I am running cufflink on 0.9M paired end sam data of human RNA and gtf annotations.
The GTF annotations were constructed based on the Ensembl, UCSC gene and refseq dataset downloaded from UCSD genome browser. There are 2.0M lines in the resulted GTF file.
I didn't generate the read alignments from tophat. Rather, I used blat and bowtie to align and pair the reads, and inferred the XS:A field from the gene annotations and the splicing signals for all the reads. I also included the SQ header in the result.
The way I ran cufflink is cufflinks-0.9.3.Linuxx8664/cufflinks --num-importance-samples 2000 --max-mle-iterations 10000 -v -G transcript.gtf -r hg19/all.chr.fa -N -o chrX chrX.sam.true.sort
Strangely, I ran cufflink multiple times on the same chrX data, and the estimated FPKM of some isoforms could be quite different. I observed for some genes, the ITERMAXs were not big enough for convergence, and so I increased the corresponding parameters, but the results still varied a lot.
I understand the MCMC in cufflink is a random process which may have different results depending on the initial state, but I thought that the latest version of cufflink stabilizes its result somehow, and so I wonder whether there is anything wrong about the way I prepared the dataset and ran cufflink?
I have heard similar stories of weirdness with cufflinks pre-1.0. Specifically though I heard that the numbers changed significantly when it was re-ran between versions pre-1.0. I do not know if the person I talked to tried it twice with the same version though. I would upgrade and try it again twice to see if you get the same issues with changing FPKM values.