Hello Adriana,

typically low read counts are much noisier than high read counts, i.e. in replicates their variance is bigger in comparison to their mean (See for example http://genomebiology.com/2010/11/10/r106 for a paper that describes this behaviour of reads).

Thus, the fold-changes of genes with few reads have a high variance and can be easily very high just by chance (e.g. 1 read vs. 10 reads) without meaning necessarily that the gene expression has changed.

The fold changes, however, can also be caused by genes that have a low abundance in one condition but are completely silenced in another condition, which one would like to detect as differentially expressed.

To properly say whether a change in read counts between two condition is is observed just by chance or reflects a true change in gene expression, an approach that is reliable is to first model the read distributions. The modelled reads distributions can then be used toidentify statistically significant differential gene expression. This approach is taken by tools such as DESeq or edgeR.

Tools that do not perform the modelling of the read distributions try avoid false calls of differential expression by not looking looking at genes with a low expression (e.g. by requiring a minimum number of reads for genes). This means, however, that these tools cannot detect differential gene expression for lowly expressed genes.

Hello Adriana,

Such few reads usually represent transcriptional noise. With so few reads it is difficult to ascertain whether these are truly actual instances of expression or randomly present due to sequencing errors. This can easily be ascertained by having multiple replicates of the same sample. If these reads are present in one, but not in other replicates then they most probably are transcriptional noise. We used this approach to ask the same question of small RNA expression library. To answer this question we used K-S statistics applied to frequency distribution plots of replicates. The rationale here was that the cut off should be the lowest number of reads where the distance between the two frequency distribution  plots of replicates should be minimized. This was a valid assumption, since if the replicates were exactly the same the frequency distribution curves would overlap and the distance would be zero. Thus we defined noise as the minimum number of reads that prevents frequency distribution curves of replicates being close to each other. I am attaching the paper that describes this in more detail.

Article Analysis of deep sequencing microRNA expression profile from...

Hello Adriana,

typically low read counts are much noisier than high read counts, i.e. in replicates their variance is bigger in comparison to their mean (See for example http://genomebiology.com/2010/11/10/r106 for a paper that describes this behaviour of reads).

Thus, the fold-changes of genes with few reads have a high variance and can be easily very high just by chance (e.g. 1 read vs. 10 reads) without meaning necessarily that the gene expression has changed.

The fold changes, however, can also be caused by genes that have a low abundance in one condition but are completely silenced in another condition, which one would like to detect as differentially expressed.

To properly say whether a change in read counts between two condition is is observed just by chance or reflects a true change in gene expression, an approach that is reliable is to first model the read distributions. The modelled reads distributions can then be used toidentify statistically significant differential gene expression. This approach is taken by tools such as DESeq or edgeR.

Tools that do not perform the modelling of the read distributions try avoid false calls of differential expression by not looking looking at genes with a low expression (e.g. by requiring a minimum number of reads for genes). This means, however, that these tools cannot detect differential gene expression for lowly expressed genes.

Hello Bronwyn,

As i know, you can call Corset with the option -m :Filter out any transcripts with fewer than this many reads aligning. Default: 10

Considering the background as explained by Philipp Drewe, you shouldn't set this value too low, depending on the quality of your data.

Another approach is to use tools which sort of assume that everything is expressed just not sequenced deeply enough and thus assign a prior. An example is BitSeq (Bayesian Inference of Transcripts from Sequencing Data) - doi:10.1093/bioinformatics/bts260.

The general discussion about influence of low expressed transcripts/genes on DE analysis you can also find in some of the Web Collection of MAQC consortium SEQC project papers: http://www.nature.com/nbt/collections/seqc/index.html

When setting arbitrary thresholds we have to be very careful that we're not introducing bias in our data. I strongly dispute the comments that low counts are simply 'noise'.

Read counts for a gene are a function of expression and length. A 10kb transcript with 10 reads is very different from a 200bp transcript with 10 reads: it's a 50-fold difference in expression. In using a simple cut-off they are being treated the same, which may be a problem in your experiment.

I agree with Thomas in that cut-offs, if any, should be performed across *all* samples not per sample. So, if a gene's average expression is ~3 reads in all samples, then it's unlikely to be informative.

I would set the threshold much lower, or 'off' even, and then analyse the cufflinks results at the end keeping in mind the average expression. Any 'significant' genes with low expression should be treated with caution.

Thank you all very much for the insight and links! This has been incredibly helpful information.