Talk:Alzheimer's disease, gene set

Great to have this PHS on Promethease. Not sure how this would be done with the Promethease software. Using a gene set? Could give the users their normalized beta score with some context on what it means.

Genetic science appears to have reached the point that such polygenic scores could be clinically and personally relevant.

I have found perfect proxies that are on 23andme for SNPs that are not on 23andme. It might be helpful to include these.

There seems to be a fair amount of online interest in this research, though people probably would prefer to pay someone else than do all the look ups and determine which are the risk alleles themselves.

I am not sure yet which alleles are actually the risk alleles. The column "Risk alleles" is actually the minor alleles. The article was not clear on this question.--J1 (talk) 17:39, 26 March 2017 (UTC)

Reverse engineer the PHS? For whatever reason ( perhaps commercial?) no distribution of the PHS appears to be available from the article. Without such a distribution one would not know whether they were in the 1st percentile or 99th percentile. There could be a way around this. We could simply use the MAF to determine the distribution.

Assuming the SNPs are independently inherited, then the distribution should be calculatable from the dbsnp and beta numbers.--J1 (talk) 17:39, 26 March 2017 (UTC)

Just ran the dataset through a program to find the beta PHS distribution.

On a 10,000 person sample, assuming simple additive betas (article mentioned beta log): 1% lowest risk was below -0.51. 10% lowest risk was below -0.03. 50% median risk was at 0.59. 10% highest risk was above 1.24. 1% highest risk was above 1.78. --J1 (talk) 22:19, 26 March 2017 (UTC)

no promises, but if you'll share the code you've got with info@promethease.com I'll see if it's possible to introduce something similar into a new tab, ala the 'Blood'. --- cariaso 03:41, 27 March 2017 (UTC)

% Not pretty, though it looks like this is getting close. Simply add the weighted log betas across all 31 SNPs. % Below program then ran a simulation with 1 million people. I was trying different things in the program so that is why I commented things % out. I had to translate the distribution because the values around zero were causing trouble with the floor function.

% Without knowing the distribution of the polygenic hazard score, the article is useless. What would a PHS of 1 mean? % The article fills in the details in detail (supplementary figure) as soon as a percentile score and age is known. % There probably will be a lot of interest in this article's result because it clearly shows that everyone is at risk of dementia. % Would like to double check that this simple method is in fact correct.

% Here's what another study using a polygenic approach had to say:

%"In the target sample, we calculated the total score for each individual as the number of score alleles weighted by the log of the odds ratio % from the discovery sample. Scores are additive across SNPs on the log odds scale and therefore multiplicative on the odds of disease % scale." PMID: 19571811 % Supplementary page 20 % So, all you seem to need to do is multiply the log weights by the allele count. % They should all be independent.

% If the above is all correct, then by running a simulation as below you should be able to find the %tiles for the beta log hazard score. % I have run these simulations and it appears that the mean is 0.59, with a standard deviation of 0.49, and the percentiles as above. % If I could upload spread sheets somewhere I could show the charts and some numbers. % When I binned everything there was a nice normal curve.

% The code needed for Promethease would be something similar though easier. % All you would need would be to sum the weighted log betas across the 31 SNPs. % You could then give perhaps a percentile value. The article seemed to concentrate on those in the highest or % lowest 1%,10% and 25%. Might also want to add in APOE, I didn't do this below though it would be easy to add in % 2 more betas.

var MAF : array 1 .. 31 of real var beta : array 1 .. 31 of real var sumsq : real sumsq := 0

a1 := 0.07 a2 := 0.14

var r1, r2 : real

MAF (1) := 0.1426 MAF (2) := 0.4091 MAF (3) := 0.3950 MAF (4) := 0.2704 MAF (5) := 0.1554 MAF (6) := 0.1104 MAF (7) := 0.2452 MAF (8) := 0.2726 MAF (9) := 0.3385 MAF (10) := 0.2115 MAF (11) := 0.0250 MAF (12) := 0.4555 MAF (13) := 0.2941 MAF (14) := 0.4012 MAF (15) := 0.2360 MAF (16) := 0.3335 MAF (17) := 0.2698 MAF (18) := 0.2879 MAF (19) := 0.3608 MAF (20) := 0.3263 MAF (21) := 0.2923 MAF (22) := 0.1090 MAF (23) := 0.2754 MAF (24) := 0.0917 MAF (25) := 0.1460 MAF (26) := 0.00747 MAF (27) := 0.4722 MAF (28) := 0.0469 MAF (29) := 0.2608 MAF (30) := 0.1745 MAF (31) := 0.0940

%beta1

beta1a := -0.47 beta2a := 1.03 beta (1) := -0.09 beta (2) := -0.08 beta (3) := -0.15 beta (4) := -0.06 beta (5) := 0.17 beta (6) := -0.11 beta (7) := -0.07 beta (8) := 0.23 beta (9) := 0.28 beta (10) := 0.11 beta (11) := 0.14 beta (12) := -0.09 beta (13) := 0.09 beta (14) := 0.09 beta (15) := 0.21 beta (16) := 0.16 beta (17) := -0.07 beta (18) := 0.07 beta (19) := 0.08 beta (20) := -0.2 beta (21) := 0.3 beta (22) := 0.18 beta (23) := -0.11 beta (24) := 0.1 beta (25) := 0.08 beta (26) := -0.23 beta (27) := 0.09 beta (28) := -0.09 beta (29) := 0.18 beta (30) := -0.25 beta (31) := 0.1

var PHS, personPHS, sumPHS : real PHS := 0 personPHS := 0 sumPHS := 0 var binnum, binnumneg,binzeronum : int var bin, binneg : array 1 .. 100 of int

for m : 1 .. 100

   bin (m) := 0
   binneg (m) := 0

end for binnum := 0 binnumneg := 0 binzeronum:=0

for j : 1 .. 1000000

   for i : 1 .. 31

       rand (r1)
       if r1 <= MAF (i) then
           PHS := PHS + beta (i)
       end if

       rand (r2)
       if r2 <= MAF (i) then
           PHS := PHS + beta (i)
       end if

       personPHS := personPHS + PHS  
       PHS := 0

   end for
   % put personPHS
   sumsq := sumsq + (personPHS  - 0.59) ** 2
   % put sumsq
   % put personPHS, " x"
   % if personPHS > 0.1 then
   %     binnum := floor (personPHS * 10)
   %     % put binnum
   %     bin (binnum) := bin (binnum) + 1
   % end if
   % 
    binnum := floor ((personPHS +2) * 10)
    % put binnum
    bin (binnum) := bin (binnum) + 1
   
   
   % if personPHS > 0 and personPHS <= .1 then
   %     binzeronum := binzeronum + 1
   % 
   % end if
   % if personPHS < -0.1 then
   %     binnumneg := floor (-1 * personPHS * 10)
   %     % put binnum
   %     binneg (binnumneg) := bin (binnumneg) + 1
   % end if

   sumPHS := sumPHS + personPHS
   personPHS := 0

end for

% put sumPHS % put sumsq

% for aa : 1 .. 24 % put binneg (25-aa), " ", 25- aa % end for

% put binzeronum, " ", 0

for a : 1 .. 50

   put bin (a)

end for

--J1 (talk) 04:50, 27 March 2017 (UTC)

The science findings are quite preliminary, as the authors themselves state, and they have ~zero clinical utility or credibility at this point. We realize it's exciting to see the headlines and to stoke enthusiasm on forums, but please keep perspective on this. Science by press release isn't science. We will be releasing genetic risk scores on SNPedia and through Promethease when the findings are robust and replicable - which these aren't. To summarize the top 3 things that are needed: (1) AUC statistics showing at least 0.8, preferably 0.85 or higher, (2) an independent replication of the results, (3) incorporation of rare variants (and not just GWAS SNPs). And obviously, it would help if the SNPs involved were actually measured (*not imputed*), and, were SNPs that are on/in either most direct-to-consumer chips or exomes, although in reality, most likely, we'll have to wait for full genome sequence availability in a lot of cases. So, please, stay enthusiastic but keep perspective as well as our "science first" guidelines in mind.Greg (talk) 17:14, 27 March 2017 (UTC)

I am glad that you added this perspective to the conversation, as at least on first read the p values and apparent clinical translation seems plausible. One of the authors of the article below gives some tentative support to the value of their PHS.

http://journals.plos.org/plosmedicine/article/comment?id=10.1371/annotation/3472f94d-a39b-4ab1-aa15-550a8935934e --J1 (talk) 17:58, 27 March 2017 (UTC)

Perhaps a compromise would be to look at the extreme outliers. For example, the 1% or 5% top and bottom of the PHS. The article compared the 10% extremes and arrived at a p value less than 10e-22 with an odds ratio between the top decile and bottom decile of 3.4. A more conservative percentile such as the 1% PHS tails might be better.

Also the article highlights the fact that Alzheimer's is really only a question of when. Above all else this message could be helpful for everyone to plan their future. As it is now 23andme reports that all epsilon 33s have markedly reduced risk for AD. The present article has tentatively shown that this is not correct. In fact, at the highest risk extreme on the PHS, some epsilon 33s would have greater AD risk than epsilon 44s. --J1 (talk) 00:05, 28 March 2017 (UTC)