We have developed FMTs using microbiota from six human samples, with each human microbiota sample transplanted into 2-3 individual mice. As these mice will be used in further experiments,we want to evaluate whether the mice successfully reconstituted the microbiome from the human FMTs. For this, we performed 16S sequencing on both human and mouse samples. Now I’m considering the best approach for analyzing the data.
My initial plan is to perform a beta diversity analysis to see how each human sample clusters with its respective mouse samples. This should help identify any mice with potentially incomplete transplantation in a straightforward way.
If I want to go further and compare the taxa abundances between human sample and mouse samples, I’m unsure of the best approach. My thought is to compare each human sample individually with each of its corresponding mice, for example, as follows for human sample 1:
human_sample_1 vs mouse_1_human_sample_1
human_sample_1 vs mouse_2_human_sample_1
human_sample_1 vs mouse_3_human_sample_1
I am also considering calculating intersections, such as the percentage of genera shared between each human and mouse sample. However, since both sample types might share taxa at different abundances, I’m wondering if a correlation analysis of relative abundances could be better to assess the transplantation quality.
Does anyone have suggestions for additional analyses, or has anyone faced similar analyses that could provide insights or alternative approaches?
I would definately start with beta diversity! Remember that different metrics will give you different ideas about what's shared. So, jaccard will tell you about presence/absence, bray curtis will tell you about the most abundant taxa, and aitchison should, in theory, tell you about things that double/half, but if you dont filter carefully, it can be affected by zeros. (Don't rarify before Aitchison, either! It will increase your sparsity and make things worse!) I would probably check the distance between the donor and recepient vs the distance between a random donor and the recepient (e.g. does the mouse look more like its donor than another human). I'd also recommend checking with in the mice: does a mouse from human 1 look more like another mouse from human 1 than another mouse from human 2. You'll probably see clustering in PCoA space, so it should be pretty obvious.
If you're looking to trace features, I'd suggest looking at the ASV vs genus level. I'd probably look at a correlation with a CLR-transform vs direct relative abundance, both because the data behaves better as CLR and becuase the CLR should be more insulated from compositional effects. You might even first look at taxa that colonize the mice (which subset of the human microbiome is in the mice) and then look at correlation there.
Thank you so much for your response! I will follow your advice for my analysis.
Regarding the correlation analysis, I have been exploring whether certain methods are preferred for compositional data and came across this pre-print article (https://www.biorxiv.org/content/10.1101/2024.02.29.582875v1). It suggests that with proper normalization, such as CLR, the performance of straightforward measures like Pearson’s correlation is similar to state-of-the-art methodologies like SparCC or Rho. What are your thoughts on this aspect? Do you have any preference based on your experience?
I dont think your preprint is appropriate to your scenario. You're not constructing networks, you're asking about the relationship between specific taxa within a group. (Does donor predict acceptor). I also diagree with the authors about their claimed advantages of the L1 transform vs CLR (namingly, I'd argue the move into euclidean hyperspace over a simplex is a feature not a bug). I also think the log transform better reflects the way we measure bacterial growth, and that stablization is an improvement. I agree that sparsity is a major concern, but I think you can solve that by first identifying what features transfer and not starting with correlation.
Hi @andresarroyo,
I just wanted to pop in and echo what @jwdebelius is saying and also suggest a tool that I am developing called q2-fmt which is qiime2 plugin that specializes in exactly what you are talking about! I think this should help you vizualize the engraftment of the microbiome!
I am currently working on a tutorial for using q2-fmt but that is still in the works. If you are interested I can circle back when it is ready.
@cherman2 the plugin seems really useful. I will check both your pre-print and the plugin. Sure I will be interested in the q2-fmt tutorial when it was ready.
@jwdebelius your comments help me to clarify this aspect. I had not taken into account that sparsity would be solved if I previously select transferred features.
I have an additional question about taxa pre-filtering in this context. Human samples were sequenced in a previous run with other human samples, while mouse samples were sequenced in a second run. At what stage should I merge the human and mouse samples? maybe after denoising and applying filtering steps to all samples together, or should I filter each mouse and human samples separately using the same parameters and then merge both feature tables and taxonomies? I’m not sure how relevant the difference in both approach would be.
I have a question that relates in part to the last question I asked in this thread. As I mentioned before, the human and mouse samples were sequenced in two separate sequencing runs. Additionally, my workflow includes using q2-sidle, as I’m working with multiple V regions.
Previously, I generated the sidle reconstruction outputs for my human samples (~70 samples), and now, 6 out of these 70 are the ones I'm using to evaluate the FMT. My question is about the comparability of the sidle reconstructions since they were performed independently for the two runs (human and mouse). Do you think it would be better to merge the DADA2 outputs and run the sidle reconstruction with all samples (human + mouse) at once? Would you expect any differences between these two approaches?
The key for Sidle should be the primers and database. I would use the same prepared database, and use it consistently. The one step Im not sure you'll be able to combine is tree construction, so maybe be aweare there?
I consistently use the same primers and the SILVA 138 database. So far, I haven't included tree reconstruction in my workflow because it was originally based on SILVA 128. However, I need to evaluate whether including it would be suitable, as I’ve seen discussions in the forum about using different SILVA versions at different stages. I plan to read up on this in more detail.
My main concern here may actually be simpler. In my two sequencing runs, the sequencing quality was different and I used different region trimming lengths for each run (for example, V4 in run 1 is 200 bp, while V4 in run 2 is 180 bp). The length differences across V regions are at most 20 bp. Could these length variations in the V regions have an important impact on each Sidle reconstruction process, or should this effect be minor and not very significant?
Sidle requires a fixed length. The longer reads can be trimmed with trim-dada2-posthoc, which will make the reads the same length. But, if the mice are the shorter reads, you will have to re-extact the database and re-process the human reads.
You can't mix database versions, and unfortunately, no one has published an updated fragment insertion backbone. So, if you wanted a tree, it would be with Silva 128.