How to use DESeq2's variance stabilizing transformation with microbiome data (CC195)

[@kylegervers1731]

Another great video, and I'm so excited that you're going to be tackling alpha diversity next! I'm assuming that the same benchmarking approach you've been using can also be applied, but the difference in richness, Shannon index, etc. can be plotted on the y-axis against the same difference in sequencing depth on the x-axis that we've been plotting. Along these lines, I'd be interested in testing how the iNEXT sample-size or coverage-based alpha diversity estimates (Hsieh et al., 2016, Methods in Ecol. and Evo.) compare to same rarefaction approach we've been using, which requires some samples to be dropped due to insufficient depth to estimate a parameter.

[@Riffomonas]

Thanks! Same approach. I’ll be sure to check out those other methods

[@Mmitish]

Please correct me if I am wrong -> (The VST in DESeq2 is commonly applied to raw counts (after filtering and handling of zero counts) to address overdispersion and stabilize the variance across the mean. This transformation is typically performed before downstream analyses that assume homogeneous variances, including distance matrices/ Euclidean or Bray or ...)

[@Riffomonas]

That's my understanding. As used by the WNWN paper and elsewhere, a pseudocount of 1 was added to all values. You can see I did this here: github.com/riffomonas/distances/blob/c748a4086e9d1811810385ee2b11e270bbf1825a/code/variance_stabilization.R#L47

[@Mmitish]

Thanks!@@Riffomonas

[@krishmen]

Thank you for the wonderful series! These videos are very useful for digging into microbial ecology. What do you think about the assumption that uneven read count might reflect the true abundances of organisms/features? Thus, if the reads are rarefied to a common number, some important data about diversity can be lost. Or am I confusing something?

[@Riffomonas]

Hey Kriss - in general the read count doesn’t mean much because we do some type of normalization when we pool the PCR products. I have seen it matter when comparing things like lung and mouth when a ton of non specific amplicons are tossed bringing down their read count

[@АлександрБолбат-ы1у]

Thank you. Are you going to publish these results? I would love to reference them, but referencing a KZbin video would likely be frowned upon. :)

[@Riffomonas]

I’m working on it! Hopefully I’ll have a preprint up in the next month or so

[@leocadio_blanco]

@@Riffomonas Hi Pat - I am again dealing with the same question from reviewers - Aitchison + Centerlog... Over the summer I was not sure if I read somewhere you had summitted the preprint? By now I just sending people the link to your youtubes....

[@Riffomonas]

Very sorry for the delay! I should have the preprint up by the end of the calendar year and hopefully much sooner

[@1973vgc]

it is not boring! I am learning a lot from you! T H A N K Y O U !

[@Riffomonas]

Ha! Thanks for watching 🤓

[@BannerOfBlasfemy]

thanks a lot Pat. Would this stabilisation approach be appropriate prior to beta diversity analysis by NMDS of Bray-Curtis distances?

[@Riffomonas]

Hi! No, I'd never encourage anyone use DESeq2. Use rarefaction of Bray-Curtis distances and then plug that in to NMDS

[@BannerOfBlasfemy]

@@Riffomonas Thanks Pat. I guess you refer to Episode CC200. From this episode's concluding remarks I took that DESeq2 is the better one at stablising variance in a microbiome dataset....My problem is a virome dataset with relatively even seq-depth, but huge variance in absolute (remapped-to-contig) read counts (100-7M virus reads per sample), obviously with many 0s in every sample. I have spent some time with the github discussions around DESeq and phyloseq, and although I learned some things, for my objective, I have left more confused.

[@JimtheEvo]

One thing I found recently was trying to use deseq2 on a new Apple silicon Mac requires you use the intel version of r. Also I’ve always said no when asked if I want to compile from source, I think this works better if you don’t have a c++ library installed on your Mac. I might just have had 1 bad experience and stuck with not compiling.

[@Riffomonas]

Good to know - thanks for sharing!

[@huikl6562]

Wonder if it is feasible to generate a mean rarefied count table so we can treat the data just the same for easy implementation to common tools and functions. There got to be a "centroid" which its vegdist() results is the same as doing avgdist() on original data, or are there are some complication that compel us to do the iteration right before calculating these metrics?

[@Riffomonas]

That would be the relative abundance

[@gimanibe]

Thanks Pat, as always. Have you tried the normalization implemented in metagenomeSeq? I want to compare with your NULL model ;)

[@Riffomonas]

No but I’m not sure I’d expect it to be any different. Let me know what you find!