Gene expression is visually different between two samples (5 replicates per sample for 10 total, only duplicates are shown marked by red line). (See attached). The TPM for each is also different 0.43 vs 46. However, the bioinformatics specialist did not include this gene in their final gene list. They used DE-Seq and I used CLC. I understand DE-Seq is more robust but the difference in expression is so apparent both by read density and after TPM normalization.
Is there a likely explanation for the apparent discrepancy, or does anyone have any experience with this?
Are you sure they used TPM? For DGE analysis in with DESeq, it seems a weird choice, as the normalization method of choice, should be a median of ratios that takes into account the RNA composition of the library. TPM accounts for sequencing depth and gene length but fails to include RNA composition, therefore being inappropriate for DGE. Another good choice would be TMM from EdgeR.
Ireneusz Stolarek They did not use TPM for DESeq2. Perhaps the better question: is there an explanation as to why DESeq2 would fail to detect strongly repressed genes (eg like those pictured above), while CLC is able to identify them.
The individual who used DESeq2 offered this initial response to this question (see below)
Briefly, reads are counted per gene, p-value is calculated adjusting for the negative binomial distribution of reads from RNA-seq, then, adjusted p-value is calculated using Benjamini-Hochberg's FDR method, then independent hypothesis weighting is applied.
The cases could occur that potential de-regulated genes could be dropped out. You may check them manually, then, q-PCR to confirm it. The goal of RNA-seq is to find out obviously de-regulated genes screening 20,000 genes.
Right, I don't know how you selected the final set, but usually, the output of DESeq is further filtered for the log2fold change (the strength of the effect) and the statistical significance of the observed effect.
The log2fold change cut off most often is selected as at least 2. Now the candidate genes are further selected for the adjusted p-value of (most often)
Ireneusz Stolarek This specific gene was 30 fold difference and fdr of 0.001 respectively using CLC. In DESeq2 it did not appear to make the cut off of log2fold > 2 and fdr
Guess another extensive benchmark might show putative biases in the current methods. Did you also check manually this gene in each of the replicates? Looking at your picture I assume the top part is one sample, bottom another. I would exclude this gene from the analysis completely, as it's barely even covered in one sample. So this gene might be lost during the QC in DESeq.
But then again, what if the gene is shut down completely. If it's a transcription factor you could look at its whole regulon with GENIE3 to see if the whole thing gets deregulated (or look for its corresponding TF and characterize the regulon this way)
Ireneusz Stolarek Yes, I did check replicates they are consistent for all 10 samples with 5 replicates per test strain. Interestingly this gene is stage specific and the open source fastq data supports the observation that this gene is almost undetectable at this stage. So, I feel relatively confident that the image is representative of the overall trend. I am just surprised that this would go undetected using DESeq2 which is considered the premier program for DEG. More so, given the fact that CLC is commercial in nature and therefore somewhat looked down upon by experienced bioinformaticians, at least in my limited experience.
"Ignorance is a bliss". Well if you are giving someone a task and then find an obvious mistake, they should fix it, and it doesn't matter which software. I think I will run some tests with DESeq to see how this problem emerges, it seems very disturbing.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS