We are relatively new to RNAseq analysis and are seeing unusual results in our cuffdiff output, where a number of our genes are reporting “-Nan” for one treatment FPKM in the final output, while the FPKM values are seemingly fine at all other steps of the analysis. We are concerned that this may be an error, as these Nan values are not associated with the same animal across a number of different analyses (animal A is problem in comparison 1, but is fine in comparison 2 where animal 2 is now a problem) nor are they consistent when the analysis is repeated, despite the constant presence (approximately 2/3 of genes reporting Nan are consistent between repeated analyses, while 1/3 are unique). I’ve provided some additional details of our analysis below.
This study involves 65 free-ranging animals from multiple locations, seasons, sex, etc. Sequencing was performed by Illumnia HiSeq, approximately 25M 100 base single-end reads per animal.
Analyses have been performed using the Tuxedo Suite of tools in a local instance of Galaxy.
Following grooming and trimming, TopHat v 2.0.9 (Galaxy tool v 0.7) was used to map the reads from each individual to the species-specific Ensembl genome.
Cufflinks v 2.2.1 (Galaxy tool v 188.8.131.52) was then run using reference annotation, bias correction, multi read correct and cufflinks effective length correction. The FPKM values for each animal at this point in the analysis seem fine, no “-Nan” values are present in the output.
Cufflinks gtfs from all individuals were combined using Cuffmerge v 2.2.1 (Galaxy tool v 184.108.40.206) using reference annotation and sequence data.
Multiple separate cuffdiff v 2.2.1 (Galaxy tool v 220.127.116.11) analyses were run (location, season, sex, etc.) using the cuffmerge gtf file and tophat generated bam files for each animal. The cuffdiff parameters used included, geometric library normalization, pooled dispersion estimation, 0.05 FDR, Min alignment count 10, multi-read correct, bias correction and cufflinks effective length correction.
In the differential expression testing and FPKM tracking output files (both gene and transcript level), one treatment will have “-Nan” for a number (5-10%) of features, resulting in a no test status for DE. From the read group tracking file we can see that this is due to a single animal from the treatment group reporting “-Nan” FPKM values; however when looking at the cufflinks output, the FPKM value there seems fine and is well within the range of FPKMs reported for other animals.
All 65 animals are used in each cuffdiff analysis, but are assigned to different groups depending on the variable tested. As such the n for each group is different, but is always >10. The cuffmerge gtf and tophat bam files used in each cuffdiff comparison are the same. However, different animals report –Nan in the various cufdiff analyses and if we repeat an analysis, the Nan values are still present and in the same animal, but the genes reporting these are different.
I apologize for the longwinded explanation. If anyone has seen similar results and can provide any insight to help us determine if these are random errors or legitimate biological results we appreciate your response.