I am working on Microbial Diversity on human skin using the V3-V4 300bp amplicon by Illumina.
All in all, I performed five independent sequencing runs. Especially on the first run, the reverse reads are of very poor quality. Even if I push out the truncations limits to a very low level I only get 1% of merged reads when using dada2-denoise-paired. Hence I decided to use dada2-denoise-single-reads.
1.) What kind/ amount of information am I losing if I use denoise-single instead of paired?
2.) Can I use denoise-single on the first run and denoise-paired on the later runs (where reverse reads look better) and later merge the table? Or will this result in a bias in downstream analyses?
You will lose some amount of taxonomic resolution and ASV resolution (impacting diversity, etc), but probably not an enormous amount.
If you want to run a test to see what kind of information you lose, you could follow this tutorial but trimming with two different primer sets (V3 only vs. V3-V4).
This will certainly lead to biases, as the paired-end reads will be longer and somewhat taxonomically different from the single-end reads. If using ASV (denoising) methods, your ASVs will all have different IDs and so no ASVs will be shared between the single and paired-end runs!
On the other hand, if you do want to follow an approach like this you could use closed-reference OTU clustering to map all of the single- and paired-end reads against full-length reference sequences, then it will be possible to compare across runs. But then you lose other information (e.g., ASV-level variation, reads that do not map to a reference) so it is still a trade-off.
All in all, I'd suggest seeing what the single-end reads get you in terms of diversity and taxonomy results, then circle back if you think that you are losing too much.