Moore Notes 12 15 10

Group Call

• Call logistics
  • Will use all calls for research topics
  • Announce to the group if you want to lead a discussion on a particular project
  • PIs will grab the time slot if needed, but we'll stop doing strict every other week (any week can be research group call)

• Taxa area relationships for niche modeling - James
  • Code up and running
  • Not much pattern at coarser resolution (phylum)
  • Will contact Josh about finer taxonomic resolution data (genus)

• Tom has maxEnt up and running on linux machines
  • Can run the analyses in parallel on the cluster now
  • Opens up computational validations (e.g., jackknife analysis)
  • Will start writing if validation looks good
  • MaxEnt has another set of tuning parameters that are being investigated and can be used if needed

• Protein families - Guillaume (performance plots)
  • Tried mapping seed sequences back to the families they were used to build (1 month compute)
  • Recall and precision are perfect for most families, but there are long tails
    • Only used top hit, second best hit might be the right one and a decent hit
    • Sam: check how far off next best hit is
    • Katie: could they actually belong in the other family and MCL put them in the wrong cluster?
    • Tom: still a good validation for the well-differentiated families (when MCL clustering was clear)
  • Analysis of "bad" families (imperfect recall and precision) - Tom
    • Can we grow the good families by adding in some of the sequences from bad families using hmmsearch
    • Or by growing MCL clusters?
    • Redo de novo MCL clustering of bad sequences only
    • Dongying: hmmsearch (in HMMR3, not HMMR2) is local search only
      • unless you increase the coverage threshold, so MCL or BLAST is better
      • can draw a cutoff on E-value cutoff on inside versus outside cluster hits
    • Katie: Could you use the non-seed sequences from big families? Guillaume - no, all were used as seeds for most families
  • What is next?
    • Modify AMPHORA2 code to work on a this database
    • In silico validations?
    • Read classification for metagenomic libraries
      • Qin et al microbiome
      • GOS (versus the peptide predictions)
      • HOT/ALOHA depth series - relative differences, common families would probably work
      • Might need to deal with genome size based bias
        • there's a correction available, but might not be generalizable
        • 16S copy number metric could be used perhaps
      • Could build trees for some families (e.g., with pplacer) and then do a PD analysis rather than abudance