You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, trio requires users to build homopolymer-compressed databases and we run one round of consensus after rukki resolution and coloring. In contrast, with Hi-C data, we run consensus, map Hi-C data, then run rukki and another round of consensus. This is similar to the architecture for GFAse as well.
This should be universal, perhaps always use uncompressed post-consensus sequence? This would allow users to build normal meryl databases rather than HPC ones, simplifying confusion. Additionally, output file locations differ between HiC and trio (e.g. trio = 6-layout*/*scfmap vs hic = 8-*/6-layout*/*scfmap). These should be promoted to top level to be the same in both cases.
The text was updated successfully, but these errors were encountered:
Currently, trio requires users to build homopolymer-compressed databases and we run one round of consensus after rukki resolution and coloring. In contrast, with Hi-C data, we run consensus, map Hi-C data, then run rukki and another round of consensus. This is similar to the architecture for GFAse as well.
This should be universal, perhaps always use uncompressed post-consensus sequence? This would allow users to build normal meryl databases rather than HPC ones, simplifying confusion. Additionally, output file locations differ between HiC and trio (e.g. trio =
6-layout*/*scfmap
vs hic =8-*/6-layout*/*scfmap
). These should be promoted to top level to be the same in both cases.The text was updated successfully, but these errors were encountered: