My samples have been 16S sequenced using MiSeq (v4), and PacBio (full-length) and I need to compare their output statistically. I calculated selected alpha diversity metrics (observed, Shannon, Pielou, Simpson) for sequences generated from each technology. Now, I have each calculated metric for both technologies for the same set of samples. I am confused about selecting the most proper statistical test to compare both technologies using each metric. I found the attached paper where they used the Kruskal-Wallis test, I don't know why since the comparison between 2 sets of data (v3v4 vs FL)? However, using Prism, I performed the Mann-Whitney test (Wilcoxon rank-sum test), unpaired, two-tailed, and I got significant differences for observed and Pielou metrics, and non-significant results for other metrics. I am glad to hear recommendations from the community.
No idea how you analysed the data (OTU-based, ASVs...), but from starters, the sequencing methods have different amplification biases as they use different primers in the first step of the whole process, even if you leave aside that they are 2 completely different technologies.
Not surprised at all that some tests come back as significant.
PS: Evar is preferred over Pielou's evenness as Pielou's puts more weight on the more abundant taxa. Chao1 preferred over "observed" as the latter is dependent on the sequencing depth. Shannon index is also sensitive to sequencing depth, and inverse Simpson is quite robust to different sequencing efforts.
Xabier Vázquez-Campos
Thank you for your response and suggestions. So, regarding your question about how I calculated the metrics, simply in R using estimate_richness function in Phyloseq. This was calculated for each phyloseq object (one from Illumina MiSeq and the other from PacBio, for the same samples). I used DADA2 pipeline in R to generate these objects. So, for the same samples, I have Observed -OTU, Shannon,...from each technology. Then, I compared the scores of the same metric from MiSeq vs PacBio using Wilcoxon rank-sum test. Since I got a big difference between both sequencing technologies in terms of no. of taxa and generated read counts, so, this makes sense why observed, and evenness indices showed significant differences between both technologies. Also, some samples were luckier in MiSeq rather than in Pacbio and vice-versa. I already calculated Chao1 but I got exactly the same values as observed OTU since it is calculated at the sample level, any suggestions here?
I will test Evar and inverse-Simpson and keep the forum informed for discussion.
Much thanks!
Xabier Vázquez-Campos
For updates, I calculated Evar, Fisher, Inverse-Simpson, Rare-abundance, Dominance-abundance matrices and performed Wilcoxon rank-sum test between MiSeq and PacBio scores. So, Evar, Fisher, Rare-abundance, Dominance-abundance were significant which agrees with the differences in the microbiome composition from both technologies. MiSeq amplified large no. of rare taxa whereas PacBio missed them or probably it is more accurate than MiSeq given the length of the amplified region.
However, inverse-Simpson was non-significant as Simpson and Shannon.
Reckon you have to deal with the fact that the two datasets are not independent of each other - you've used two technologies on the same set of samples. Might be better using something like a paired samples t test (which tests that the average difference between the two sets of samples is zero)
Paul Somerfield
Thank you for your comment. I quickly re-tested my dataset using paired instead of unpaired, but still non-parametric because my data do not follow normal distribution. Also, the paired-sample t-test (parametric) tests that all difference between the 2 sets of samples is zero, and this not the case here since I am using 2 diff technologies. The results were the same either using paired or unpaired non-parametric test (Wilcoxon match-pairs signed rank test). Any recommendations here?
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Modeling
- Analyses of Variance
- ANOVA