4 ms·
It is very easy to find correlations between phenotypic traits (getting MS, developing cancer, autism etc) and "your favourite biological variable". See John Io
by Gatsky 9y ago
It is very easy to find correlations between phenotypic traits (getting MS, developing cancer, autism etc) and "your favourite biological variable". See John Ioannidis talking about microRNA studies for example. Part of the reason for this is that high throughput data like gene expression measures, gene methylation or microbiome sequencing gives you p << n data which invariably is also low dimensional ie a few factors explain most of thr variation. It is therefore easy to find a variable or a 'signature' which correlates with one of these explanatory factors and the phenotype, but doesn't tell you anything much about what is going on, or how to treat the disease. But it does allow thousands of papers to be published with pvalues <0.05.
Not to downplay this data, but caution is required not to over interpret the results, and avoid making the sane mistakes over and over again.
- _of 9y agoResearchers in genomics are aware of this and that is why researchers are expected to correct their p-values for multiple testing, for example using q-values.
- Gatsky 9y agoNo that isn't the point. These variables are often associated with an outcome or phenotype in a statistically rigorous manner. The point is that knowing this doesn't get us anywhere.
- _of 9y agoGet us anywhere? That is research, small pieces to the larger puzzle. Biology is complex, a lot of molecules doing different things. We don't have perfect tools to study those molecules or their interactions. Genomics and statistics offer one path to their study, and they complement already existing methods.
- Gatsky 9y agoI don't understand what you are trying to say. My point is also that the tools are imperfect, particularly when it comes to generating useful insights, as opposed to just correlations.
- _of 9y agoIf your tools don't generate useful insights, then you are not using them right. A properly designed study, employing high-dimensional genomics data, can for sure generate mechanistic insights, not just correlations. It's all about your study design.
- Gatsky 9y agoSome tools will never give you useful insights because they have inherent limitations, or the model being studied doesn't lend itself to an informative study design. Results get published anyway. Which is fine. My point was that many high throughput studies do not materially contribute to the advancement of knowledge. This is entirely compatible with your point that well designed high throughput studies can contribute.