Dear to whom it may concern,
I would like to ask people who are interested in univariate analysis in metabolomics. Now, I am proceeding my metabolomics data using univariare analysis, namely p-values and FDR-adjusted p-values.
However, as far as I know, the calculation of a p-value for each feature depends on two factors: (a) distribution of each feature and (b) variance of each feature between case and control group. To be more specific, the first step is that we need to apply a statistical tool (I do not know which tool can help me to check this issue) to check whether one examined feature is normally distributed in both these groups or in only one of them, and of course, there are two scenarios as follows:
1. If this feature is normally distributed in both these group, we proceed to use F-test as a parametric test to check whether the variance of this feature in both these groups is equal or unequal. If it is equal, we can do a t-test assuming equal variance, otherwise, a t-test with unequal variance must be taken into account.
2. If not, a non-parametric test will be applied to obtain a p-value for this feature. In this case, may you please show me which tests are considered as non-parametric tests?
I am unsure that what I mention above is right because I am a beginner in metabolomics. In case, this procedure is right, that means that each feature will be processed under this step by step one to obtain a p-value because all features are expressed differently in the distribution and variance way between these groups (case and control).
I hope that you may spend a little time correcting my idea and give me some suggestions in this promising field.
Thank you so much.
Pham Quynh Khoa.
Hello, first of all, what's the sample amount in each group? In general, I suggest to test non-parametric if n < 6. For smaller sample size it is not really serious to assume parametric conditions, independently from test results. As non-parametric alternative for t-test, I suggest u-test (Mann-Whitney-Test). Best regards, Max
Hello Quynh Khoa Pham and Maximilian Blum ,
I find this question very interesting as I like to use univariate methods with metabolomics data. I am familiar with PCA, PLS, PLS-DA, OPLS (multivariate) and t-testing/Volcano plots (univariate). Within the multivariate methods I think that it makes a fundamental difference if they are used for making predicitions on unknown samples or to identify relevant variables: Predictions can be convincingly evaluated by external test set validation, RMSEP, Q2, or scramble test, whereas I have some concerns for using these evaluation approaches to validate the significane of identified relevant features. Thus I prefer the univariate before multivariate approach for identifying relevant features for reasons of simplicity, at least when comparing two groups. I suppose that your problem is about identifying variables, because the univariate approach cannot make predictions on unknown samples.
Either way, an important consideration is how to finally evaluate the features that appeared important. For example, if a feature clearly prevails in one group it is not clear if this is because of an increased expression or because it has not been degraded or because of something completely different. Consequently I would argue that the outcome of any variable selection based on metabolomic data can only be a hypothesis that would require further experimental confirmation anyway. This unfortunately long text should justify a comparatively relaxed evaluation of the p-values achieved from univariate analysis of metabolomic data. I would say that your careful considerations on normal distribution and homogeneity of variance are appropriate when experimentally confirming the rather hypothetical results of the metabolomic approach. They appear exaggerated for the metabolomic data themself.
I have particular concerns on the usefulness of the FDR-adjusted p-values. I work with LCMS data: the feature detection settings have a strong impact on the number of detected features and in turn the number of detected features strongly impacts on the FDR-adjusted p-values. As a consequence, the FDR-adjusted p-values may become absurdly strict with sensitive settings. I believe that it would be necessary to strongly limit the number of included features before calculating the FDR-adjusted p-values. Maybe limited to only those needed to substantiate a particular conclusion. I am happy about corrections when this consideration is wrong.
- Badges
- Science method
- Similar topics
- Chemistry
- General Biochemistry
- Biochemical Techniques
- Analytical Chemistry