|
Author
|
Topic: Royal Truman: Avida, a biologically unrealistic model for neo-Darwinian Theory
|
Rex Kerr
Member
Member # 632
|
posted 16. April 2004 03:22
Sometimes, however, it is a good idea to be selective with your hypothetical axioms to avoid being led down too many dead-end paths. We could suppose that CSI could be created, and we could suppose that f = m*a, and we could suppose that starting in 2005 all batteries will never run down and therefore supply power forever.
Some of these suppositions are backed by more evidence than others--either direct or indirect. f=m*a has been backed up pretty well. I'm going to wait to see an everlasting battery, though, before I change my plans for 2005.
IP: Logged
|
|
michaelgoodrich
Member
Member # 393
|
posted 16. April 2004 10:42
Kerr writes:
michaelgoodrich wonders whether Avida's lack of exposure of reproduction infrastructure invalidates it as a model of biological evolution.
In brief: no. Digital organisms in Avida do, actually, handle their own reproduction. That part of the code is very much subject to deleterious mutation.
-----------------------------------------
I am late in seeing this comment, sorry.
I do not think I understand this comment. Are you saying that Truman is incorrect?
Or do I have Truman incorrectly on this point?
In any case, how is "that part of the code" subject to the effects of deleterious mutations?
regards,
[ 16. April 2004, 10:43: Message edited by: michaelgoodrich ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 16. April 2004 12:34
Truman was complaining that he thought that more of the genome should be devoted to replication. The digital creatures in Avida have to execute instructions to copy themselves. It's not terribly hard--you can get it down to about 15 instructions (I forget exactly)--but then again, the organisms don't have to be terribly large (e.g. 22 instructions total). If a mutation affects the replication code, chances are the digital critter won't replicate.
Biological organisms also have to handle other essential tasks, most notably metabolism. The Avida critters don't--metabolism is assumed. If Avida critters had to use instructions to handle their own metabolism, it would provide yet more instructions that were vulnerable to mutation. But we've already got some, so there's no real reason to expect that the simulation results would differ greatly. If you die and leave no offspring because you can't metabolize or you die and leave no offspring because you can't reproduce, you're still dead and have no offspring. That's what's important, and that is something that Avida does include by making replication vulnerable. Adding metabolism, to a first approximation, would just slow the simulations down without adding any new information.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 16. April 2004 17:23
michaelgoodrich:
Truman complained that in the Lenski, et al, study, which used a 50-instruction Ancestral critter which had just 15 replication instructions and 35 "junk" instructions, in his view the proportion of the total instructions (15/50) that were replication code and were exposed to mutations was too small. But in the run I described earlier in this thread, avida critters rapidly shed the "junk" instructions before they begin to produce lineages capable of performing a logic function. That is the common pattern, not a unique case. I've seen it in every run I've made starting from a long Ancestral critter.
From my earlier posting: quote:
At the moment when the grid is just filled with 3,600 critters, the average length of the 3,600 critters' genotypes is 21.2 instructions and the length of the dominant (most frequent) genotype is 20 instructions. The average critter is executing just 17.5 instructions, indicating that the replication code in a number of lineages has already evolved to be tighter than the Ancestor's 18 instructions.
In other words, by the time the grid is filled, most of the critters in the avida world are just three or so instructions longer than the Ancestor's replication code (18 instructions in the starting critter I used). The critters are mostly replication code, and all the replication code is exposed to potentially lethal mutations. That disposes of two of Truman's complaints: (1) relatively little replication code exposed to mutation, and (2) lots of "junk" instructions for evolution to use to evolve critters capable of performing logic functions. Neither is the case.
RBH
[ 16. April 2004, 17:25: Message edited by: RBH ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. April 2004 02:00
A few more notes on Truman's original critique. He wrote quote:
(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 rewarded with extra metabolic energy
(2) Extremely high mutational rates are used
* Point mutations occur at a rate of 0.0025 different "instructions" ("codons" or "amino acids") per instruction position per generation; deletions and insertions have a probability of 0.05 per genome copied
(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
* New, more complex functions, like EQU, once formed are essentially guaranteed to perpetuate, i.e., "fix" in the population.
Most of these have been dealt with before, so I'll only briefly indicate why they are irrelevant. First, awarding SIPs in proportion to genome length renders length, in terms of number of instructions, selectively neutral. However, there is still selective pressure in favor of shorter genomes such that the population rapidly sheds the "worthless genetic material." Why that occurs is left as an exercise for the reader. (Note: It's not confined to the initial phase of a run. I've also seen it - shedding accumulated "junk" instructions - happen as far into a run as 150,000 updates when most lineages in the population were performing 6 or 8 or all 9 logic functions.) This means that Truman's point (3) is a non-starter; very soon after an avida run begins most of the length of the critters' instructions strings is associated with replication and is vulnerable to disruption by mutations. Truman's extended discussion of the benefits of shortening the genome is therefore irrelevant; in fact an avida population does that very rapidly when one starts a run with a single Ancestor of 50 instructions. Nevertheless, that run will also yield lineages that eventually evolve to perform complex functions and over generations expand their genomes in the face of selective pressure for shorter genomes.
Regarding the high mutation rate, one can do the same sort of experiment with lower mutation rates and find similar outcomes. It just takes longer. quote:
More complex functions, often separated from simpler ones by a single mutation, are rewarded by accelerating the rate of replication of that whole 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 (!) ensures such evolutionary progress will be almost impossible to lose later.
In fact, when an individual that performs a more complex function first appears in the population it almost never immediately fixes. While an individual that performs a logic function that its peers do not perform may have a substantial reproductive advantage, it is also a lone individual and is vulnerable to mutational disruption and/or being over-written by a neighbor before it has the opportunity to reproduce successfully. A reproductive advantage for one individual does not translate to automatic fixation. It's quite unexceptional to see individuals capable of performing a more complex function flicker in and out of the population, perhaps hanging on for a while at a low frequency for some time before their reproductive advantage finally carries them to a stable high frequency in the population. Even though they may have a substantial reproductive advantage, they are not invulnerable and fixation is not guaranteed.
At the end of his critique Truman offered six suggestions. quote:
1. Selectively favor deletions, leading to more efficient, streamlined genomes
2. Increase significantly the amount of DNA necessary to keep the organism alive and to reproduce
3. make new functions statistically more challenging to attain. This would capture the notion of also needing additional regulatory elements, new or modified protein domains, and so on
4. Significantly lower the mutation rates per instruction position each generation
5. Provide smaller rewards for new derived functions, such as s = 0.2
6. Decrease the initial proportion of superfluous DNA significantly.
(Numbering added)
So far I have run various pilot runs implementing four of those suggestions: #2 and #6 jointly, by reducing the "junk" instructions in the Ancestor as a proportion of genome length; #3 by eliminating various combinations (though not yet all combinations) of three of the intermediates (Lenski, et al. ran control conditions eliminating all single intermediates and all pairs of intermediates); and #4 by editing the appropriate control file. In each instance, lineages capable of performing each of the nine complex logic functions (except for those eliminated as controls in #3) evolved in one or more lineages. For the runs with lower mutation rates I increased the total updates to 500,000. #1 is already covered - all else being equal, there is at least a slight selective bias toward shorter genomes. So only #5 remains to be looked at. I suspect that it'll be innocuous, but it's worth testing and I'll get around to that soon.
I'll also note that enabling the capability for avida's version of horizontal gene transfer increases variability in the population and appears to raise the likelihood of evolving lineages capable of performing the various logic functions, particularly under variability-constraining circumstances, like low mutation rates. Since Truman is concerned with biological realism, though, allowing HGT in a population of digital organisms that reproduce by mutable self-copying seems appropriate.
Finally, I'm coming around to Rex Kerr's view that the instructions in avida are just that: discretized instructions, and that it's a mistake to try to identify them with a particular level of biological entity - gene, codon, base pair, or whatever. It's that they're discrete that allows them to be appropriate entities for heritable (and mutable) information representation in a generic evolutionary system.
Given that this thread has ramified some, I'd appreciate it if someone would let me know if there are criticisms from Truman (or anyone else) that haven't at least been addressed, if not answered - I'm still popping pain pills and don't really trust myself to read for content yet.
RBH
[ 19. April 2004, 03:09: Message edited by: RBH ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. April 2004 17:12
To take a fast look at the last suggestion that Truman made (#5, reduce the selective advantage for performing more complex logic functions) I freed up two fast machines last night and made three evolutionary runs with three different randomization seeds. The runs were of a maximum of 250,000 updates with the following fitness function [Function: My Computational Merit (Lenski Computational Merit)]:
quote:
NOT: 1.25 (2)
NAND: 1.5 (2)
AND: 2 (4)
OR_N: 2 (4)
OR: 2.5 (8)
AND_N: 2.5 (8)
NOR: 3 (16)
XOR 4 (16)
EQU 4 (32)
The Computational Merit values I used were assigned on no particular basis other than to have them at least slightly increasing for more complex functions. This fitness function is considerably flatter than that of Lenski, et al. I used the 50-instruction Ancestor described above and default values for mutation rates.
In the three runs, one run produced lineages capable of performing all 9 logic functions and one run evolved lineages capable of performing 7 of the 9 (all except XOR and EQU). The third run evolved no lineages at all. Recall that I start runs with a single Ancestral critter rather than with a matrix filled with identical Ancestors. The Ancestral critter in that third run incurred an immediate point mutation to its replication code that rendered it incapable of reproducing, so the population never left the initial state of one Ancestral critter.
The conclusion is that given some reproductive advantage that is sufficient to offset the selective bias toward shorter genomes (explaining the latter still being left as an exercise for the reader), the avida population does not require that extraordinarinarily large selective advantages be assigned to the more complex logic functions. The general result of the Lenski, et al., study are robust in the face of a range of mutation rates, in the face of varying topography of the fitness landscape, and in the face of the unavailability of a variety of combinations of specific 'intermediate' forms.
RBH
IP: Logged
|
|
Janitor@MIT
Member
Member # 125
|
posted 19. April 2004 17:19
The problem in investigating models is to make them real. By "real" I have very very high standards, which are not my own. They are "pre-specified" one might say. Evolution and design confront the same problem here: Make it real!
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. April 2004 21:44
J@MIT,
Even with the heightened emphasis on pain management in medicine, with the consequent availability of good drugs, that's too cryptic for me. As we all know, you don't elaborate, but would you generate some free associations to your post to kind of narrow down its reference?
TNX,
RBH
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 19. April 2004 22:46
Maybe he's asking for an all-atom molecular dynamics simulation of the entire biosphere for millions of years?
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 19. April 2004 23:00
Rex,
That'll have to wait until my former colleague who's a quantum information theorist gets a really big quantum computer up and running.
RBH
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 20. April 2004 17:23
RBH said,
quote:
Finally, I'm coming around to Rex Kerr's view that the instructions in avida are just that: discretized instructions, and that it's a mistake to try to identify them with a particular level of biological entity - gene, codon, base pair, or whatever. It's that they're discrete that allows them to be appropriate entities for heritable (and mutable) information representation in a generic evolutionary system.
If this is true, then we can't draw any conclusions about the evolvability of complexity in biology. Imagine that we found that yes, the flagellum evolves quite easily as long as we have some way of encoding entire subunits (eg, the hook, filament, driveshaft, and basal bodies) as "discrete" units that can be swapped out by evolution (with no mutational damage to the units aside from swapping them out). Or if we find that multicellularity evolves quite easily as long as we provide "discrete" packages of genomic information encoding a body plan: genes that regulate head morphogenesis, ventral/dorsal polarity generation modules, or limb-forming modules. Sure, if you provide most of the complexity in the units of evolution, complexity can probably evolve. But how relevant is that to the discussion?
It seems to me that if RBH and Rex want to take their own assertions seriously, they must seriously question the inferences that Lenski et al want to draw regarding biological evolution.
So: what, precisely, in biology, corresponds to the "instruction" of an Avida program? Is it biologically relevant?
John
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 20. April 2004 20:35
John Bracht asked quote:
So: what, precisely, in biology, corresponds to the "instruction" of an Avida program? Is it biologically relevant?
It's obvious that Bracht's invocation of "the hook, filament, driveshaft, and basal bodies" of the flagellum as somehow corresponding to discrete avida instructions is vacuous. As I said above with respect to those instructions, quote:
It's that they're discrete that allows them [instructions] to be appropriate entities for heritable (and mutable) information representation in a generic evolutionary system. (Emphasis added)
Hooks and filaments and driveshafts are not the heritable information representations of a flagellum-equipped bacterium. The heritable information representations in biological organisms are at the level of DNA/genes, and in avida they are at the level of instructions and sequences of instructions. That does not mean that instructions are exactly analogous to genes - obviously they are not - but that they are in that neighborhood rather than at the level of phenotypic expressions as Bracht incorrectly suggests.
Taking "evolutionary system" to mean a system in which populations of imperfect replicators reproduce in a non-uniform selective space characterized by limited resources, where at least some of the variation among the replicators is heritable, then biological systems and avida systems are both instances of that more inclusive category. Specifics of the two instances differ in some degree, just as the specifics of falling apples and orbiting satellites differ in some degree, but both have the essential defining features of evolutionary systems. Lenski, et al., showed that structures that meet Behe's definition of irreducible complexity can readily evolve in avida in a constrained context employing just a subset of the evolutionary mechanisms known to operate in biological systems. To make the case that the complexity of organic life cannot arise by those same evolutionary processes, augmented by the other evolutionary mechanisms known to exist in biology, IDists cannot appeal to generic arguments about evolutionary systems; they must show that specifically biological evolutionary systems have some unique property that insulates them from that finding. Truman's remarks about mutation rates and distributions in protein spaces are gestures in that direction. That's clearly an area for research, and some already exists as I noted early in this thread.
RBH
IP: Logged
|
|
John Bracht
Member
Member # 5
|
posted 20. April 2004 21:49
RBH:
quote:
Lenski, et al., showed that structures that meet Behe's definition of irreducible complexity can readily evolve in avida in a constrained context employing just a subset of the evolutionary mechanisms known to operate in biological systems.
But my point (which seems a bit lost on RBH) was that given the way that Lenski et al set up the system, irreducibly complex systems aren't that hard to evolve. They swap rather complex "instructions" instead of mutating individual codons in genes as evolution does. As I argued above, if one defines the system such that large blocks of complexity (the instructions themselves) are pre-encoded and only swapped out (never themselves built by the evolutionary process or destroyed by that process) then sure, irreducible complexity is rather easy to achieve. You just have to throw together the right blocks (instructions).
If biological evolution were like this, I doubt anyone would give a second's consideration to Behe's arguments. If evolution were powerful enough to drop whole proteins (RBH says instructions don't equal proteins but are at about that level of complexity, and I agree), de novo from nowhere (or rather, from a set of all relevant, ready-to-interact, pre-adapted proteins pre-defined somewhere) into various positions in the genomes of organisms, irreducible complexity would evolve rather easily.
So the whole issue is what exactly Lenski et al modelled. RBH wants to have it both ways: argue that Lenski et al didn't really model biological reality (and they didn't if you can't draw the kinds of analogies I'm trying to do between their simulation and biotic reality) and yet also argue that they proved that evolution can produce irreducible complexity. Maybe "Lenski evolution" can build IC systems/structures, but can "REAL" evolution?
John
[ 20. April 2004, 22:06: Message edited by: John Bracht ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 20. April 2004 22:56
John Bracht wrote quote:
f biological evolution were like this, I doubt anyone would give a second's consideration to Behe's arguments. If evolution were powerful enough to drop whole proteins (RBH says instructions don't equal proteins but are at about that level of complexity, and I agree), de novo from nowhere (or rather, from a set of all relevant proteins defined somewhere) into various positions in the genomes of organisms, irreducible complexity would evolve rather easily.
This is similar to arguments Bracht made in a Literature Forum thread on the same topic: quote:
I've noticed that genetic algorithsms that succeed at generating complicated outcomes often start with the high-level building blocks for those outcomes and just shuffle them around until the desired end product is produced.
charlie d answered it appropriately: quote:
John:
yours are sensible objections, and yet totally irrelevant. This was not a simulation of biological evolution, nor did the authors ever claim so.
This work proves one thing and one thing only: that within a population of simple, inaccurate replicators subject to some selective constraint, mutation, selection (and cooption) can result in the generation of complex features (i.e., for what we are discussing here, systems with the hallmarks of IC) without such features being initially "specified", "smuggled in", etc.
Thus, it does not matter what the original replicators looked like, what the mutation rate is, what the encoding language is, what kind of mutations are allowed, etc etc etc.
Again: this is not meant as a simulation of biological evolution in any way, shape or form, but as a proof of principle that IC systems can evolve spontaneously through evolutionary processes. (As the authors state at the beginning of the article, darwinian processes apply, by default, to any entity endowed with the ability to replicate and mutate in a selective environment.)
Of course, one can object, as you are doing, that the computer simulation is more efficient at this process than living organisms are, and everyone (including the authors and myself) will agree (that's why they used computers and not real live elephants). However, in that case, the issue becomes entirely quantitative and debatable; the qualitative uniqueness of IC, that somehow supposedly placed it beyond the reach of naturalistic evolutionary mechanisms, has evaporated.
and I responded (to a similar point later) quote:
Are you suggesting that the evolution of the flagellum did not have necessary building blocks available via cooption or recruitment, or that those building blocks are not themselves evolvable for other functions and thus become available for cooption? Are the proteins that comprise the flagellum unique to it and not available otherwise? Are there no other known biological systems that use proteins (or similar protein structures) in configurations similar to the functional components of the flagellum - are there no cilia elsewhere doing other things, for example?
Also in that thread Bracht wrote quote:
This is all a probabilities game, and if given all the components of a complex system plus a high shuffling rate, chance alone can shuffle them around into something selectable (like EQU).
The answer to that was obvious: The necessary "probabilities game" control condition was run and showed that given all the components and a high shuffling rate, chance alone can't generate something selectable like EQU. Chance can generate short segments of code that perform simpler functions, and those simpler functions, or parts of the code that underpins their performance, can then be co-opted to perform more complex functions, and so on. And that's the kind of account that Matzke, for example, provides for the flagellum.
In fact, the avida system makes available precisely the kind of data necessary to validate the core ID methodologies. One has all the necessary information to calculate Behe's "number of unselectable steps" metric in his modified definition of IC (and it is clear in the Lenski study that unselected steps occur). Similarly, it provides all the necessary information to calculate Dembski's specified complexity metric. If ID proponents actually take those metrics seriously, avida provides the ideal context in which to generate research that could convince "evolutionists" of their utility and validity.
Clearly the avida platform doesn't represent all the levels of analysis in a biological system. The authors of the program state that in several papers. But it does demonstrate that there are conditions under which evolutionary mechanisms can generate IC structures, and so the argument by ID must shift from "evolution cannot generate IC systems" to something like "for these specific reasons, biological evolution cannot generate IC systems." And those reasons have to be more than "because it's easy in avida and really really hard in biology." Neither the alleged ease in avida nor the alleged difficulty in biology have been shown by the IDists central methodologies.
RBH
IP: Logged
|
|
Argon
Member
Member # 276
|
posted 20. April 2004 23:55
John Bracht writes:
"But my point (which seems a bit lost on RBH) was that given the way that Lenski et al set up the system, irreducibly complex systems aren't that hard to evolve. They swap rather complex "instructions" instead of mutating individual codons in genes as evolution does."
I do not think the 'mutation-of-a-single-basepair-at-a-time' notion of genomic evolution adequately covers the full breadth of available mechanisms. Gene duplication, fusion and recombination all can swap complex 'instructions' and functional subunits. It is pretty clear that many of the IC systems cited by Behe (e.g. blood clotting, immune system & etc.) seem to have components that were 'borrowed' from pre-existing sequences at some point in the generation of 'new' IC systems.
[ 20. April 2004, 23:55: Message edited by: Argon ]
IP: Logged
|
|
|
Printer-friendly view of this topic