4D Nucleome Network Wiki

Hi-C read filtering. Presentation by Anton, Max, Mirny Lab.

(Minutes by Burak; Feel free to edit.)

4 types of filters

• - PCR
• - Dangling ends / self circles
• - Poorly mapped reads (MAPQ)
• - Random ligations (distance to restriction site)

Data set:

• IMR90 HindIII Jin et al
• IMR90 MboI Rao et al

1. PCR: looking at pairs vs. max-bp differentce

• - large peak at 0
• - flattens out at max_delta=3
• - mismatch of 1 or 2 does not have a lot of excess but they might be worth removing.
• - Bing's question on tool: Using home-grown pairsamtools.
• - Burak's question on comparison to juicer: basically agree.

2. Dangling ends / self circles

• - evidence in read pair orientation discrepancy at short distances.
• - Dangling ends: ~+- pairs out to 1jb
• - self circles: ~-+ pairs out to 10kb in HindIII; not so evident in MboI since molecules don't circle back as much.
• - Conclusion: We cannot trust contact within ~10kb for 6 cutters or within ~1kb for 4 cutters.
• - Yunjiang: The error structure is to some extent library specific. Anton/Max: All HindIII data we've looked at has similar trends with roughly 10kb.
• - Max: restriction efficiency can also have an effect in determining features. So we should look at ?2-3 restriction fragment scale
• - Burak: Did you look at restriction fragment instead of bp distance. Max: Yes, in a paper before.\ * - Anton/Max: propose to keep the pairs because they can be useful. But we might want to flag them. But don't include them in corrections (normalization) * - Bing: need to make sure these are removed in matrix. 3. MAPQ * - cis/total ratio does have a MAPQ dependence out to MAPQ=60. * - Try to estimate sensititivy/specificity vs. MAPQ threshold. * - MAPQ>0: 17% of filtered reads comes from mismappers; MAPQ>=60: 3-5& of filtered reads comes from mismappers;\
• - MAPQ>=60: 20% of reads are removed
• - the effect of MAPQ>0 vs MAPQ>59 does not appear to be big in terms of cis contact probability distributions
• - Conclusion: MAPQ>0 or MAPQ>10 might be the optimal cutoff.
• - Followup
  • - - One should compare the matrices as a follow up.
  • - - Leonid asks: any A/B compartment features?
• - Another solution:
  • - - report pure and less pure outputs, similar to juicer.
• \Hoachen: Similar studies in WGS from same cell type may be informative 4. Distance to restriction site * - very close to rest-site (1-few bp): dangling ends * - very far from rest-site (: random ligation * - Also note low rate of within 30bp; because of bwa-mem requirement. very close:\

• - can't really study cis/trans.
• - cis contact probability distributions: some bias in very short distances, but overall have similar features to total.

very far:

• - similar approach to above; use cis-trans to estimate sensitivity/specificity of a distance threshold. Low
• - cis contact probability distributions: appears normal.
• - 4cutters: you can't even assign to the right restriction fragment. Very very few reads at large distances
• - 6cutters: 15% are far from cut site, but don't show bad properties; and are estimated to have relatively low noise.

Conclusion:

• Propose: remove duplicates and short distance pairs.

Action items:

• Build matrices for different filtering outputs and compare.

Next call:

• 2 weeks. We will follow up on read filtering.