I have done shotgun proteomics experiment with a defined number of samples. Now, I have to validate the shortlisted proteins.
I had a query that whether the validation to be initially conducted in samples (used in discovery phase) before going for other samples of cohort or not and if has to be performed, then those samples be counted in total count of validation samples?
The basic statistical issue here is whether or not you meet a statistically-valid confidence threshold. This depends on the analysis technique you plan to employ for the analysis. You have generated a series of potential markers (using Beysian statistical methods, like PCA) through an open study with too-few samples (one that does not meet a statistical validation threshold). If you have 10 putative biomarkers and want to employ a t- or F-test, you basically will need about 10 x 80 (880) independent patient samples (including 80 controls) for validation of these biomarkers. If you want to examine the cross products of these 10 biomarkers, you will need way more. 80 is a general rule of thumb, the actual number depends on the POWER of the test post-acquisition of the data. Sorry, but the data drives the answer, not the other way around, which is why you see the FDA requiring more clinical data for NDAs. The more overlap of the biomarker concentration distributions between the control and target group, the more patient samples will be required to reach statistical discrimination. If you are going to use an ROC curve for the analysis you may need more. Sit down with a real statistician BEFORE you start you study and write out the protocol that you will be employing to see if it makes sense. You can apply some estimates of the biomarker concentration distributions from your open study to see what sample load will be required to validate each of the markers.
In terms of re-using the open study data, in your validation set, this also depends on the analysis technique. AI methods generally frown on this since in AI, the system trains itself and you don't really know what it decided was important. This is generally also true in multivariate cluster analysis methods. In more block design (hypothesis-driven) studies it is more acceptable to include the preliminary data in the final set (i.e., the FDA requiring the collection of additional clinical trial data). In this case, the same hypothesis is applied to both sets of data, so both sets can be included. Think about "metastudies" that pull data from multiple smaller clinical data sets. And, no, you can't just say I'm testing the same 80 samples for each of ten biomarkers independently. You only get one hypothesis for each set of 80 samples, so pick it well.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics