Tweet
Hello,
I have been exploring which transcript quantification method to use for isoform level analysis. To date I have been performing gene level analysis with a STAR-htseqcounts-edgeR pipeline. However, I am now interested in quantifying isoform levels. To this end, I have mapped with STAR as before, but allowed the maximum number of reported mappings and reported out the transcriptome covering reads with the option available in the latest version of STAR. I have then used the default and appropriate read orientation settings for RSEM, salmon, and eXpress to see how much these tools agree with a single library of about 130million paired end human reads. When I compare the Salmon NumReads with either eXpress effective counts or RSEM estimated counts I find very good agreement with high Pearson correlation, (Salmon/RSEM read count figure attached) Although RSEM and Salmon are in much greater agreement linearly (Pearson correlation ~0.97 (RSEM/Salmon) vs ~0.70 (RSEM/eXpress, Salmon/eXpress), while Salmon reads are monotonically closer to eXpress values than RSEM ones (Spearman correlation ~0.92 (Salmon/eXpress) vs ~0.82 (Salmon/RSEM). However, it seems with Salmon and RSEM that both the FPKM, and more importantly, TPM values reported (another figure attached) disagree in a way that appears to be largely dependent on transcript length. Is there something in either model that significantly changes the way TPM is calculated to bias the transcript counts based on length when the read approximations produced by the two models are so similar? I wasn't able to identify such a bias from the documentation and literature.
I should say I do notice that RSEM tends to estimate much higher TPM than either eXpress or Salmon for transcripts around 100-200bp (read length is 51bp); however, the agreement with eXpress TPM values for anything larger than this is profound - while the Salmon and RSEM values diverge as a function of decreasing transcript length as seen in the second attached figure. Any discussion would be much appreciated.
Warm regards,
David
I have been exploring which transcript quantification method to use for isoform level analysis. To date I have been performing gene level analysis with a STAR-htseqcounts-edgeR pipeline. However, I am now interested in quantifying isoform levels. To this end, I have mapped with STAR as before, but allowed the maximum number of reported mappings and reported out the transcriptome covering reads with the option available in the latest version of STAR. I have then used the default and appropriate read orientation settings for RSEM, salmon, and eXpress to see how much these tools agree with a single library of about 130million paired end human reads. When I compare the Salmon NumReads with either eXpress effective counts or RSEM estimated counts I find very good agreement with high Pearson correlation, (Salmon/RSEM read count figure attached) Although RSEM and Salmon are in much greater agreement linearly (Pearson correlation ~0.97 (RSEM/Salmon) vs ~0.70 (RSEM/eXpress, Salmon/eXpress), while Salmon reads are monotonically closer to eXpress values than RSEM ones (Spearman correlation ~0.92 (Salmon/eXpress) vs ~0.82 (Salmon/RSEM). However, it seems with Salmon and RSEM that both the FPKM, and more importantly, TPM values reported (another figure attached) disagree in a way that appears to be largely dependent on transcript length. Is there something in either model that significantly changes the way TPM is calculated to bias the transcript counts based on length when the read approximations produced by the two models are so similar? I wasn't able to identify such a bias from the documentation and literature.
I should say I do notice that RSEM tends to estimate much higher TPM than either eXpress or Salmon for transcripts around 100-200bp (read length is 51bp); however, the agreement with eXpress TPM values for anything larger than this is profound - while the Salmon and RSEM values diverge as a function of decreasing transcript length as seen in the second attached figure. Any discussion would be much appreciated.
Warm regards,
David
Comment