I have constructed a PWM and want to test its accuracy. Scanning an entire chromosome (chr3, hg18) yields very high false positives (magnitudes higher than the true positives). Scanning the entire genome is much worse.
From some literature and biostars posts, its recommended that I only scan sections of the genome to get an understanding of the accuracy of my PWM. Some posts talked about just scanning the promoter regions with my PWM (or say, a 1000bp neighbourhood around the promoter region).
So, is there a database or some sort of annotation that tells me where the promoter regions are?
I have a feeling that this question may not be well formed because to get "promoter" regions, I need to be talking about genes. Since I am only working with the MEF2 family of transcription factors, is there a way to retrieve promoter regions in which these transcription factors bind?
My idea: Since the TF binds upstream of the gene, usually in a promoter region, AND I know the true binding sites, would it be sufficient to just extract a 2000bp neighbourhood around the binding site?
Here is what my data (only the first couple of lines) looks like:
>chr1:6585537-6585547(-) ------ctatttatag------- >chr1:228916512-228916522(+) ------ctatttatag------- >chr12:50731227-50731237(+) ------ctatttatag------- >chr13:49597702-49597712(+) ------ctatttatag------- >chr14:103058980-103058990(+) ------ctatttatag-------
So again, is it sufficient to take
chr1:6585537-6585547(-), extract a 2000bp neighbourhood around it, and scan this neighbourhood with my PWM. Clearly, I know there is one true binding site and I'll compare it with the number of false postives it gives me.
I have information (coordinates) for a total of 1875 binding sites. From these, only about 140 coordinates to binding sites on chromosome 3. These are my "true binding sites". If I scan chromosome 3 with my PWM, it gets around 85,000 hits.. ie, 85000 sequences that have scored very high that could be potential binding sites. They are, for all intents and purposes, false positives unless they fall (or overlap) the true binding site. Out of 85000 hits, only a very overlap the true binding region.
Now, literature recommends that instead of scanning an entire chromosome (or the entire genome), its better just to scan promoter regions. I guess there is confusion here, because the question can be "promoter regions for what gene?"
But I guess we have bonafide promoter regions if we take say 1000bp neighbourhoods about our true binding sites.