HTseq-count feature type choice for RNAseq

shangzhong0619 · 08-06-2014, 03:15 PM

Hi all,
I am using HTseq-count to count reads in my bam files. I used ncbi mouse annotation file.
I wonder which feature type (third column of gff3 file) I should use?
My understanding of how htseq count reads is that if I choose 'exon', then it will count reads only mapping to exons and sum those up for a gene. If I choose 'gene', it will count all the reads mapping to introns and exons of that gene. Theoretically, for RNAseq I should choose exons, and ignore reads mapping to introns.
In my sample, I know I knock out the gene ext1. I tried both choices and used DEseq2 to do differential expression analysis. In my results which I chose the 'gene' feature, the ext1 gene was the most significant gene. However, in the results for choosing 'exon', there are over 200 genes more significant than ext1. So now I am confused, it seems that the 'gene' feature are better than 'exon'? Anyone has this situation before? thanks.

dpryan · 08-07-2014, 01:44 AM

You'll want to use 'exon', otherwise you increase the noise from pre-mRNAs. I'm not sure why you would be disturbed to find other genes with more significant p-values than the knocked out gene. There are a few different things that affect one's ability to determine variance and the signal level (in this case, how highly expressed a gene is) is one of them. Since you're knocking the gene out, its expression should be very low in some of your samples (I'm assuming a complete knock out rather than just hemizygous), so you'll then have less power there to begin with. If this causes some other decently expressed gene to drastically increase expression then said other gene will probably have a smaller p-value. That's not really a problem.

shangzhong0619 · 08-07-2014, 11:24 AM

Quote:

Originally Posted by dpryan

You'll want to use 'exon', otherwise you increase the noise from pre-mRNAs. I'm not sure why you would be disturbed to find other genes with more significant p-values than the knocked out gene. There are a few different things that affect one's ability to determine variance and the signal level (in this case, how highly expressed a gene is) is one of them. Since you're knocking the gene out, its expression should be very low in some of your samples (I'm assuming a complete knock out rather than just hemizygous), so you'll then have less power there to begin with. If this causes some other decently expressed gene to drastically increase expression then said other gene will probably have a smaller p-value. That's not really a problem.

Hi, dpryan. Thanks for your reply. Actually my knock out is hemizygous, so in this case, it is possible to have genes more significant than the knock out gene?

dpryan · 08-07-2014, 01:29 PM

Sure and for the reasons that I mentioned. Remember that the p-value just relates to believability of the finding given the data. Further, changes in biology are typically non-linear, so a 50% decrease in the knocked out gene could easily lead to larger changes in others. Not to mention that a hemizygous deletion will often not halve the expression.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
htseq-count	paolo.kunder	Bioinformatics	10	10-22-2014 05:45 AM
HTseq-count of zero	halffedelf	RNA Sequencing	14	04-17-2014 11:31 AM
Is my parameter choice for HTSeq-count is right?	wmseq	Bioinformatics	15	11-07-2013 01:24 PM
Which ID should be used for HTSeq-count?	syintel87	Bioinformatics	11	02-07-2013 01:16 AM
multiBamCov or htseq-count to count read per feature ?	NicoBxl	Bioinformatics	1	07-03-2012 03:05 AM

08-06-2014, 03:15 PM	#1
shangzhong0619 Member Location: La Jolla Join Date: Nov 2013 Posts: 17	HTseq-count feature type choice for RNAseq Hi all, I am using HTseq-count to count reads in my bam files. I used ncbi mouse annotation file. I wonder which feature type (third column of gff3 file) I should use? My understanding of how htseq count reads is that if I choose 'exon', then it will count reads only mapping to exons and sum those up for a gene. If I choose 'gene', it will count all the reads mapping to introns and exons of that gene. Theoretically, for RNAseq I should choose exons, and ignore reads mapping to introns. In my sample, I know I knock out the gene ext1. I tried both choices and used DEseq2 to do differential expression analysis. In my results which I chose the 'gene' feature, the ext1 gene was the most significant gene. However, in the results for choosing 'exon', there are over 200 genes more significant than ext1. So now I am confused, it seems that the 'gene' feature are better than 'exon'? Anyone has this situation before? thanks.

08-07-2014, 01:44 AM	#2
dpryan Devon Ryan Location: Freiburg, Germany Join Date: Jul 2011 Posts: 3,480	You'll want to use 'exon', otherwise you increase the noise from pre-mRNAs. I'm not sure why you would be disturbed to find other genes with more significant p-values than the knocked out gene. There are a few different things that affect one's ability to determine variance and the signal level (in this case, how highly expressed a gene is) is one of them. Since you're knocking the gene out, its expression should be very low in some of your samples (I'm assuming a complete knock out rather than just hemizygous), so you'll then have less power there to begin with. If this causes some other decently expressed gene to drastically increase expression then said other gene will probably have a smaller p-value. That's not really a problem.

08-07-2014, 01:29 PM	#4
dpryan Devon Ryan Location: Freiburg, Germany Join Date: Jul 2011 Posts: 3,480	Sure and for the reasons that I mentioned. Remember that the p-value just relates to believability of the finding given the data. Further, changes in biology are typically non-linear, so a 50% decrease in the knocked out gene could easily lead to larger changes in others. Not to mention that a hemizygous deletion will often not halve the expression.