Hi,
Lets give a vivid example to explain my question: suppose we have a list of universal genes including both oncogenes (OG) and tumor suppressor (TS), and the purpose is to test the enrichment of conditions when we select some genes by a method (which we want to test its randomness or not). Here, the "method" refers to a highly selective approach which is capable of identifying the cancerous behavior, i.e. the so-called method tries to identify as many oncogenes as possible and tries to avoid the selection of TS. Here, a possible way to analyze the predictive ability of this method is to employ hypergeometric p-value; for oncogenes, this would be perfect. However, for TS, it is a bit misconducting (at least for me), since as the "number of successes in the sample" drops, the p-value increases. Hence, lower number of TS in a sample (which clearly shows the high predictive power of the method), wrongly indicates insignificant hypergeometric p-values.
So, my question is: how to reconcile this situation and modify this problem? Is it reasonable to report 1 - calculated_p_value ?
Thanks in advance.
I analyzed many cancer genes by t-test. It is not helpful because of the outliers.
See my papers on RG.
Dear Raid,
Yes, both OG and TS are present in the sample because it is a set of cancer-related genes. So, we have a mixture of cancer genes (population sizes 1496), which 700 of them are TS (number of success in population) - and the others may be OG or genes with loss of function mutations. We have a presumed algorithm which claims that has the capability of choosing a cancer-related behavior (which should have as many OG as possible with lowest number of TS - since they are absent/not-active in cancer cells). So, if we pick a sample consisting of 1080 genes, lets say it contains 20 TS (success in the sample size), which clearly shows algorithm's "high prediction power". To examine the randomness of algorithm's selection method, we can calculate hypergeometric test. But, of course the output p-value is 1 (Probability of drawing 20 successes (TS) or more from a sample of 1080). This is due to the fact that "lower" number of TS (successes) in the picked sample shows the higher predictive power as opposed to OG selections (higher better). Hence, how can I assess the significance of my so-called algorithm by hypergeometric test in this condition for TS?
Best
- Badges
- Science topic