Wesley R. Elsberry
Posts: 4991
Joined: May 2002
|
Congrats to Cryptoguru for reading some of the responses and making some changes to his argumentation. This is not commonly observed in these sorts of exchanges. Plus, welcome back to the beginning of the second farewell tour. (I wrote that last sentence before the second flounce.) I'm still not inclined to accept changes in the original challenge, though, so I will be specifically noting those.
| Quote (cryptoguru @ Jan. 19 2015,07:45) |
[...]
Section D:
comparison of AVIDA to biological evolution
1) an analogy must be drawn between AVIDA commands and the genome
2) it possibly makes most sense (I concede) to assume that AVIDA commands are analogous to codons (and not proteins), so that any mutation will always create a set of valid codons.
3) the level AVIDA is selecting at is therefore analogous to a folded functional protein.
|
A problem with (3) is that I've already pointed out the level of selection is the Avidian, not any sub-sequence of its genomic instruction sequence.
| Quote (cryptoguru @ Jan. 19 2015,07:45) |
Section E:
problems with model
1) the multi-dimensionality of the genome due to multiple reading frames means that a point mutation in the genome will likely affect the expression of multiple proteins (19K coding genes code for at least 100K different proteins in the human genome). Mutations in AVIDA are mutually exclusive and therefore don't have a regressive effect on the expression of other COMMANDS. This is not a trivial difference, it is analogous to the difference between a bisection method (AVIDA mutation) and a bisection method where the root can change at each iteration (Genome mutation).
|
This is not in the original challenge, and is thus irrelevant to answering the original challenge. The Avida instruction set includes "mov-head", "jump-head", and "set-flow", which can and do change expression dramatically. Avida itself has been used to perform in silico experimentation on overlapping genes:
| Quote |
One consequence of overlapping genes is to reduce the tolerance for mutation. Virtual experiments conducted within the past several years using a software system called Avida have indicated that overlapping reduces the probability of accumulating so-called neutral mutations in a gene (mutations that have no effect). Neutral mutations are unlikely with overlapping genes, because the mutation must have no effect on two genes with different reading frames.
|
But like I said, that's all irrelevant to the original challenge.
| Quote (cryptoguru @ Jan. 19 2015,07:45) |
2) It is not just multiple reading frames that introduce polymorphism into the genome, but regulatory genes can effect the expression of an entire coding gene. This non-linearity is not modelled in AVIDA, which is a linear sequential code (like assembler). That is, the Genome executes a higher-level language than a sequential instruction set.
|
Again, this concern is nowhere to be seen in the original challenge, and is thus irrelevant to the original challenge. Plus, the Avida documentation notes:
| Quote |
One major concept that differentiates this virtual assembly language from its real-world counterparts is in the additional uses of nop instructions (no-operation commands). These have no direct effect on the virtual CPU when executed, but often modify the effect of any instruction that precedes them. In a sense, you can think of them as purely regulatory genes. The default instruction set has three such nop instructions: nop-A, nop-B, and nop-C.
|
But like I said, that's all irrelevant to the original challenge.
| Quote (cryptoguru @ Jan. 19 2015,07:45) |
3) AVIDA enforces selection by rewarding at the functional level ... it identifies a function as a logical operation and rewards the organism that is presenting it. This is equivalent to natural selection providing feedback scores to the organism on a per-protein level. e.g. protein 1 7/10, protein 2 3/10, protein 3 9/10. This kind of micro-management can't happen in real-life as natural selection is blind and is applied not even to an organism but an entire population. This feature artificially boosts the productivity of hopeful combinations of commands, which otherwise wouldn't be encouraged.
|
Bogus in many particulars. Group selection is about as dead a concept as it is possible to get in biology. Not so long ago, Cryptoguru asserted this:
| Quote |
1) evolution mutates DNA on a nucleotide level affecting function (gene level) and selects on the organism level.
|
which appears to be that rarity in logical fallacies, a contradiction. It also indicates that Cryptoguru's conceptual movement on this particular is in a bad direction.
Avida does not examine the genomic instruction sequence to recognize something. It examines the output from the IO instruction of the Avidian that indicates that it correctly performed a behavior, in the Lenski et al. 2003 paper the rewarded behaviors comprised a set of nine logic operations. An Avidian can internally compute every logic function around and receive exactly zero extra CPU cycles of merit if it fails to output the results to the environment via the IO instruction. I've already mentioned this before. The assertion that things are otherwise is a persistent misunderstanding on Cryptoguru's part.
Avida's award of merit for Avidian behaviors is analogous to biological organisms getting better/more nutrition, or greater movement efficiency, or better artifact construction (nest, hive, or tools) due to a favorable trait. There is really nothing to object to on this point, and this is no problem for the model.
As far as the final quoted sentence goes, some traits have to have relative benefits in order to simulate natural selection. This has to happen in the model since it happens in biology. This is not a problem for the model.
| Quote (cryptoguru @ Jan. 19 2015,07:45) |
4) The level of variability in AVIDA compared to the genome is like comparing solving a Rubik's cube with cracking 2048-bit RSA encryption. The logical functions which AVIDA selects and guide the optimisation process are trivial ... and also non-distinct. There are infinitely many ways to implement the EQU function using the AVIDA instruction set. Proteins are specific in their form, not just in an abstract functional way. The likelihood of randomly selecting a combination of AVIDA commands that performs a logical function is extremely high, I don't know if it's even possible to work this out considering there are infinitely many combinations that could represent each logical function. (e.g. inc dec inc dec inc dec is equivalent to leaving a register unchanged). In the genome the possibility of just mutating 3 neighbouring nucleotides anywhere in the genome to produce a different codon is less than the chances of winning the national lottery twice in a row, and two proteins with different amino acid chains can likely never be equivalent in function. (unless randomness of function is the required function for the protein)
|
Again, this concern is nowhere to be seen in the original challenge, and is thus irrelevant to the original challenge. The original challenge made no restrictions on the size of a novel function.
Proteins are not absolute, on-off switches of function for a given amino acid sequence/fold configuration. Proteins often exhibit partial functionality. Physiologists also know that proteins work better and worse for a given function at different temperatures. Proteins also often tolerate substitutions of amino acids without drastic changes in function. Even Douglas Axe's work shows that he had to go to swapping out large swaths of amino acids in order to almost entirely eradicate barnase function. So perhaps the two proteins are not equivalent; that does not mean that they are significantly different, which is what Cryptoguru's argument requires. See also Dayhoff's Atlas of Protein Sequence and Structure, which shows the differences in the proteins used for the same purpose across a variety of organisms. Those often show multiple amino acid differences in even a relatively short protein like cytochrome-c. That's been around since the 50's, so there's not much excuse for not knowing that before spewing.
About "In the genome the possibility of just mutating 3 neighbouring nucleotides anywhere in the genome to produce a different codon is less than the chances of winning the national lottery twice in a row"...
That is really unclear. Does Cryptoguru mean the odds of getting a different codon given three neighboring nucleotides mutate? Three nucleotides specify a codon, given four nucleotides, that's a total of sixty-four possible codons. Randomly picking new values for each of the three nucleotides will only yield the same codon in 1/4*1/4*1/4 = 1/12 of the time, or about 8%. The complement of that is about 92%. Does he mean instead, as I suspect, that single nucleotide mutations are rare, and that getting three adjacent ones are even rarer? True, but it doesn't work out as he seems to wish. Changing a single nucleotide would give a different codon by definition, and given a uniform sampling of codons, that yields about an 80% chance of a different codon that yields a different amino acid as a result. Yeah, the SNP event is ~1e-9, but a change in amino acid resulting is pretty common as a result of an SNP. One does not need to change all three nucleotides to get a different codon, nor to get a different amino acid as a result.
Nor do I buy the "probability is extremely high" gambit based on arm-waving. Cryptoguru should either show his work or give it up. Towards that work...Sure, there are infinite ways to get EQU function. And for each one of those, Cryptoguru needs to estimate the number of Avida programs of the same length that do not provide EQU function. For his claim, he needs to show that number is far from L^26-1, where L is the program length. A probability will incorporate that other number, and not just count the "hits". A sample of program lengths from the minimal EQU length to, say, 50, should suffice. My own assessment of probability of EQU, based on actually having used Avida, programmed Avida, and programming and testing Avida-ED changes, is that hitting on EQU randomly is a tiny, tiny probability. That's arm-waving, too, but with some experience to back it. Then there is the result in the Lenski et al. 2003 paper that directly addresses this concern experimentally:
| Quote |
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 x 10^-9, Fisher's exact test). However, these populations tested more genotypes, on average, than did those in the reward-all environment (2.15 x 10^7 versus 1.22 x 10^7; P < 0.0001, Mann-Whitney test), because they tended to have smaller genomes, faster generations, and thus turn over more quickly. However, all populations explored only a tiny fraction of the total genotypic space. Given the ancestral genome of length 50 and 26 possible instructions at each site, there are ~5.6 x 10^70 genotypes; and even this number underestimates the genotype space because length evolves.
|
The only real numbers in this particular aspect of the discussion indicate that for Cryptoguru "extremely high likelihood" can refer to a probability smaller than 1 in 2.15 x 10^7. Your mileage may vary.
But like I said, that's all irrelevant to the original challenge.
| Quote (cryptoguru @ Jan. 19 2015,07:45) | | DISCUSS! |
Cryptoguru's original challenge has been met. If Cryptoguru wants to move on to a different challenge as his objections seem to indicate, he should at least acknowledge that his prior one was met before stating a new challenge.
Much of the discussion that ensued after Cryptoguru's original challenge was due to Cryptoguru's inability to focus on the terms of the challenge he himself wrote. There seems to be a curious vagueness about whether Cryptoguru wants a model of natural selection, a model of genetic inheritance, a model of codon replacement, or a model of abiogenesis itself. Certainly the objections raised afterwards have touched upon all of those. There is a concept called salience that Cryptoguru ought to get familiar with. A model aimed at determining whether new information can arise in computer code via an evolutionary process doesn't need to be freighted with most of the irrelevancies that he has discussed. In the extreme, it appears that Cryptoguru wants the equivalent for biology of the "theory of everything".
The other significant fraction of Cryptoguru's output concerned things that he believed were true, but weren't. For instance, the majority of Cryptoguru's claims concerning the Avida system were and are, charitably speaking, bunk. This didn't prevent Cryptoguru spouting falsehoods with fervor and vehemence, and ironically insisting that others had gotten their facts wrong on that score.
--------------
"You can't teach an old dogma new tricks." - Dorothy Parker
|
