Dinh/Dinh 2013/NOTES/2013-1-16
Jump to navigation
Jump to search
N37 - Genome segmentation
- Since the previous models did not work well for genome segmentation, we should re-do our assumptions
- Based on previous high-coverage data, we see that LMR sites are clustered together. This means that we can and should perform smoothing.
- Our samples are from tissues and not cell lines, the measurements that we get are averages from a population of largely heterogenous cells. This may mean that we can only consider a two states model, where one is fully methylated and the other is low to unmethylated.
- This mean that we should use any valley/"peak" finding method to find LMR: loess, poisson model, two-states hidden markov models, etc.
Binning with a sliding window approach
- Start with the first measured CpG position in the genome
- Try to merge this to the next position if they are <= 2kb apart.
- If either one of the CpGs have fewer than 10 reads, then expand the first one left and the second right ("borrows" information from sites nearby that are <=250bp away) until the minimum read depth is achieved.
- Compare the methylation levels of the two using either probability (soft thresholding) or minimum difference (hard thresholding) if the threshold for merging is passed, then merge the two CpG positions.
- After the merging decisions have been made, create merged windows & label.
- Merge unlabeled CpGs to the left and right of each windows if they are within 250bp & are > 2kb from the next nearest window
- Merge two similarly labeled windows together if they are within 2kb of each other.
Probability approach #1:
- Compare a window with it's non-overlapping upper and lower neighbors using Chi-square statistics & requiring a minimum 0.2 difference.
Brute force #1:
- We define a threshold for where the borders are going to be set. If there is not point above or below the threshold, we merge the windows together.
Annotation
- Label a window FMR if <=0.2 distance from 1 and LMR if >0.2 from 1.