gregpoore
commented
1 week ago Hi @hermidalc, apologies for the delayed reply. I'm happy to help provide you with files but first have some comments:
- Newer, microbiome-specific batch-correction tools have been introduced since the 2020 paper, including ConQuR, which we recommended using in the Oncogene 2024 work for addressing the TCGA batch effects. Technical details aside, a key advantage of ConQuR is that it provides corrected outputs in counts, permitting usage of downstream microbiome analyses (rather than the log-adjusted values from Voom-SNM). However, its main limitation is the usage of only one batch variable at a time, so we had to subset TCGA data to Illumina HiSeq-only and either WGS or RNA-Seq to use it. This means there are separate ConQuR-corrected tables for WGS and RNA-Seq data, and the total number of samples is less due to the Illumina HiSeq subsetting.
- The additional rounds of stringent host depletion (GRCh38, T2T, HPRC, GENCODE) most affected RNA-Seq samples from UNC, which were sequenced less deeply than other samples in TCGA. This was complicated by how many RNA-Seq samples were generated with 50-bp sequencing, which can make host filtration overly sensitive (ie, removal of putative microbial reads with any homology to human regions). Since we aimed to be very conservative in the manuscript (ie, preferring high specificity of microbial read calls despite lower sensitivity), these processes cause the UNC samples to have poor quality, and we did not perform further analysis of them. Related to ConQuR, since the only other RNA-Seq center beyond UNC was Canada's Michael Smith Genome Sciences Centre (CMS), the CMS RNA-Seq data was not batch corrected.
- As another conservative measure, we focused only on human-associated species, and further restricted to those with high aggregate coverages (≥50%). These considerations reduce the read counts but lend greater confidence in their biological plausibility.
- As another conservative measure, we implemented direct alignments of host-depleted data to cleaned microbial genomes in RefSeq release 210 (ie, "RS210-clean"). These alignments should be more specific than Kraken's k-mer-based approach, although they may be less sensitive.
To summarize, the tables we currently recommend would be (i) raw or ConQuR-corrected, (ii) separate for WGS and RNA-Seq, (iii) comprise human-associated taxa with high coverage, and (iv) derive from direct alignments against RS210-clean. Table S13 of the paper provides the raw (WGS and RNA-Seq) abundances of RS210-clean taxa with ≥50% aggregate coverage (note: Table S6 contains the taxonomic names for the genome IDs in Table S13). However, I can find and share the ConQuR-corrected WGS table here if that would be helpful. Is that what you would prefer?
Edit: There are creative ways to get ConQuR to correct for >1 batch variable at a time, such as concatenating multiple factors into a single vector in R, but we did not publish this data (in part because of limited time to reassure ourselves the correction acted appropriately). In theory, doing this would help create the kind of single abundance table you're asking about. If this is something you want, I can point you in the right direction.
hermidalc
Dear Greg - thanks for your continued work on this and for this follow up paper with updated pipeline. I'm the lead author of Hermida et al. Nat Commun 2022 which used your original Poore et al. Nature 2020 dataset, specifically the "Kraken-TCGA-Voom-SNM-Plate-Center-Filtering-Data.csv” and “Metadata-TCGA-Kraken-17625-Samples.csv” from ftp://ftp.microbio.me/pub/cancer_microbiome_analysis.
Do you have the microbial abundances from this updated Oncogene 2024 pipeline for all 32 TCGA cancers in a format similar to the two files mentioned above?