We have a proteomic dataset where many proteins have been identified in samples derived from two classes of samples with three biological replicates each. Each of these samples was analysed (injected into an LCMS instrument) twice, generating a technical duplicate (this means, that we have 12 LCMS results in total, six sets of duplicates). We have used a technique (SWATH) where we have essentially no missing values for any of the proteins.
When playing with the data (e.g. making a x-y graph where two different replicates are compared to each other), it is clear that (as expected) technical noise is much lower than biological noise (differences in protein abundances between different sample sources).
Can anybody advise me how this observation could be quantified? As we only have duplicates for each sample, we cannot work with standard errors when trying to capture technical variability. If we just take the difference of a value between technical duplicates, we neglect that we have duplicates for EACH sample. What does the average of the six differences tell us and how can this be related to the average of three means in each class?
1. See if the software such as Skyline that accounts for both technical and biological replicates
2. If the technical replicates are equal (or appear to be so to your instinct), then discard one of the replicates. You have only three biological replicates for each group because technical replicates provide no real information about the treatment or method but you care only about the biological replicates.
3. If you want to still include the replicates then there try to include, as there are six samples per group
4. If you don’t have any technical variability or your biological samples are similar then one of the injection in a group should be equal to the other 5 samples in a one sample t test.
5. Apart from any statistics, use the leave-one-out method to remove any one of the six mass spectra (per sample) to know if just leaving one injection can provide you a different information this will rule out any black sheep.
6. the average of six injection must be similar to the samples (three biological replicates)
I hope this helps.
19
The technical variability question can be perhaps tackled by asking the question of how well two measures of the same property agree with each other. Since the two measurements are interchangeable, this would be a type I intraclass correlation.
In terms of analysis, whether you have missing data or not, you can take technical variability into account by using variance estimates based on the clustering of data caused by having two measurements on each entity. In an analysis, the technical variability isn't a nuisance but a bonus, giving you two measures of each property and thus reducing measurement error. However, in an analysis, the focus will be on predictor factors with the requirement to adjust variances for the duplicate measurements rather than to quantify it. (You can, of course, quantify it, but I think it's best measured by the level of agreement, as above, rather than by the effect of the clustering on estimates of variation.)
Hi Achim, difficult one since you cannot really tell the degree of technical variation form a sample set like this. For this you would still need to run more technical replicates, even if it is only on a reference sample that you run side-by-side with the actual samples. This would allow you to quantify technical variance better.
I like to use non-directed PCA to visualize if biological variance is bigger than technical variance. Since PCA is a weighted approach (larger values, larger impact) it is not a complete measure e.g. of what happens at lower abundances, and it is not really "quantitative", but it still is a good indicator.
Hi Achim
Interesting question. Obviously the smaller the biological difference the more important the control of technical variability becomes.
Just some random thoughts.
Are you wanting to quantify the proteins relative to each other or are you looking for something that must change (other than concentration) between the biological sets?
If you elected a particular peptide in your data set that is generally abundant and compared the data for just that peptide between the replicates, then do the same for a low abundance peptide you may get some idea of whether the technical variability is related to abundance or not.
If the variability is the same between these peptides of different abundance you could assess whether two closely eluting peptides show similar variability between the duplicates and again for peptides eluting at quite different retention times from each other.
The problem is that there is no universal type of internal standard that can be added to remove some of the instrument and sample handling variability but the earlier (so that all subsequent manipulations are performed in unison) the method of comparison can be made or introduced the smaller the technical variability would be.
- Badges
- Science topic
- Similar topics
- Chemistry
- General Biochemistry
- Proteomics