5/Mar/26 - Lab meeting presentation feedback

[ppreshant]

initial discussion about m3 conference and giving a rigorous talk that stands scrutiny by big guns in the field

Demo/Presentation feedback

mmlongs2 features

• tiara: what's the point in detecting / removing eukaryotic contigs before binning?
  • Maybe they needed it to make their GTDB-tk classifier to work smoothly?
• Why 4x iterations on the binning? : what gives this arbitrary magic number ; was this data dependent when chosen and would it be optimal for any dataset?
  • #somatem #features : We should be able to implement a proper convergence criteria. Austin came up with a good metric to use for convergence (As a goodness of binning): Use pigeon's score

  Felix: do their parameters change across these 4 iterations : Maybe that should be useful/easy to implement with a fixed number of iterations (PK)

Documentation

Could be AI assisted but make sure to clean it up and that no "AI detectors" catch onto it

• Maybe Grammarly AI would be nice + AI reuser/ paraphraser to clean it up

• Features of AI writing: (apart from Em-dash) excessive bolding / of random words, (corporate speak) lot of filler text that doesn't make any point / is unsupported

• Need to add images / screenshots @microbemarsh Not sure what would be best to include in our case?, maybe visual outputs?

Tool notes :

Taxonomic profilers

• lemur: Mike will have a new DB out before 20/Mar that we can test and incorporate
• sylph/savont : will be key for fast running #features #somatem
  • Is there a point in implementing savont? @microbemarsh (look at/paste old discussion for choice between lemur and sylph..)
• What was the issue / discussion about taxburst plots again? was the plot that came up bad or something because of the 6 month old/unmaintained repo? Any changes we need to make for this or was it just a passing comment? @microbemarsh

Design choices

• Optimize for user experience ; don't do docker instead of conda for this reason..
  • Think of people giving up after 2 debugging missions within half a day
• Docker vs conda: conda for regular self-deployment and docker for cloud managed ones? supporting both will be challenging though.. maybe seqera's wave helps here?
  • docker makes it harder for people to retrieve intermediate files that are not explicitly declared as outputs in the nextflow process
    • Experience from STAMPS course : not sure what they were running within docker here ; likely not Nextflow?
  • I don't even know where to find any intermediate files right now ; I guess it depends on the executor (local, cloud) rather than the environment (conda vs docker)
  • (debugging internal files) Find out if nextflow makes such intermediate files accessible ; generate a report to Todd. example: meta-Flye's 50 odd files?
• Vet modules for good labs + well maintained ones before incorporating into Somatem
  • Discard unmaintained packages eventually
  • Trust the open-source/bioinformatics community to maintain their packages within a couple days/reasonable amount of time

Take-aways

• Get onto conda ASAP ; maybe polish versions and have one ready by the 20th Mar for the talk
• Minimize making more features : maybe just 1 (from the take last slide / discuss with Todd?)

[ppreshant]

tax-profiling: Bakeoff thoughts

Could run multiple profilers with Eddy's unified database(s) and optimal parameters and combine their results into additional analysis

Tools

to run by Todd?

• sylph
• lemur
• Kraken2
• Centrifuger
• Gannon2

(from Bakeoff paper) We evaluated six taxonomic methods spanning k-mer, alignment-based, probabilistic, and sketch-based strategies: Kraken2 (v2.1.3), Centrifuge (v1.0.4), Centrifuger (v1.0.9), Ganon2 (v2.1.1), Sourmash (v4.9.0), and Sylph (v0.8.1). Tool selection was guided by three criteria: (i) widespread use in metagenomic taxonomic profiling, (ii) demonstrated compatibility with long-read data, and (iii) support for customized reference databases based on NCBI RefSeq or related resources.

Kraken2 and Centrifuge are k-mer–based methods that assign reads via exact or approximate k-mer matching, whereas Centrifuger extends this framework with a probabilistic reassignment model that redistributes ambiguous k-mers to stabilize abundance estimates. Ganon2 employs hierarchical interleaved Bloom filters and assembly-aware indexing to improve scalability and control false positives. Sourmash and Sylph are sketch-based profilers that rely on MinHash and FracMinHash containment sketches, respectively, to compare query read content against large reference catalogs. Unless otherwise noted, all tools were executed with their default parameters using 40 threads, and outputs were converted into a unified result schema for downstream evaluation. A concise summary of tools and their default databases is provided in Table~\ref{tab:methods_overview}, with full command-line details in the Supplementary Methods.

How to ensemble?

How exactly will this emsemble be analyzed?

1. merge all their OTU tables using taxpasta (as initial collation for user inspection and custom analysis)
2. Pick a criterion of taxa presence in 2+/5 or 3+/5 tools for stringent presence ; compare this to magnet outputs?
3. Big question: merging quantitative profiles - maybe for later?