Hi all, I'd appreciate any thoughts that will help us with the experimental design for our project.
Our experimental design involves collecting paired samples (case-control) for multiple patients. We use RNA-Seq data to measure differential expression. Getting biological replicates is difficult and at present we just have the 'typical' unreplicated paired design:
Patient 1 case, Patient 1 control
Patient 2 case, Patient 2 control
etc.
In our case, we really want to focus on individual-level differences. However, without biological replicates we can't compute any robust statistics describing the statistical significance of the findings in a direct n=1 case-control comparison.
We can carry out the analysis using edgeR (I'm sure other packages such as deseq2 also support a similar process - it's just that edgeR has a good section in the user guide about this scenario). We first pool our samples over all patients to compute the coefficient of variation (either genewise or common) with just case and control groups. This gives us an estimate based on a number of 'effective biological replicates', then we can use this estimate to compute p values in the individual comparisons.
Those p values reflect the probability of observing the single paired sample difference in expression, based on a background model constructed from a pool of all pairs.
This isn't ideal, I know. But would you, as a reviewer, accept a paper describing that approach to DGE analysis?
Are you aware of any studies using this approach, or analysing the effectiveness?
By my reckoning, this will be a *conservative* approach, since it is likely to overestimate the genewise variance when a reasonable number of patients is included, on the assumption that variation between patients is greater than variation between biological replicates. Therefore we are more likely to have an issue with false negatives, rather than false positives. Does this help to justify it?
Thanks for any thoughts.
Our experimental design involves collecting paired samples (case-control) for multiple patients. We use RNA-Seq data to measure differential expression. Getting biological replicates is difficult and at present we just have the 'typical' unreplicated paired design:
Patient 1 case, Patient 1 control
Patient 2 case, Patient 2 control
etc.
In our case, we really want to focus on individual-level differences. However, without biological replicates we can't compute any robust statistics describing the statistical significance of the findings in a direct n=1 case-control comparison.
We can carry out the analysis using edgeR (I'm sure other packages such as deseq2 also support a similar process - it's just that edgeR has a good section in the user guide about this scenario). We first pool our samples over all patients to compute the coefficient of variation (either genewise or common) with just case and control groups. This gives us an estimate based on a number of 'effective biological replicates', then we can use this estimate to compute p values in the individual comparisons.
Those p values reflect the probability of observing the single paired sample difference in expression, based on a background model constructed from a pool of all pairs.
This isn't ideal, I know. But would you, as a reviewer, accept a paper describing that approach to DGE analysis?
Are you aware of any studies using this approach, or analysing the effectiveness?
By my reckoning, this will be a *conservative* approach, since it is likely to overestimate the genewise variance when a reasonable number of patients is included, on the assumption that variation between patients is greater than variation between biological replicates. Therefore we are more likely to have an issue with false negatives, rather than false positives. Does this help to justify it?
Thanks for any thoughts.