Input Data in Training

[ekokrek]

Hello Wang,

I tried to predict chr12 with pretrained models (both HindIII_40000 and model_12000).
When I compared the enhanced matrix with the original and down-sampled (1/16) matrices, it was similar to the down-sampled. So the enhancement was quite unsuccessful.

Down-sampled input is generated as follows: GSE63525_GM12878_primary_intrachromosomal_contact_matrices.tar.gz file (HIC001-HIC018) is downloaded and raw contacts at 10kb resolution are down-sampled by a ratio of 1/16. In terms of sequencing depth, reads from HIC001-HIC018 make up ~4B reads and down-sampled version makes up ~220M reads. So I guess your model(s) should be appropriate for this situation.

Technically, what I down-sample is the pairs not reads but would that affect the result that much?

Can I learn which exact GM12878 data was used in training these pretrained models?