|
Author
|
Topic: Nature Refutes ID?: The Evolutionary Origin of Complex Features
|
Josh
Member
Member # 405
|
posted 08. May 2003 12:12
The following article has captured alot of attention very quickly. It is published in the current Nature:
Articles
Nature 423, 139 - 144 (2003); doi:10.1038/nature01568
The evolutionary origin of complex features
RICHARD E. LENSKI*, CHARLES OFRIA†, ROBERT T. PENNOCK‡ & CHRISTOPH ADAMI§
* Department of Microbiology & Molecular Genetics, Michigan State University, East Lansing, Michigan 48824, USA
† Department of Computer Science & Engineering, Michigan State University, East Lansing, Michigan 48824, USA
‡ Lyman Briggs School & Department of Philosophy, Michigan State University, East Lansing, Michigan 48824, USA
§ Digital Life Laboratory, California Institute of Technology, Pasadena, California 91125, USA
Correspondence and requests for materials should be addressed to R.E.L. (lenski@msu.edu).
A long-standing challenge to evolutionary theory has been whether it can explain the origin of complex organismal features. We examined this issue using digital organisms—computer programs that self-replicate, mutate, compete and evolve. Populations of digital organisms often evolved the ability to perform complex logic functions requiring the coordinated execution of many genomic instructions. Complex functions evolved by building on simpler functions that had evolved earlier, provided that these were also selectively favoured. However, no particular intermediate stage was essential for evolving complex functions. The first genotypes able to perform complex functions differed from their non-performing parents by only one or two mutations, but differed from the ancestor by many mutations that were also crucial to the new functions. In some cases, mutations that were deleterious when they appeared served as stepping-stones in the evolution of complex features. These findings show how complex functions can originate by random mutation and natural selection.
They manage to write an entire article about a "longstanding challenge" without once mentioning and or citing any ID theorists. Anyway, some links forwarded from the ASA listserve:
'Digital organisms' illuminate evolution (three versions of the article)
http://www.newscientist.com/news/news.jsp?id=ns99993706
http://www.eurekalert.org/pub_releases/2003-05/msu-aes050503.php
http://news.nationalgeographic.com/news/2003/05/0507_030507_digitalorganisms.html
This is obtaining alot of attention and should be relevant to those here. Any comments?
[ 14. May 2003, 12:11: Message edited by: Moderator ]
IP: Logged
|
|
nobody
Member
Member # 145
|
posted 08. May 2003 12:31
"Any comments?"
Yes. It took intelligence to write the computer program mentioned in the article. However it takes much greater intelligence to write the programming of life. In fact it requires more intelligence than all the scientists in the world put together currently have.
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 08. May 2003 15:58
The paper, supplementary materials, etc., are all online here:
http://myxo.css.msu.edu/papers/nature2003/
So, nobody, when someone develops a weather simulation, does that mean that weather requires intelligent design also?
Also, please show where solutions that the digital organisms came up with were input by an intelligent designer. Dawkins' "METHINKS" program this ain't...
IP: Logged
|
|
Nel
Member
Member # 614
|
posted 08. May 2003 18:55
So, where is the source code to this program?
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 08. May 2003 23:40
yersina,
What makes you so confident? My reading is just the opposite:
Dawkins' "METHINKS" program this IS...
As a friend of mine from Europe pointed out, all this simulation does is show that genetic algorithms can perform a simple hill-climbing exercise. A smooth pathway towards a desired target. Not very awe inspiring from the computer science side of things.
It is a bit unfortunate that so many people (including Nature) have embraced this without discernment.
IP: Logged
|
|
charlie d.
Member
Member # 159
|
posted 09. May 2003 09:44
Micah:
if it's a smooth path, it's actually even worse for ID, because it'd still be a smooth path to an irreducibly complex set of instructions, something that by definition should be impossible.
Actually, the authors argue that the EQU function evolved by cooption of unrelated, simpler functions, more in tune with what would be predicted based on observations of the evolution of biological complex systems, though I haven't had the time to read the paper carefully (and won't for a while).
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 09. May 2003 12:13
Hey Charlie,
quote:
if it's a smooth path, it's actually even worse for ID, because it'd still be a smooth path to an irreducibly complex set of instructions, something that by definition should be impossible.
Actually, this isn't right. When the smooth path in question is set up in the experiment as a red carpet towards the target, we shouldn't be impressed. In fact, we should be no more impressed by this system than Dawkins' Weasel program. Just because the researchers developed a more convuluted system, which is more difficult to analyze, doesn't mean we should automatically stand back in awe.
It is sort of like the black box of complex multi-layered neural networks. Hard to analyze, easy to admire. I'll take back my words later if on a second reading, the system seems to have done something extraordinary. I have no problem with that. What I do have a problem with, is the speed with which people have embraced the project without a measure of skepticism...a skepticism which is well deserved given what we know about computation and modern computer systems.
Hey, if anyone's interested, maybe we could talk about this paper in more detail sometime next week. It would be cool to have an informal reading discussion group. I'll have to go over the paper one more time in more detail, but my intuitions so far have been expressed above.
[ 09. May 2003, 12:15: Message edited by: Micah Sparacio ]
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 09. May 2003 13:28
Howdy,
In the "METHINKS" simulation, the target is pre-specified, and the program always reaches the exact same answer. Here is Dembski's critique:
quote:
Although Dawkins uses this example [METHINKS] to illustrate the power of evolutionary algorithms, the example in fact illustrates the inability of evolutionary algorithms to generate specified complexity. We can see this by posing the following question: Given Dawkins's evolutionary algorithm, what besides the target sequence can this algorithm attain? Think of it this way. Dawkins's evolutionary algorithm is chugging along; what are the possible terminal points of this algorithm? Clearly, the algorithm is always going to converge on the target sequence (with probability 1 for that matter). An evolutionary algorithm acts as a probability amplifier. Whereas it would take pure chance on average 10 to the 40 tries to attain Dawkins's target sequence, his evolutionary algorithm on average gets it for you in the logarithm of that number, that is, on average in only 40 tries (and with virtual certainty in a few hundred tries).
But a probability amplifier is also a complexity attenuator. For something to be complex, there must be many live possibilities that could take its place. Increasingly numerous live possibilities correspond to increasing improbability of any one of these possibilities. To illustrate the connection between complexity and probability, consider a combination lock. The more possible combinations of the lock, the more complex the mechanism and correspondingly the more improbable that the mechanism can be opened by chance. Complexity and probability therefore vary inversely: the greater the complexity, the smaller the probability.
It follows that Dawkins's evolutionary algorithm, by vastly increasing the probability of getting the target sequence, vastly decreases the complexity inherent in that sequence. As the sole possibility that Dawkins's evolutionary algorithm can attain, the target sequence in fact has minimal complexity (i.e., the probability is 1 and the complexity, as measured by the usual information measure, is 0). In general, then, evolutionary algorithms generate not true complexity but only the appearance of complexity. And since they cannot generate complexity, they cannot generate specified complexity either.
Explaining Specified Complexity
But this algorithm can attain a very large number of different EQU functions -- even though working EQU functions are such a small proportion of all possible "genomes" that the "tornado in a junkyard" runs that the authors did, where only EQU functions were rewarded, never produced EQU.
But, if there are systems with simpler, alternative functions, then EQU functions -- different each time, so *no* pre-specified sequence, but still EQU functions -- were produced regularly (not all the time, though). Furthermore, these EQU functions were irreducibly complex (or had IC cores, if you like), in that any of a large number of null mutations would abolish the originally identified function of EQU.
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 09. May 2003 13:33
Nelson,
The source code and executable for Avida can be downloaded by following the links on their page, and the parameter files they used as input are also available. I presume one could replicate the whole thing on a PC if you used the same random number seeds (which they specify on the page also).
IP: Logged
|
|
charlie d.
Member
Member # 159
|
posted 09. May 2003 13:55
Micah:
the problem is that a smooth, direct path to an IC system simply shouldn't exist (if there were, the system wouldn't be IC, would it?).
The way I read the paper, and to go back to a biological parallel, the target EQU function would correspond to a selectively advantageous trait, e.g. "bacterial motility". Depending on evolutionary contingency, different solutions to this problem exist, and each can be reached based on what other functions happen to be available to individual evolved program lineages (bacterial lineages). These lower-order components evolved independently, for advantageous functions other than EQU (in bacteria, for instance, protein secretion, or adhesion, etc), and were later coopted into the EQU trait in various ways. It also so happens that in some cases these lower functions combined in a way that the EQU solution is dependent on each and everyone of these lower components, and can't function at all without (ie. it's IC, like flagellum-based motility).
Again, this is just after a cursory reading.
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 09. May 2003 13:58
Well, considering that they used dozens upon dozens of linux boxes for this experiment, I doubt it could be run on your average PC with any success. Who knows though. But it shouldn't matter. We can just assume that everything they say about the simulation (outside of the implications) is accurate. Skepticism regarding the implications is still highly warranted.
*****************************
yersina, it is quite possible, in the context of GA search, to run into "sticky" points and maxima that the algorithm can't get out of. My suspicion is that the variability of success is merely a function of the nature of the algorithm. So the road to the target class isn't guaranteed.
But this shouldn't matter. There might be contingency regarding whether the algorithm finds the prepared path or not. But the path is still prepared and relatively smooth.
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 09. May 2003 14:02
But charlie, an algorithm can be developed to follow a smooth path towards the discovery of any one instance of a constrained target set of solutions and then implement a solution for putting a mouse trap together. This says very little.
![[Frown]](frown.gif) Forgive the mouse trap example.
[ 09. May 2003, 14:16: Message edited by: Micah Sparacio ]
IP: Logged
|
|
YZ2
Member
Member # 91
|
posted 09. May 2003 15:41
Obviously, the paper is not an easy read. I have a number of questions in my mind for comments here.
EQU is considered a "complex" instruction, because it cannot be computed sequentially using the defined instruction set in a stack machine in a simple manner.
My question is: how complex is it, if I have, say a parallel machine instead of what is defined there? Is EQU just a rehash of the
existing semantics from the instruction set, or a truly new instruction? Another way to put it is: is there a simple pathway to construct EQU?
In any case, even if EQU cannot be computed from existing instruction set in one shot, the claim that it is a "complex" feature comparable to biological complex feature is, seems to me
a gross exaggeration. At most it can claim that it is a complex instruction that is generated. Complex instruction as defined in the paper could be quite different from biological complex feature.
My second question is: can similar procedure originate complex instruction
"semantically" different from the original instruction set? An example I can think of
(which I am not sure is a good example with respect to mathematical properties but may be a good starting point for discussion) is the "TestPrime" instruction.
Any ideas?
[ 09. May 2003, 16:28: Message edited by: YZ2 ]
IP: Logged
|
|
Erik
Member
Member # 160
|
posted 10. May 2003 06:43
Micah Sparacio informed us that "When the smooth path in question is set up in the experiment as a red carpet towards the target, we shouldn't be impressed". Please clarify in what sense the smooth path was set up as a red carpet towards the target. Exactly what should the authors have done differently in order to better investigate the question the authors were interested in? Exactly what would you have done differently in order to better investigate the question you are interested in?
In a later post Micah Sparacio stated: "My suspicion is that the variability of success is merely a function of the nature of the algorithm". Exactly which algorithm is being referred to here? Is it the basic Avida platform for executing the instruction codes of a population of self-replicating programs? Is it the algorithm (aka "selective environment") that determines how much processing time each self-replicating program gets? If it is the latter, then I note that this is indeed important and some results are reported by the researchers:
quote:
"We carried out experiments to examine the effects of different selective environments on the propensity to evolve the EQU function, with all other conditions held constant. Ten further populations evolved under each of 36 possible regimes in which one or two simpler functions were not rewarded. In all environments, at least one population evolved EQU. Evidently, neither any particular simpler function nor any pairwise combination of functions was required to evolve this complex feature. In these 36 environments, the overall fraction of populations that evolved EQU was 124 of 360 (34%), only slightly less than in the ‘reward-all’ environment (P = 0.0764, one-tailed Fisher’s exact test).
At the other extreme, 50 populations evolved in an environment where only EQU was rewarded, and no simpler function yielded energy. We expected that EQU would evolve much less often because selection would not preserve the simpler functions that provide foundations to build more complex features. Indeed, none of these populations evolved EQU, a highly significant difference from the fraction that did so in the reward-all environment (P ~ 4.3 * 10^-9, Fisher’s exact test)."
Lenski et al., Nature, 423:139-144
What the experiment proves is that Irreducible Complexity (in the sense of the 1996 definition) can evolve by cooption of different, simpler functions. The identification of IC in the clinical Avida world is much more reliable than in the messy world of biology and the experiment shows that IC can evolve in an environment where different, simpler functions increase the reproductive success. ID advocates should also note carefully that there are numerous paths to the function of the IC system in question (see the "Discussion" section of the paper). Although there may be important differences between the Avida world and real-world biology, none of these differences play any part in Behe's argument. Behe's argument is just as applicable to both. Therefore the evolution of IC in Avida is a valid counterexample.
Erik
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 10. May 2003 13:49
What I thought was interesting ab how it showed that essential mutations arose initially through hitchhiking of detrimental mutations.
This makes the calculations of pathways in real biological examples even more complicated since we cannot just reject a mutation just because it has some detrimental impact without looking at the overall picture.
The re-use or co-option of simpler components, the hitchhiking of detrimental mutations which later become essential in the function of EQU seem to all explain why the at first sight simple 'mutation and selection' can be quite inventive.
It also shows that the two components are essential, mutations are needed for variation but are internal to the cell and selection which is also essential but is external to the cell. The combination of an internal and external mechanism makes it possible for information/complexity to be transfered from the environment into the genome.
While Dembski's 4th law of thermodynamics (or should we call it the 2nd law of thermodynamics for closed systems) applies to closed informational systems in which entropy can only increase, opening up the system leads to an efficient way to transfer entropy out of the genome.
It is important that we do not underestimate the power of the rather simple mechanism for evolution, the fact that the two components are located in different environments is what allows for complexity/entropy to increase/decrease in the genome at the expense of the complexity/entropy in the environment. This suggests that in RMNS, natural selection is the designer that 'injects' (C)SI into the genome. I would also suggest that RMNS being a natural designer may not be in contradiction with Dembski's ID inference since the ID inference tries to establish (intelligent) design without identifying the designer involved. But it does seem to show that the NFL theorems may be irrelevant as suggested by findings released after the publication of No Free Lunch.
I would love to see Dembski's calculations applied to this example to determine what the probability calculcations using perturbations, localization etc would suggest for the probability of the EQU gene to evolve?
IP: Logged
|
|
|
Printer-friendly view of this topic