Background

  • Hint PMC7087379 and HiCnv PMC8210825 calculate CNV profiles from Hi-C
  • NeoLoopFinder adds segmentation module using HMM PMC8191102
  • Eigenvector decomposition (ICE) PMC3816492
  • The ICE is biased when applied to cancer data with CNV variation, which can be corrected using NeoLoopFiinder

Problem Description

How to Account for CNV in 3D Contact Normalization using HiChIP learning from HiC approaches. </br>When:

  • C[i,j] : unknown contact map between i and j
  • D[i,j] : observed contact counts between i and j
  • B[i] : Bias term
  • S[k] : the average of nonzero marginal sums of bins with copy number equaling to k.

The corrected Contact map C can be calculated by:

equation description
C[i,j] = B[i] x D[i,j] x B[j] estimate C
B[i] = B[i] x S[k] / sum(D[i,]) update B[i] := k-copy proportion over all contact flux from i

NeoLoopFinder incorperates a S[k] term for the bins w/ CNV value is k. In brief, this additional term, relative to the marginal contacts sum(D[i,]). panelizes the C[i,j] in proportion to the CNV amount. I interpret this as a CNV-aware track, where biases are factorized into contributions corresponding to multiple CNV levels.

Results

  • Focus on Type II CNV Loss
    • segments expanded with 20k binning
    • Condition: CNV input > 0, and for methods in {Blue, Red, Green}, the difference between CNV method and CNV input > 0.5.
  • Color Codes:
    • Gold: Input ChIP
    • Blue: CopywriteR HiChIP
    • Red: Our Multi-peak HiChIP
    • Green: NeoloopFinder (3D) HiChIP
Title figure
35 CNV loss regions image
IGV snapshot at chr8:85,605,879-136,041,554 image
  • Mpeak recovering the loss
  • The 3D-aware method tends to overestimate CNVs in regions with a high degree of 3D contacts.

Conclusion

Applying a 1D method to HiChIP with proper peak calling reduces Type II errors, whereas incorporating 3D-aware information tends to introduce Type I errors.