with In the previous section I used some very simple three or eight bits to codify each probability. Tourna- tractable models to illustrate the phenomenon of in- ment selection of size 5 was used and a lower bound of duced symmetry breaking. In this section I will 0.005 for mutation imposed. The success of the self- present some more non-trivial examples. For more de- adapting system in converging to the time dependent tails I refer the reader to the original articles. global optimum was compared to that of an “optimal” i) Self-Adaptation: It is well known that mutation fixed parameter system with p =0.01 and pc =0.8. and recombination rates are not uniform throughout The upper curves show the relative frequencies of the a structure such as a protein. One may well wonder optima using 8-bit and 3-bit codification and also what why certain values are found rather than others and if happens when pc = 0 and only the mutation rate is or not there is any adaptive value in it. In fact, in the coded. There are several notable features: first of all, case of the HIV virus it can be shown that preference the optimal fixed parameter system was incapable of for non-synonymous mutations in the neutralization finding the new optimum whereas the coded system epitope of the virus is directly due to an induced sym- had no such problem. For the case pc = 0 the curve metry breaking [13]. 40, 41 shows the relative frequency of the strings asso- Normally one thinks of the mutation and recombi- ciated with the optimum at 40 and 41. Before the land- nation rates as exogeneous parameters. However, if scape “jump” this optimum is local being less fit than one considers a system where they are coded in the the global optimum at 10 and 11. After the “jump” chromosome, but are not directly selected for, then it is fitter but less fit than the new global optimum 63 one has a completely autonomous system wherein one which is an isolated point. may examine whether the mutation and recombina- concentrated on this local optimum and starts to cool 1.0 down again only to find that this is not the global op- CR-Mut8b 40,41 0.75 CR-3b timum, whereupon the system heats up again to aid CR-8b the removal of the population to the true global opti- 0.5 mum. It is clear that there is a small delay between
0.25 the population changes and changes in the mutation rate. This is only to be expected given that there is 1-0 no direct selective advantage in a given generation for a particular mutation rate. The selective advantage 0.75 Cros 8b Cros 3b Average Rates / CR of a more mutable genotype over a less mutable one 0.5 can only come about via a feedback mechanism. It is precisly this feedback process that is described and 0.25 8b-Mut Mut 3b measured by the effective fitness function. Mut 8b 0 0 5 10 15 20 25 30 35 40 45 50 The average mutation rate also grows due to another Generation effect which is that the new optimum is more likely to be reached by strings with high mutation rates which Figure 2: Graph of relative concentration of the global then grow strongly due to their selective advantage. optimum (CR) (upper graph) and average crossover Thus, high p strings will naturally dominate the early and mutation probabilities (lower graph) as a func- evolution of the global optimum. After finding the op- tion of time for the “jumper” landscape. CR-3b and timum however it will become disadvantageous to have CR-8b are the results for 3-bit and 8-bit encoded al- a high mutation rate hence low mutation strings will gorithms. CR-Mut8b is the result for coded mutation begin to dominate. Induced symmetry breaking here is with pc = 0, with 40, 41 being the relative concentra- once again manifest in a most striking way. Although tion of strings associated with the local optimum at there is no direct selective benefit to differently coded 40 and 41. Mut 3b, Mut 8b, Cross 3b and Cross 8b strings their ability to produce offspring that can adapt are the average mutation and crossover probabilities to the changing landscape is very different. in 3-bit and 8-bit representations. The solid line for ii) Neuro-genetic models: In this case an analysis was Mut8b is the average mutation rate in the case pc = 0. made [15] of the population dynamics of a variant of Kitano’s neurogenetic model [16, 17] wherein the One thus sees that the global optimum was found in chromosome encodes the rules for cellular division and a two-step process after the landscape change. First the phenotype is a 16-cell organism interpreted as a the strings started finding the optima 40, 41 before connectivity matrix for a feedforward neural network. moving onto the true global optimum, 63. Immedi- Specifically, an artificial ecological environment was ately after the jump the effective population of the studied which consists of a single species composed of new global optimum is essentially zero. The number neural networks as individuals. Every chromosome, or of strings associated with 40 and 41 first starts to grow genotype, is used to produce a particular architecture substantially at the expense of 10 and 11 strings. At its for a feedforward NN that consists of 12 input neu- maximum the number of optimum strings is still very rons, 4 hidden and 1 output neuron — the phenotype. low, however, very soon thereafter the algorithm man- A genetic algorithm is then applied to the chromo- ages to find the optimum string which then increases somes present in the population at each epoch which very rapidly at the expense of the rest. The striking induces a search of the connectivity matrix space de- result here can be seen by comparing the changes in termined by the structure of the NN. Environmental the relative frequencies with the changes in the average effects are included in the fitness function that mea- sures the learning capacity of a particular individual. mutation rate, especially in the case pc = 0. Clearly they are highly correlated. First, while the population A chromosome consists of eight blocks of four genes is ordering itself around the original optimum, there each one of which is a three bit structure. The blocks is an effective selection against high mutation rates as themselves are labelled from a to h. The reproduction one can see by the steady decay of the average muta- process always begins with block a. Thus the first four tion rate. After the jump there is a noticeable increase genes have a priviliged role as they label the cells that in the mutation probability as the system now has to are going to be reproduced in the second step of re- try to find fitter strings. As the global optimum is production. As an example consider the chromosome an isolated state it is much easier to find fit strings baea.dcaa.defa.becd.aaea.aafh.haec.fgaa. The two step associated with 41 and 40. The population is now reproduction process specified by this chromosome can tations. A genotype is a cellular automata with binary be written elements which gives rise to a giraffe neck size, i.e. a d c ba phenotype, given by the number of automata elements b a a a e a that are “switched on” at the fixed point (steady state) a −→ −→ e a aa b a of the automata dynamics. As there are many differ- e ae a ent automata that can evolve to the same fixed point the genotype-phenotype mapping is highly degenerate. 011010001000 One “master” gene in particular plays a special role as 000000100000 ↔ (6) it governs the way in which the Boolean rules used in 000000001000 the evolution mutate. 100000100000 Thus the first block, baea, codes for the division of the Each member of the population is selected for the next original cell a into four cells. The first of these cells, generation with probability Pi = fi/ j fj, where fi b, then divides into four more which form the upper is the fitness of phenotype i. Initially,P there are ample left quadrant, dcaa, of the matrix. The second cell, a, resources available from both small and large trees, maps block a of the chromosome into the upper right the only selective criterion being that giraffes prefer quadrant etc. Finally, one constructs the connectivity to choose a mate among those that have similar neck matrix by reading left to right, row by row. Thus a 1 size. This “social pressure” landscape is modelled by specifies a connection between an input neuron and a defining the fitness of the ith giraffe to be a function hidden neuron and a 0 its absence. of its neck size ni and the average neck size of the pop- ulation, hni, with value one if hni− δ 190 180 In this contribution I have tried to briefly lay out the 170 case for induced symmetry breaking as an origin of or- 160 der in biological systems. Without doubt it exists, as 150 has been conclusively demonstrated. It is possible to average neck size see it at work in simple analytic one and two locus pop- 140 ulation models, and also numerically in several much 130 more non-trivial examples of artificial life system as I have briefly touched upon here. The extent to which it 120 exists in real biological systems is a question for future 110 0 200 400 600 800 1000 1200 1400 research. The chief difficulty in applying these ideas number of generations to the latter is that it is very difficult to assure one- self that apparent selection for a particular genotype Figure 3: Graph of average giraffe neck size (in ar- is due to an effective selection, via a symmetry break- bitrary units) as a function of time for a population ing effect, and not via some direct, yet unobserved, of 1000 giraffes. The drought starts at generation 240 selective factor. For this reason I believe it is well and lasts for 30 generations. Subsequent neck growth worthwhile continuing with the examination of math- lasts for another 1000 generations. The parameter val- ematical models of increasing complexity wherein one ues used were: µ = 0.0025, ν = 10−6, δ = 2.0 and may better control the fitness landscape and the na- ǫ =1.0. ture of the genotype-phenotype map, and also to con- sider artificial life systems where there is much more control over selective factors. neck size grows very quickly. After it ends it continues One might enquire as to why bother introducing the to grow, albeit more slowly, for a substantial amount concepts of effective fitness and induced symmetry of time until a steady state is reached. These results breaking. There are several reasons: first of all they can be explained quite simply: In the period before allow one in a quantitative sense to understand the the draught, and before expression, type one chromo- different mechanisms by which order may arise in bi- somes increase due to the effect of neutral drift. After ological systems. Secondly, they provide a framework expression they are effectively selected against due to within which neutral evolution and natural selection their tendency to produce giraffes with longer necks can be understood as different sides of the same coin, that pass outside the tolerance threshold and there- and in particular under what circumstances neutral fore cannot reproduce. Thus, before the drought the mutations may lead to adaptive changes. Thirdly, in- effective fitness of type one chromosomes is low. How- duced symmetry breaking may well lead to more ro- ever, due to the effect of mutations type one chromo- bust adaptive systems. It is no good having an ex- somes are not eliminated totally but constitute about tremely fit phenotype if when subjected to mutation 1−5% of the total population. After the draught starts at the genotypic level it typically mutates into an unfit the effective fitness of type one chromosomes increases phenotype. Rather one requires that an organism not substantially, given that they lead to giraffes of longer only be fit but that it gives rise to fit offspring which necks. The result is that the population becomes dom- in their turn give rise to fit offspring etc. Induced inated by type one chromosomes, with a small fraction symmetry breaking can pick out precisely those evolu- of type zero remaining due to the effects of mutation. tionary pathways that possess this type of robustness After the end of the drought as type one chromosomes as is found in the neurogenetic model of section 4. tend to produce longer necks the average neck size in- To what extent the different possible sources of order creases until a steady state is reached and it cannot predominate will depend very much on the landscape grow anymore. considered and is as open to debate as the standard In the giraffe model there is absolutely no direct selec- selectionist/neutralist argument. I believe that arti- tive difference between type one and type zero chro- ficial life research can play an important role in this mosomes. The only advantage of one versus the other debate by examining the generic properties of land- is in how they produce well adapted offspring, a quan- scapes and populations that admit as dominant one titative measure of this being the effective fitness. source of order versus another. Acknowledgements [12] C.R. Stephens and H. Waelbroeck, Effective De- grees of Freedom of Genetic Algorithms and the This work was partially supported through DGAPA Block Hypothesis, Proceedings of the Sixth Inter- grant number IN105197. The basic idea of induced national Conference on Genetic Algorithms, Mor- symmetry breaking was developed in collaboration gan Kaufman, San Francisco (1997). with Henri Waelbroeck to whom the author is grateful for many stimulating conversations. [13] C.R. Stephens and H. Waelbroeck, Codon Bias and Mutability in HIV Sequences . National Uni- versity of Mexico Preprint ICN-UNAM-97-09, (adap-org/9707), to be published in the Journal References of Molecular Evolution, (1997). [14] C.R. Stephens, I. Garc´ıaOlmedo, J. Mora Vargas [1] S. A. Kauffman, The Origins of Order, Oxford and H. Waelbroeck, Self-Adaptation in Evolving University Press, Oxford (1993). Systems, Artificial Life 4.2, 183-201 (1998). [2] M. Kimura, The Neutral Theory of Molecular [15] O. Angeles, H. Waelbroeck and C.R. Stephens, Evolution, Cambridge University Press, Cam- Emergence of Algorithmic Language in Genetic bridge (1983). Systems, Biosystems 47, 129-147 (1998). [3] S. Wright, Letter to M. Kimura, see W.B. [16] H. Kitano Designing Neural Networks Using Ge- Provine, Sewall Wright and evolutionary biology, netic Algorithms with a Graph Generation Sys- University of Chicago Press, Chicago (1986). tem, Complex Syst. 4: 461-476 (1990). [4] J. Maynard-Smith, Natural Selection and the con- [17] H. Kitano Neurogenetic Learning: an Integrated cept of a protein space, Nature 255, pp 563-564 Method of Designing and Training Neural Net- (1970). works Using Genetic Algorithms, Physica D75: 225-238 (1994). [5] M.A. Huynen, Exploring phenotype space through neutral evolution, J. Mol. Evol. 43, pp [18] J. Mora, H. Waelbroeck, C.R. Stephens and 165-169 (1996). F. Zertuche Symmetry Breaking and Adapta- tion: Evidence From a Simple Toy Model of a [6] L.L. Gatlin, Information Theory and the Living Viral Neutralization Epitope, National Univer- System, Columbia University, New York Press, sity of Mexico Preprint ICN-UNAM-97-10 (adap- N.Y. (1972). org/0797) to be published in Biosystems (1997). [7] V.A. Ratner, Molecular Genetics, Principles and Mechanism, (in Russian). Nauka Novosibirsk (1983). [8] M. Eigen Self-Organization of Matter and the Evolution of Biological Macromolecules, Natur- wissenschaften 58: 465 (1971). [9] S. Wright, The roles of mutation, inbreeding, crossbreeding and selection in evolution, In: Jones DF (ed) Proceedings of the sixth international congress on genetics, vol 1, pp356-366, Brooklyn Botanic Garden, New York (1932). [10] Z. Taib, Branching Processes and Neutral Evolu- tion, Springer Verlag, Berlin (1992). [11] C.R. Stephens and H. Waelbroeck, Analysis of the Effective Degrees of Freedom in Genetic Al- gorithms, Physical Review E57, pp 3251-3264 (1998).