tag:blogger.com,1999:blog-3017697080484068141.post6828401741820548976..comments2016-10-17T12:50:27.511-07:00Comments on No DNA Control: It's more delicious because it's more nutritiousChanghttp://www.blogger.com/profile/12291718994939895064noreply@blogger.comBlogger4125tag:blogger.com,1999:blog-3017697080484068141.post-86222025351334925792010-08-21T11:58:43.931-07:002010-08-21T11:58:43.931-07:00Right. If you normalize to an 'expected' ...Right. If you normalize to an 'expected' matrix then the cross should mostly disappear. What remains will be the evidence for true interactions, which then need to be investigated individually.Rosie Redfieldhttps://www.blogger.com/profile/06807912674127645263noreply@blogger.comtag:blogger.com,1999:blog-3017697080484068141.post-37265735970600833992010-08-21T11:49:31.161-07:002010-08-21T11:49:31.161-07:00Right. Palate.
It is appropriate to think of the...Right. Palate.<br /><br />It is appropriate to think of the 2D plot above as an "importance" plot, rather than an interaction plot. But it's not simply made by multiplying single frequencies together... That would be the "expected matrix" for the analysis you suggest above (I think).<br /><br />I started with this view, because it can be directly related to the simpler barplot above it, so had some intuitive appeal for me.<br /><br />The reason this is not truly an "interaction plot" is because I am only reporting the observed enrichment/depletion of pairwise mismatches. As I alluded to above, if I normalized these values to an expected matrix as you describe above to produce "interaction plot", the cross should mostly disappear... Having a mismatch at position 7 does not help you predict the chance of a mismatch anywhere else.<br /><br />I suspect, based on staring at this thing, that the AT-tracts and position 4 will show an interaction. <br /><br />And then of course I have to examine what the mismatches actually are, but baby steps over here... baby steps. When I try to jump straight in, I get confused...Changhttps://www.blogger.com/profile/12291718994939895064noreply@blogger.comtag:blogger.com,1999:blog-3017697080484068141.post-77865357317967183942010-08-21T07:53:41.022-07:002010-08-21T07:53:41.022-07:00So, I think the covariation analysis might best be...So, I think the covariation analysis might best be done this way:<br /><br />Take the pairwise variation data (the 2D graph) and normalize it by the expected pairwise variation (the product of the two single-position variation values). <br /><br />Plot that. Pairs with scores >1 and scores <1 are candidates for interaction effects.<br /><br />For the candidate pairs, compare the frequencies of each combination to the expected frequencies given the single-position values. <br /><br />Plot a 4x4 array showing these. For positions where the overall normalized analysis showed interaction score <1, we expect to see that the consensus pair is underrepresented and some combination of bases other than the consensus is overrepresented. For positions where the score was >1, we expect to see that the consensus pair is overrepresented and all non-consensus pairs are underrepresented.<br /><br />Or maybe vice versa, if I've gotten this confused.Rosie Redfieldhttps://www.blogger.com/profile/06807912674127645263noreply@blogger.comtag:blogger.com,1999:blog-3017697080484068141.post-66937404743349951082010-08-21T06:32:38.813-07:002010-08-21T06:32:38.813-07:00Palette? Palate?
Is this really a matrix of inte...Palette? Palate?<br /><br />Is this really a matrix of interactions, or just of the products of importances? I think those are quite different things. If you made a matrix of the products of the numbers of the one-dimensional graph, would you get the two-dimensional graph?<br /><br />Are the inportances in the 2D plot normalized to the importance of each base considered singly in the 1D plot? If yes, then the colours in the 2D plot tell us that there is an interaction effect between these positions. But if they're not, then this is just the product of the importances.<br /><br />To get the full interaction effects, I think you need to consider more information. Specifically, when considering the cases where two specific bases don't match the consensus, I think you need to consider WHICH non-consensus bases are at those positions, and whether there are any correlations between them. For example, in the 2D graph, cases where positions 4 and 26 are mismatched might occur a bit more often than one might expect (a bit more often than the product of the individual expectations). To say that this is because of an interaction effect, you'd want to get the frequency of each base combination, and then see if any non-consensus combinations are overrepresented.Rosie Redfieldhttps://www.blogger.com/profile/06807912674127645263noreply@blogger.com