GAMOVA vs STRUCTURE?

Posted by The putrid stench of gnxp on April 07, 2007, 03:27 PM | #

I don’t have the time to look this paper over as carefully as I would like, but note that ‘clustering’ can be viewed as a metric different from genetic similarity/differences.

Remember the Jorde paper from the Utah group discussed here recently.  The “w” measure of genetic similarity was a more stringent measurement than clustering, and is more in line with information need for determination of genetic interests.

Posted by James Bowery on April 07, 2007, 05:04 PM | #

You are referring to the MR article “Genetic Similarities Within and Between Populations” wherein the academic paper by the same name by David J Witherspoon, Stephen Wooding, Alan R Rogers, Elizabeth E Marchani, W Scott Watkins, Mark A Batzer and Lynn B Jorde is discussed.

In that MR discussion you stated:

as the post states, “w” ignores structure

Could you present a definition of “structure” and how it differs from “clustering” if at all?

It sounds to me like you’re, perhaps indirectly, criticizing the use of correlation structure measures to strengthen the theory of EGI.

Posted by The putrid stench of gnxp on April 08, 2007, 08:38 AM | #

“It sounds to me like you’re, perhaps indirectly, criticizing the use of correlation structure measures to strengthen the theory of EGI.”

That couldn’t be more mistaken.

Clustering, essentially, is based upon comparing an individual’s gene frequencies to the “centroid” of a particular population; that is, if Europeans are characterized by a set of gene frequencies, and Africans by another, how close is the individual to either set.

I would define structure as composing of those elements of genetic information above and beyond a one-by-one evaluation of gene frequencies and including as well genetic structures beyond that of the single allele.

Easy examples of “structure” would include copy number variation, insertions, deletions, inversions, etc.  But these are NOT the major determinants of structure, nor what we here usually refer to as structure.

That would be the coinheritance of particular alleles (or, even, gene sequences) between individuals and groups.  If one reads the “materials and methods” section of the Jorde paper, they are defining “w” as the average distance of each allele taken in turn - that is _not_ what we would consider structure.

The basic component of structure would thus be defined as the coinheritance of gene sequences/alleles.  Linkage disequilibrium is therefore one aspect of structure, and new LD patterns are established, not surprisingly, upon admixture.

To put it even more simply and to use a crude analogy (very crude), it is not the allele frequencies for, say, eye color, skin color, hair color, nose shape, skull shape, etc. each taken in turn and averaged.  Instead, structure concerns itself with the frequency that these various types of alleles happen to be co-inherited in the same individual or group, the frequency by which particular _combinations_ of the allele type show up in the specific set.

As far as I can see, this is not being measured by Jorde’s “w” distance metric, nor the “c” clustering metrics.

If I am mistaken, then I stand open for correction; perhaps Jorde et al., were not as clear in their paper, and I misinterpret their methodology.  I would, however, require clear and unmistakable evidence that such a misinterpretation has taken place.

Posted by The putrid stench of gnxp on April 08, 2007, 08:47 AM | #

Note: I mistakenly placed the next two posts on this topic in the amygdala thread.

Posted by James Bowery on April 08, 2007, 09:50 AM | #

The misplaced comments were:

From Jorde:

1. Methodology:

“Pairwise Genetic Distance
We use the “shared alleles” genetic distance (Bowcock et al. 1994; Chakraborty and Jin
1993; Mountain and Cavalli-Sforza 1997), which defines the distance between two individuals at
a locus as one minus half the number of alleles they share. The genetic distance between
individuals is the average of their per-locus distances. Pairs of individuals are classified as
“within-population” or “between-population” according to whether the individuals were sampled
from the same or different groups of populations as defined above.
8
Dissimilarity fraction ?ˆ
Let ? be the probability that a pair of individuals randomly chosen from different
populations are genetically more similar than an independent pair chosen from any single
population. We compute all possible pairwise genetic distances, classify them as within- or
between-population distances (the sets dW or dB, respectively), then calculate the frequency with
which dW > dB (that is, a within-population pair is more dissimilar than a between-population
pair.) This fraction, ?ˆ , is an estimator of ?. The expected value of ?ˆ ranges from 0 to 0.5
(regardless of the number of populations). At ?ˆ = 0, individuals are always more similar to
members of their own population than to members of other populations; at ?ˆ = 0.5, individuals
are as likely to be more similar to members of other populations as to members of their own. The
distributions of pairwise genetic distances implied here resemble the Common Ancestry Profiles
proposed by Mountain and Ramakrishnan (2005), who use a different measure of genetic
distance. The shared-alleles distance used here generally yields slightly lower values of ?ˆ .
Centroid misclassification rate CC
The centroid classification method is also based on pairwise genetic distances, with one
critical difference: every individual is compared to the centroid of each population, rather than to
every other individual. The centroid is the genetic average of a population, an individual whose
pseudo-genotypes at each locus are the frequencies of the genotypes in that population (not
including the individual being compared to the centroid). This genetic distance is equivalent to
the average of the genetic distances from an individual to all other individuals in the target
population. Each individual is then assigned to the population with the closest centroid, as in
Cornuet et al. (1999). These assignments are compared to the known populations of origin, and
the proportion of individuals misclassified is reported as CC. The expected classification error for
9
random assignment of individuals to populations is 1 – 1/n, where n is the number of
populations.

2. Problem with clustering:

“Nonetheless, it is instructive to consider the analogy using panel C as a guide.
For example, an African individual x with qx = 0.52 will be more similar to a European y with qy
= 0.55 than to another African z with qz = 0.4. Yet that individual x will still be closer to the
population mean trait value for Africans (qA ? 0.46, the African centroid) than to the mean value
of Europeans (qB ? 0.61).”
Posted by The putrid stench of gnxp on Sunday, April 8, 2007 at 12:44 PM | #

“The genetic distance between
individuals is the average of their per-locus distances.”

This is _not_ how I would define “structure.”
Posted by The putrid stench of gnxp on Sunday, April 8, 2007 at 12:46 PM | #

Posted by The putrid stench of gnxp on April 09, 2007, 10:48 AM | #

A review of “On Genetic Interests” by van den Berghe, who considers himself a “political anarchist”:

http://lists.paleopsych.org/pipermail/paleopsych/2005-May/003227.html

One can contrast van den Berghe’s essentially fair review to the ludicrous “hatchet job” by David B of gnxp.  Of course, Dr. van den Berghe seems to be an independent entity, and not an extended phenotype of devious south asians.

Also, read the last paragraph of van den Berghe’s review.  Sound familiar?  “Life has no interests”, “who cares”, “toffee and Mahler.”

Posted by Dasein on October 13, 2009, 11:35 AM | #

The problem with this approach, as it concerns correlation structure (‘allelic covariance’), is that the distance measurements they use are based on single loci.  It seems to me, as has been pointed out here before, that incorporating correlation structure is the major task for updating Salter’s work.  I’m not a population geneticist, but from what I can determine, the commonly-used distance measurements are based on single loci (e.g. the FST distances used by Salter/Harpending).  Is there an established way to quantify distance that considers correlation structure?  A couple of ideas come to mind, but perhaps someone more knowledgeable in the area can correct me if I’m wrong.  The first would be to consider haplotypes instead of alleles (this should presumably increase the proportion of inter-group variability).  I’m not sure how one would decide which, and how many, haplotypes to check, though once the relevant SNPs were known it would be trivial to get types for different races and admixes using the HapMap data (although I don’t know that the SNPs are phased, which might pose problems).  A second possibility, more of a ‘hack’, would be to evaluate random combinations of loci- in effect creating random haplotypes in silico.  This could be tried with different numbers of loci, and it would be interesting to see how the proportion of inter-group variability increases as a function of number of loci in the artificial haplotypes.  I don’t know how one would determine a ‘natural’ or ‘valid’ cutoff (typically, when clustering you can look for an elbow, or flattening-out, of the variability captured by the number of clusters).  Anyhow, it seems like an important topic and one I’d be interested in exploring.  Like I said, I’m not a population geneticist, so I am hoping to get some feedback from James and others who may know more about this.

Posted by James Bowery on October 23, 2009, 02:14 PM | #

My work on compression-as-intelligence has led me to the working belief that the best currently practical approach is to treat structure as algorithmic information.  What this means is you take a set of data—any data at all—and create the minimum length computer program that outputs that data. 

Its easy to state this as the goal but it is, in fact, provably uncomputable to come up with said minimum length program—although it is asymptotically computable with exponential computing resources.

There are other approaches involving generalization of relation arithmetic—hence quantum information systems—that may prove superior if or when quantum “computers” (a misnomer really) become a practical reality.

Posted by Dasein on October 23, 2009, 03:14 PM | #

Thanks, James.  That’s a lot to think about. 

From what I’ve been able to discern, one of the biggest problems with existing measures of variability is that, so long as there is a decent number of alleles, they are guaranteed to show high intra-group variability.  This necessarily limits the proportion of inter-group variability.  And for this reason, my in silico haplotype approach would actually decrease the proportion of inter-group variability.  I know that Salter’s use of Fst underestimates ethnic kinship, I’d just like to find a way (preferably based on existing methodologies) to quantify this.  I’ve ordered some books on measuring genetic distance, and I’ve also been looking at some work by Lou Jost, who is critical of existing distance measurements.  I’ll come back to this thread when I’ve had a chance to get through it.

Posted by James Bowery on October 23, 2009, 04:08 PM | #

There is an enormous amount of work on compression of genetic data.  Indeed, it may be the most active area of bioinformatics.

The primary value of what you call “existing distance measurements” is that they offer guidance as to what sort of a priori (or Bayesian prior) structures may offer optimal compression. 

The test is whether they do, in fact, enable the production of shorter programs that output the known data.

Posted by Loriver on October 27, 2009, 07:12 PM | #

Dasein:

From what I’ve been able to discern, one of the biggest problems with existing measures of variability is that, so long as there is a decent number of alleles, they are guaranteed to show high intra-group variability.  This necessarily limits the proportion of inter-group variability.  And for this reason, my in silico haplotype approach would actually decrease the proportion of inter-group variability.  I know that Salter’s use of Fst underestimates ethnic kinship, I’d just like to find a way (preferably based on existing methodologies) to quantify this.

The real problem you should be considering is that of inter-ethnic similarity, not the magnitude of differences between ethnic groups. The differences are big enough already - you think 1,000,000 in child-equivalent reproductive losses fails to make the point? But the degree to which a random Englishman might be, in EGI terms as they stand, somehow an ‘Italian’ or a ‘Moroccan’ has not been investigated fully. This variability in itself is the issue, not any secondary concern that it reduces the apparent scale of ethnic differences.

Also, I believe that kinship has a formal definition and Fst is, in strict mathematical terms, equivalent to average kinship. Just take another look at http://majorityrights.com/index.php/weblog/egi

So it’s not quite true that Fst underestimates ‘ethnic kinship’, but it may well be true that Fst is not the only measure of ethnic genetic identity that we should use, especially with a view towards limiting inter-ethnic similarity.

Do consider: to what extent could ‘structure’, or any non-Fst measure of ethnic similarity, be compared to family kinship (e.g. as child-equivalent losses)? How important is this?

Posted by Guessedworker on October 27, 2009, 07:40 PM | #

Thanks, Loriver.