kissDE diffExpressedVariants Error in nls

Question

[ovidiu]

Dear Community,

I try to use kissDE diffExpressedVariants on a count table from paired end RNA Seq Data. The input file is kissplice type0a format. I have 4 samples with three replicates each. When I try to run diffExpressedVariants I get the following error. Any suggestions are welcome! Thank you!

Error in nls(modelNB, data = event.mean.variance.df, start = list(theta = 100)): step factor 0.000488281 reduced below 'minFactor' of 0.000976562\n An error occured, unable to fit models on data.


head(snp$countsEvents)
       events.names events.length counts1 counts2 counts3 counts4 counts5 counts6 counts7 counts8 counts9 counts10 counts11
1 bcc_99967|Cycle_0            83      50      55      35       0       0       0      77      66      75        0        0
2 bcc_99967|Cycle_0            83       0       0       0      82      88      88       0       0       0       45       33
3 bcc_99898|Cycle_0            83       0       0       0      52      62      59       0       0       0       25       37
4 bcc_99898|Cycle_0            83      52      37      67       0       0       0      49      76      49        0        0
5 bcc_99803|Cycle_0            83       0       0       0     113     121     153       0       0       0       81      122
6 bcc_99803|Cycle_0            83     110      70      81       0       0       0      62      85      82        0        0

  counts12
1        0
2       76
3       39
4        0
5      116
6        0

[ovidiu]

UPDATE

I found out is that the error is caused by the starting value (theta=100) when calling the nls function Letting nls doing the cheap guessing of the starting values solves the problem.

I'm not sure how this changes the results at the end.

[ovidiu]

Dear Vincent,

Thank you for your response and sorry for the delay. Regarding the conditions, yes I knew that kissDE is comparing only two conditions. I modified my code in the way, that I compared all conditions pairwise. Thank you for the hint!

Best, Ovidiu

Answer 1

[vincent.lacroix]

Dear Ovidiu, We double checked on several datasets and using the cheap guessing of the nls function indeed works well as it does not affect the final results. We will likely update the code accordingly in the next release of kissDE. Thanks for pointing this out. In the meantime, other users confronted with a similar bug can safely use Ovidiu's fix.

On the other hand, I notice that your differential analysis concerns 4 conditions.
KissDE has only been tested for comparing 2 conditions. When comparing 4 conditions, KissDE should output variants whose frequency is significantly different in at least one condition compared to the three others. However, we never tested it explicitely. In most of the datasets we used, even when more than 2 conditions were available we were focusing on the difference between 2 conditions (or two groups of conditions). If you are interested in the difference between only 2 conditions out of the 4 you have, you can use the same input data and launch kissDE using the following conditions vector (as explained in Section 2.1.1 of the manual)

myConditions <- c(rep("condition_1", 3), rep("*", 3), , rep("*",3), rep("condition_3",3))

This will compare condition 1 with condition 4.

Best,

Vincent