Evaluation of neo-Darwinian Theory using the Avida Platform. Part 1

Avida is a computer software platform designed to test neo-Darwinian scenarios including speciation, evolution of novel complex functions, and genetic drift. A large number of parameter settings are available to the researcher. We evaluate here published tests and intrinsic Avida characteristics in an attempt to extrapolate towards real biological organisms. Computer based models and simulations must capture enough realism to facilitate valid decisions. Neglect of key factors or unrealistic parameter settings permit conclusions to be claimed which merely reflect what the decision maker intended a priori. The long-term effect on genetic systems of random insertions, deletions and point mutations over deep time may be degradation of function and specificity; or the generation of true biological novelty. Competing factors are involved and only detailed quantitative analysis with real world experiments can firm up our opinions in this matter. In Part 1 we examine some of these issues and in Part 2 explore some more quantitative details based on Avida runs. It was found in this two part series that reports which demonstrate rapid increase in complex, genetically coded novel functions are an artefact intrinsic to how the Avida computer runs are planned and some neglected constraints found in nature. Using biologically realistic parameter settings and including neglected factors leads to the opposite view, both theoretically and when using Avida experiments. Mutations plus natural selection in rapidly reproducing, asexual, small genomes fails to explain the origin of ever increasing cellular functionality, including molecular machines. Introduction Avida was initiated at Caltech in 1993 <2> inspired by T. S. Ray’s tierra program. In simplified terms, strings of symbols representing any of usually 26 <4> (sometimes 28 ) computer logic instructions (shift-l, inc, add ...) are mutated and undergo Darwinian competition. Runs are initiated with a single organism endowed with the necessary instructions to self-reproduce and extra no-function instructions which can also be mutated. The original genome replicates and builds up to a stable population usually set to 3600 organisms. Detailed explanations of the research platform are available <2>, . Criteria for selecting favored members are established via parameter files environment.cfg and genesis, under stochastic conditions. Data to be recorded can be defined in the events.cfg file. The definition of evolution. Dr. Lenski and Dr. Adami are leading figures in the Avida artificial life project. They quote philosopher of science Daniel Dennett that, “...evolution will occur whenever three conditions are met: replication, variation (mutation), and differential fitness (competition)”. PCID 3.1.1, November 2004 © Royal Truman and ISCID 2004 At face value, this appears to be tautologically true, but misleads since “evolution” is a word with many connotations including change, and development to better, more complex replicators. Let us rephrase. Assume three conditions hold: (i) A population can replicate for many generations (ii) the copies are not all identical with the original (iii) a criterion establishes preferential selection of one lineage What results? Change, yes. More sophisticated replicators? Not inevitably. Suppose the original state was optimal at some point in time and all subsequent variation degrading. Differential fitness would select the less damaged on average, but the long-term outcome will still be an inferior population. Dennett’s criteria do not ensure neo-Darwinian processes result in the complex biological structures we observe. Using Avida to test neo-Darwinian notions. Leading Avida-based researchers believe their digital experiments are analogous to cellular evolution. Adami and colleagues explain, “Computations can be carried out by evolved programs if they develop sequences of code (“genes”) that perform logical, bitwise, computations on random numbers provided in their environment. Such genes evolve spontaneously if the performance is rewarded with bonus CPU cycles (their unit of energy). The complexity of the environment can be controlled by changing the number of logical operations whose performance is rewarded.” He states elsewhere, “Because it is these computations which provide the organism with the “energy" (in the form of CPU time) it needs to replicate, we can think of this computational code as the genes that code for the organism's metabolism. To this extent, we are able to observe the emergence of metabolic genes in self-replicating organisms, and thus the evolution of complexity.” Then to Dennett’s list additional criteria must be satisfied if “evolution” is to mean the grand conceptual scheme leading from simple molecules to living multicellular organisms: (iv) the replicators must not all self-destruct after a few generations (v) differential fitness must not be accomplished by simplification or net degradation of existing biological functions (e.g., parasites, which do reproduce faster than the progenitor) (vi) opportunities to attain more complex fitness states must be feasible: the organisms are not trapped in local fitness peaks, and irreducible complexity does not pose statistically prohibitive barriers. Avida run details. The Avida computer instructions are generally analogous to codons or amino acids. Any of 26 different letters are present at each position of the digital sequence, reminiscent of the 21 (including ‘Stop’) in the genetic code. A mutation is assumed to be equally likely to change any of the 26 possibilities to any other. In a recent paper in Nature, many combinations of instructions (“codons”) in a particular environment can produce 9 different logical functions (“genes” or “operons”), e.g. Table 1. A single initial organism with a genome of 50 instructions reproduces to a final fixed population of 3600 members. Several one letter symbols at the beginning and end of the genome, 15 positions in total, contain all instructions <13> required for self-replication: qucavc...utzcarvab [1] The remaining 35 positions are originally identical and have no use, thus serving as raw material to evolve novel functions. Default settings in the environment.cfg file include: point mutations lead to copying errors at a rate of 0.0025 per instruction copied; singleinstruction deletions and insertions occur with 0.05 probability each per genome copied (multiple mutations can occur in a single generation). The organisms compete conceptually for energy, which is provided in quanta called SIPs. One SIP suffices to execute one instruction. “In Avida, organisms can acquire energy by two mechanisms. First, each organism receives SIPs in proportion to its genome length. Second, an organism can obtain further SIPs by performing oneand two-input logic operations on 32-bit strings” p. . Digital “organisms” evolving new logic functions are rewarded in proportion to the complexity of the derived function. Calibrations necessary to ensure biological realism. Mutations in the symbols of [1] can produce logic functions (And, Nor...) when tested against numeric strings. The proportion of digital genomes possessing such patterns builds up rapidly for reasons which will be discussed in this two part series. Simple logic functions are statistically easier to obtain, and once significant amounts of these exist, provide the basis for evolution to the more demanding logic functions. The runs reported have not been calibrated against real organisms nor validated with biological experiments. The necessary sensitivity analysis tests were not performed. We conclude that the simulations “work” because of several judicious assumptions and parameter settings which ensure this outcome a priori: (1) Miniscule genomes are assumed with far more superfluous than necessary genetic material • The ratio of indispensible / total genetic material is initially 15 / 50 • Increasing the genome size with worthless genetic material is compensated with extra metabolic energy instead of penalized, as should be the case in nature. (2) Extremely high mutational rates are used • In default settings point mutations occur at a rate of 0.0025 instructions per instruction position per generation; deletions and insertions have a probability of 0.05 per genome copied. This permits generation of logic functions before genome truncation would remove superfluous genetic material. (3) Critically important functions for “life” are difficult to destroy by mutations • All functions necessary for an organism to replicate and survive are carried out initially by only 15 instructions. The physical machinery to transcribe, translate and perform metabolic processes are not coded genetically and thus not subject to mutational damage. (4) Statistically insignificant sequence space distances are assumed between novel, more complex functions • This is an artefact of logic functions and not protein sequence space. (5) Dramatic rewards are provided for new logic functions • New, more complex functions, like EQU, once formed are essentially guaranteed to perpetuate, i.e., fix in the population due to very high relative fitness rewards. (6) No path for graceful degradation of functions is provided. • Less effective protein variants which are still functional are more numerous than the more specific ones, and often decrease fitness but minimally. With logic functions no such variants exist, which permits “natural selection” to easily identify degraded functionality with 100% accuracy. Miniscule genomes with much superfluous genetic material. What kinds of biological organisms might correspond to Avida digital ones? The generation of new genetic information by the neo-Darwinian mechanisms under discussion can only be relevant for very small, asexual, free-living organisms, with short generations times. In the absence of mutations, Avida lineages as usually generated will never go extinct. In biological terms, the organisms are assumed to possess significant repair mechanisms and to execute all biochemical processess effectively to survive. Clearly a 56 base pair (15 X 3 X 26/21) stretch of nucleic acid cannot represent any real living organism (the extra 35 nooperation codons perform no services). The smallest living organism known is the parasite Mycoplasma genitalium with a genome of ca. 578,000 bp, and a mutational rate of about 5.9X10. Gene knockout studies <18> in a nutured laboratory setting suggests a mutant might limp along for a few generations on perhaps 250 to 400 genes (outside the laboratory 500 to 600 genes is considered more likely). A theoretical lower limit of about (250 genes) X (1000 bp/gene) implies about 250,000 bp. Nevertheless, for computation convenience generous parameter settings could be used, but the results must be followed by extrapolations to real world constraints. Our analysis must be limited to very small, self-contained asexual organisms able to replicate almost flawlessly: the smallest of bacteria. Large amounts of additional genetic material must be made available if new logic functions are to be coded for. Indeed, “The 23 pivotal genomes ranged in length from 49 to 356 instructions; the ancestor had 50 instructions and so there was a strong tendency for increased length to precede the origin of EQU.” <6>, p. l42 In nature, brand new metabolic networks, which the logic functions are analogous to, don’t evolve into existence every few mutations. A key but unstated assumption, is that the unneeded genetic material would have to remain present for a very large number of generations. This crucial requirement would not be fulfilled in nature since streamlined variants of very small genomes out-reproduce those with unnecessary coding material, as required for evolutionary experiments. For example, the bacteriophage Qβ parasite genome uses a single strand of RNA with 4,500 nucleotides. Sol Spiegelman placed Qβ, a necessary replicase and certain salts in a test tube. Portions of the genome dispensible in this environment were lost, permitting the descendants to reproduce more quickly. After about 70 generations everything superfluous had been eliminated, and a single species with a single RNA strand 550 nucleotides long remained. Only 70 generations were enough to remove almost 88% superfluous genetic material, which incidently is far more than the 35/50 = 70% initially made available in the major paper under consideration. In a second experiment, RNA already shortened optimally was allowed to replicate in the presence of a drug that slowed down replication. The drug attached to a specific three nucleotide sequence. After a few generations all members possessed the same new sequence, in which the recognizer address for the drug had mutated to 3 other nucleotides, preventing drug binding and permitting faster reproduction. Other studies <23> show that currrently unnecessary DNA added to genomes tends to be removed quickly, sometimes due to the loss of a large number of genes. For realistic Avida simulations, generation times must be inversely proportional to amount of superfluous genetic material, ceteris paribus, at least until an advantageous logic function evolved. The authors permitted deletions, but fail to provide any reward when they occur: generation times would be shortened; and less (biological) energy and material (nucleic and amino acids to produce proteins) would be needed. Some Avida researchers seem to be aware of this fact, as further discussed in Part 2, but a parameter is not provided in the genesis file to include this effect during simulations. The proportion of members possessing a genetic change, given budding or binary fission reproduction and constant population size, can be modelled mathematically : where x is the fraction of a population possessing a mutation; s is the selectivity factor; and the unit of time, t, is the generation interval. Let’s approximate the selectivity coefficient, s, by assuming the offspring with one codon less than the original 50 would replicate in about 49/50 the amount of time (and lineages with an insertion would suffer 51/50 times longer generation times). Since initial sequence [1] does nothing else than replicate, this is reasonable. The favorable selectivity assumption is also consistent with other biological advantages: less energy and material requirements, and less dangerous, interfering polypeptides. About 3.5% of the 3600 organisms would receive a harmless deletion during just the first generation. Subsequent generations will offer additional opportunities to remove more ‘junk’. Many experiments could be run with the Avida software and some may indeed undergo slower genome truncation on average, but for very large populations the Central Mean Theorem predicts such deviation will be small. Assume just 1% of one of the lineages had at some point but one fewer “no-operation instruction”. Using equation [2] produces the results shown in Figure 1. On average, only 576 more generations shifts the population from 1% to 99.9% possessing the shorter genome. This result is independent of population size. The population would actually initially have a distribution of members with some having several more and others several fewer unnecessary codons. This broader spread merely strengthens our argument. x0e 1 + x0[e -1] x = [2] 1% genomes having shed two codons would on average require 288 generations for 99.9% of the member of any population size to have streamlined their genome, see Figure 2. Before 1% of the population has obtained a 49 instruction genome, one member on average will undergo yet another harmless deletion (0.01 X 3600 X 0.05 X 34/49 > 1 deletion). This new lineage has a yet greater advantage, leading to a snowball effect. The authors assume mutational additions can also occur. Note that additions after the final instruction to divide would not be incorporated in the next generation. At some point a distribution of genome lengths would be present. Figure 3 illustrates how rapidly the population would shift from 1% to >99.9% members with 10 fewer codons: in just 70 generations. Of course, the selectivity estimates, s, might be too large. We can use a very modest s = 0.001 for the member having shedded several percent junk coding material. Then on average within 11,512 generations the population would shift from 1% to >99.9% streamlined genome, see Figure 4 (ignoring the mentioned compounding effect, which would shorten the generations considerably). For a proto-bacteria of generation time 10 minutes, this means far less than 80 days, for any population size involved. Figure 2. Population fraction having 4% (2/50) fewer codons. (Fission reproduction model, initially 99% have the larger genome, selectivity s=(50-48)/50 = 0.04 Figure 1. Population fraction having 2% (1/50) fewer superfluous codons. (Fission reproduction model, initially 99% have the larger genome, selectivity s = (50-49)/50 = 0.02) Figure 4. Population fraction having fewer codons, assuming selectivity s = 0.001 Figure 3. Population fraction having 20% (5040)/50 fewer codons. selectivity s = 0.2 This demonstrates that we cannot assume a stable starting population with genomes possessing much unnecessary DNA, and expect natural selection to ignore this. Extremely high mutational rates Having demonstrated how quickly small genomes eliminate superfluous DNA, we must consider the mutational rates as typically used in Avida runs. Low values provide much time and many generations for the streamlining process to take place. Such lineages, which reproduce more quickly, would take over the population before opportunities to evolve new functions became available. High rates of mutation and small sequence distances between logic function are typically assumed, to permit higher development. These also facilitate escape from local fitness traps: “The presence of deleterious mutations along the line of descent is more surprising. Fifteen of the 18 deleterious mutations reduced fitness by <3% relative to the parent... However, two mutations reduced fitness by >50%.” p. 141 This lineage resulted in the dominant genotype. A primitive, free living (non-parasite) bacteria <1/5 the size of E. coli at some point during the evolutionary scenario would have about 1000 coding genes, each averaging 300 codons plus regulatory regions, all critical to life, with little worthless DNA overhead. The genome must code for the translation machinery (ribosomes, DNA polymerase, transcription factors); tools to transcribe DNA to RNA (RNA polymerase); DNA coding for enzymes to permit metabolic reactions; tRNA and accompanying enzymes; and the equipment to duplicate the chromosome during fission. Based on typical (default) parameter settings for Avida runs p. , about 1000 X 300 X 0.0025 = 750 point mutations, plus (2)(0.05)(1000*300)(/50)(21/26) = 485 indels (insertions or deletions) would occur per genome every generation when extrapolated to a ca. 1000 gene organism. The latter would produce many reading frameshifts in every individual virtually every generation. These model assumptions would lead unavoidably to error catastrophe in biological organisms. Error correction machines must also be coded for to prevent this, but this poses addition questions: how would a miniscule genome survive until these evolved; and these additional genes would also be subject to damage from mutations. Nevertheless, the authors wrote, “Hence, in the ancestral genome of length 50, 0.225 mutations are expected, on average, per replication. Various organisms from nature have genomic mutation rates higher or lower than this values.” p. 143 However, no free living organisms have genomes on the order of 50 codons with such mutational rates, and the values need to be scaled to real biological sized free-living organisms to test for reasonableness. From the reference cited by the authors, the simplest kinds of single cell free living organisms, such as E. coli (genome ca. 4.6X10 bp base pair -) and S. cerevisiae (genome ca. 1.2X10 bp) have mutation rates on the order or 5 X 10 per base pair each generation, or about 0.002 to 0.006 total mutations per genome each generation. Properly scaled, we see that a vastly greater rate is being assumed by the authors for Avida runs. Literature estimates <28>, <29>, <30>, <31> for various living organisms (which exclude parasite virus) demonstrate that the mutational rates used in the simulations are not realistic. A much greater than 3600 member population size could tolerate a higher mutational rate with less risk of total extermination, but this exacerbates the preceding objection: fortuitious mutations would require more generations to fix, and thereby provide yet more opportunities to streamline away superfluous DNA. Functions critical for life are virtually impossible to destroy by mutations Biologically unrealistic mutation rates to not lead to complete population extinction in Avida runs due to a fact subtle enough to escape the attention of many specialists. Recall from above that the molecular machines needed for translation, transcription, metabolism and so on must be coded for genetically in real, biological free-living organisms. Adami and collegues claim that, “The Avida system controls populations of self-replicating computer programs in a complex and noisy environment within a computer’s memory.” <32> This is misleading. The only thing mutating are short strings of symbols such as [1] and not any computer programs. The software shell actually or figuratively performing critical functions (“replicate your genome from here to there”; “generate metabolic energy”; etc.) are written in C++ and are neither coded for by the genome (unlike real organisms) nor subject to mutation. The rest of the infrastructure is provided for free in the form of computer technology (stacks, registers, electricity, cables, chips, etc.), unlike real cells in which physical ribosomes, cell walls, ATP and so on must be genetically coded, and thus in the Avida platform cannot be damaged by mutations. The starting strings such as [1] usually soon increase in size under typical Avida settings (documented experimentally in Part 2). The minimal 15 instruction string, assumed to be necessary to support all life processes autonomously, is but one pattern from among many alternatives able to code for the cellular replication (Repl.) function (Table 2). (ID is the mutant lineage; the first column refers to the position in the genome on which the 1-letter instruction is found). Over 99.9% of the indispensible coding material needed for a minimally sized real organism to survive has been shielded from the possibility of any kind of mutation. These functions are provided by the external hardware and software. For real, compact genomes it is thousands of times more likely that a mutation would be deleterious than that a portion of the critical 15 instruction string would be affected, as implied by the Avida runs. Therefore, such Avida runs ensure a priori that a 3600 member population will almost always easily survive the mutational onslaught for endless generations even without the development of new logic functions. By generously rewarding instruction patterns which produce logic functions, the net effect of Avida type mutations over many runs is virtually guaranteed to ensure increased functional complexity over time (demonstrated in detail in Part 2). This is not some kind of law of nature, but an inevitable result of how the Avida system and runs are designed. Statistically small sequence space distances assumed Novel functions are easy to attain, compared to living organisms, and these offer immediate, dramatic selection advantages. For the dominant genome the authors report: “After the origin of EQU, another 233 steps occurred along the line of descent leading to the final dominant genotype. Of these, 62 were beneficial, 132 neutral and 39 deleterious.” p. 141 (underlines added). Within very few attempts many new logic functions developed, since so many useful mutations are available. In one example, “Forty-five of the steps increased overall fitness, 48 were neutral and 18 were deleterious relative to the immediate parent.” The logic functions are assumed to be close together in coding sequence space. How close is examined statistically in Part 2. In one report, “Of the 35 instructions required for EQU, 22 were needed for simpler functions. Three instructions required for EQU were also essential for replications; these were conserved from the ancestral sequence, as were five others.” p. 141 We now see why digital organisms manage to derive a series of ever more complex logic functions, culminating in the final EQU one. “The number of instructions required for EQU ranged from 17 to 43, with a median of 23 instructions” p. 142 . EQU was the most complex logic function to be evolved, and there is a large number of available intermediate plateaus leading up to it. When these intermediate stepping stones are removed, none of the organisms managed to evolve EQU p. . “Some readers might suggest that we ‘stacked the deck’ by studying the evolution of a complex feature that could be built on simpler functions that were also useful. However, that is precisely what evolutionary theory requires, and indeed, our experiments showed that the complex feature never evolved when simpler functions were not rewarded.“ p. 143 While this is indeed what current neoDarwinian Theory requires, the simulation cannot use biologically implausible distances to imply the theory has been validated. We can contrast this with the estimated proportion of cytochrome c sequences which would be minimally functional: <10 <25>, <26>, . This is a rather small and not too demanding protein. Until a probabilistic barrier of this rough order of magnitude has been bridged, no selection to fine-tune the sequence and fix this gene in a population would be possible. Let’s put things in perspective. Evolutionary theory must explain the origin of hundreds of new, unrelated protein families, based on different folding characteristics. Since increased complexity is under discussion, we are not interested in relatively trivial protein variants which are able to catalyze virtually identical molecules. Professor Sauer at MIT systematically deleted small pieces from viral proteins and inserted altered pieces back into the genes at the sites of the deletions to determine how much variation at various portions of the sequence can be allowed . He concluded that the likelihood of finding a folded protein by a random mutational search is about 1 in 10 for the examples studied. That is equivalent to guessing correctly one particular atom out of our whole galaxy. Clearly the Avida computer logic functions assume much too great a proportion of useful sequence variants, compared to biological random polypeptides <40> . A new domain, to be stably folded, should contain at least 50 amino acids. However, Table 3 illustrates the shortest sequences able to mimic each of the nine novel logic functions that do not depend on the initial content of stacks and registers. Three functions require only six instructions or “codons” although “nearly every major process in a cell is carried out by assemblies of 10 or more protein molecules” and not merely a portion of a single protein. As an example, 60 independent polypeptides form the pyruvate dehydrogenase complex, which does not function if one is missing. From a web site we find a few examples of different sequences from one lineage which produced the logical function NAND, in various manners, within 42 mutations (Table 4). The first column is the position at which the necessary 1-letter instruction is found, ID is the mutation number. In one lineage, the new and more valuable OR-N function was derived from the NAND function by a single point mutation at position 37, with no intermediate loss of NAND function, see Table 5. In the example in Table 6, NAND and AND functions already work. A new, more complex function, rewarded with a factor of 8, could arise from a single mutation at position 19. (Red boxes indicate the instruction is necessary for the logical function shown in green). No other instructions were necessary. Dramatic rewards are provided for new logic functions Each higher, more complex function type (often separated from simpler ones by a single mutation) are typically rewarded by accelerating the rate of replication of that new lineage by about a factor of two. Since the population was doing very well without the new function, this is unrealistic. A function such as EQU, with a computational merit factor of 32 (!) makes the probability very high that successful mutants will rapidly generate new lineages and fix in the population. The opportunity to degrade functions gracefully Avida logical functions either work with 100% effectiveness, and are rewarded richly, or have 0% effectiveness. This has considerable theoretical significance. On average, Avida-type lineages which lose these functions are very likely to be out-reproduced. Alternatively, lineages with intact functions reproduce between about 2 to 32 times faster and on average are very likely to dominate. In cells, the exact 3-dimensional topology of proteins, location, and their proportion during the lifetime are important. Compared to an optimal solution, there are statistically far more inferior possibilities. We can group these variants into categories, each less effective but more numerous than the preceding. As the population becomes more distant from the optimal, the next degradation step carries decreasing fitness penalties, as long as a necessary function is still performed. This argument can be extended to a large number of genes, resulting in a snowball effect: multiple degradations over many loci makes additional degradation harder to remove by natural selection. This fact demonstrates how easy it would be to slide slowly down a fitness slope in real biological organisms, as illustred in Figure 5.