Dilute mock data with reference.

[ndreey]

Does not have to be perfect, a simple script will suffice. Number of reference reads / total number of reads = host contamination

I believe benchmarking of 0, 50, 80 and 95% host contamination will be statistically significant (maybe skip 80% depending on time, but would be nice for correlation).

[ndreey]

Diluting the already mock data seems more efficient than having NGSNGS generate 10 samples using P_zinji ref + mock data reference genomes.

[ndreey]

Mentioned this in a different issue. "I were to dilute a 4.5GB mock data file to reach 90% P. zijinensis the file would become ~45GB??."

This is not good..

From the paper:

• The 21-sample 315GB plant-associated dataset includes 894 genomes. Of these.
  • 224 are from the proGenomes terrestrial representative genomes.
  • 216 are newly sequenced genomes from an A. thaliana root rhizosphere.
  • 55 are fungal genomes associated with the rhizosphere.
  • 398 are plasmids or circular elements.
  • 1 A. thaliana genome.
• Of these genomes, 15.3% (137) have at least one closely related genome present.
• Note that 90% of metagenome sequence data originate from bacterial genomes, 9% are fungal genome sequences and 1% is from A. thaliana.

This leaves me with these options..

• A. use CAMISIM to generate mock data the same way CAMI did to generate the root rizhosphere data but instead change the abundance profiles so that 90% become P. zijinensis. Could also change so we get more fungal endophytic abundance than bacterial.

  • Cons: Might be resource heavy, need to learn new program CAMISIM

• B. The A. thaliana reads are already removed from the mock data. Thus 90.5% of the data is bacterial and 9.5% is fungal. The idea now is to reduce this data by taking a subsample of this filtered mock data (1gb) and then dilute the subsample so i reach 90% P. zijinensis reads.

  • Cons: Smaller subsample = sampling bias and increase miss rate on rare taxa.

[ndreey]

The most logical will be to run CAMISIM as it was designed to create benchmark datasets. I will therefore explore option A.