FAQ

1. Range of LS score?
   
   • As explained in our papers, the LS score itself is not bounded between -1 to 1. With in one dataset, one way to compare the relative strength of the association is to standardize it by dividing all LS scores by the absolute value of the max/min of all scores, such that the score become -1 and 1 bounded. The statistical significance of the association is however not affected. To compare aross datasets is more tricky. One potential way is to transform the LS scores into qunatiles with signs unchanged, such that it is between -1 and 1. Then the strength can be compared in terms of the association's relative strength amon datasets.
2. How the fillMethod works?
   
   • For fillMethod in eLSA:
     none: keep na until the data is normalized, then fill it up with 0
     others: pass the fillMethod to the interpld function in scipy,
     see "http://docs.scipy.org/doc/scipy/reference/tutorial/interpolate.html" for how it works.
     
3. Why tab delimited file produced by Excel not accepted by eLSA program?
   
   • Most likely it is an issue due to different carriage return character across platforms. Two methods to go:
     install dos2unix and convert the input file to unix format
     issue an '%s/^M/\r/g' in VI and save as a new file
     
   • Update: "check_data" is now available for checking your input format consistency with eLSA expectation.
4. The alogrithm has been running about 5 days now - it does appear to be runnning smoothly - i submitted the input file i sent you i.e 715 factors, 39 samples would you suggest cutting this file down? shouldl the program run this long?
   also - how does the program work ? would it go faster if i allocate more memory and can it be run on parallel processors?
   
   • The program is currently running single process. It would be helpful if you can run it natively from Linux OS with a faster processor. Memory is typically not a concern. How many permutations (P) you put? What is the bootstrap number (B). Time ~ (B+P)*F^2*N. For 715 factors (F) and 39 samples (N) at P=1000 it probably will take a while, depends on your processor speed. One way is you can reduce either B, P. Another way is to use the bleeding version of the LSA program, which provides basic parallel support. You can download the bleeding release from the source code site and install it on your blast. Subset your input data into, say 10, trunk files and run eLSA on each pair (i,j: 0<i<=j<10) of trunk files in the upper triangle (so i<=j) and specify the parameter to tell the program if the pair is on diagonal line (i==j). The you can concatenate all the result files to get the whole result. This probably can be automated by scripting, which I haven't got time to write yet. One thing is that, you have to recalculated all the Q-values after you get all the results. You can collect all the P-values and using the R package Qvalue to do it.
     
   • Update: My suggestions are reduce the precision and OTU number and try to get a hold of how long it takes to run your analysis using the formula provided above before serious jobs. There are also "par_ana" and "fix_qv" commands available in the bleeding version which divide the jobs and run them parallel if you have a PBS system and access to multiple nodes.
5. problem with q-values
   
   • Sometimes the input set of p-values cannot be used to estimate q-values successfully using the default storeyQvalue function implementation using the interpolation core from scipy. The same might also be found using R's qvalue package. This is because, in some cases, all the pairwise p-values are quite small, a setting where the assumption of storeyQvalue and R's qvalue - the majority p-values is non-significant for q-values so it can be used to estimate the uniform rate does not hold. However, q-values can be calculated in different ways, while the storeyQvalue and R's qvalue package all basically implemented J.D.Storey's method in Statistical significance for genomewide studies PNAS 2003. There are other methods available but not implemented in the eLSA package. A few of them are listed at eLSA's homepage http:meta.usc.edu/softs/lsa . The current trade-off is to continue on LSA computation without the q-values (putting na in). If needed, p-values can be collected and sent to other tools for q-values calculation. The new version is pushed to bitbucket and tagged as bleeding.
     
   • Note important if python q-value is not working most likely R's qvalue package will not work either. The only difference between them is the interpolation function used. One is from scipy and the other is from R lib.
6. What is "-p mix"? or what is the mixed approach for p-value estimation?
   
   • "mix" is a combination of theoretical and permutation methods. Theoretical p-values are not perfectly linearly correlated to p values of permutation methods, thus it is true that "mix" p-values are not linearly correlated to permutation methods either. "mix" method saves your time by employing some "intelligence" - only passing into permutation evaluation of Pperm when the Ptheo is significant (<0.05). The p-values from this method are also trusted, and the conclusion you get based on the this method should be correct. If you still bother the difference between "mix" and permutation-only, you can always use the permutation-only method.
7. How to cut off using "P" and "Q" values?
   • Cutting off on q-value is basically cutting off on p-value with a different strategy and an easier explanation. And if you apply both P and Q cutoffs, you get the combined filtering effect: less than probability P to occur by random and with probability Q to be false positive. Basically q values are the transformation of p values. So they are telling the same story, but with different perspective and interpretations. You can use either P values or Q values or a combination of them, whichever you think fits your needs best. If you don't know what to choose, read relevant papers, and find out the conventional cut-offs they use and have a try yourself. For more information about P and Q values, see Storey et al. Statistical significance for genomewide studies PNAS 2003.
8. Have install eror with rpy2 related problem?
   • rpy2 is no longer a dependency.
   • If you have previously installed eLSA, completely remove it !!! (find it at your python site-packages)
   • Git clone the latest "master" branch or download from https://bitbucket.org/charade/elsa/downloads (Go for Tags version >v1.0.4)
   • Install ELSA: python setup.py install
   • Try lsa_compute see if it works
9. How many time points are needed to use fast theoretical p-values
   • you can use theoretical p-values with time points >10 with no delay and >20 with delay. Please refer to Xia et al. Efficient statistical significance approximation for local similarity analysis of high-throughput time series data Bioinformatics 29(2):230-237 for details
10. How to choose among normalization schemes?
    • we recommend the percentile normalization + Z-score normalization (i.e. percentileZ option in the command line) unless you have other concerns. Percentile is simply rank normalization. PercentileZ applies a Z-score transformation after Percentile normalization (this is described in recent papers). robustZ uses the median instead of mean and mad instead of standard deviation for Z-score, otherwise the same as PercentileZ. pnz is percentile normalization without zeros.