I sent a test plate to the seq facility we are working with for ddRAD sequencing of a non-model organism without a reference genome. These 50 samples came from 3 field sites from 1 region. They ran these 50 in one lane of an Illumina HISeq. When the seq data returned, it appeared fine in preliminary popgen analyses after deNovo assembly and I recovered around 8,000 SNPs after filtering.
I then sent a large batch of samples (300) to be sequenced following the same protocol. These were from ~15 sites across 2 regions (5 in one region, 10 in another). The facility ran my samples as 300 in a single Illumina HiSeq lane, but ran the single lane twice, as opposed to 100 in each of 3 lanes, once, as I thought they would do. They said 300 in one lane ran twice would provide better quality than 100 in each of 3 lanes, ran one time.
When I received the 300 back, I combined those samples with my 50 from the test plate and proceeded to filter and call SNPs together, getting significantly less SNPs, but that was expected with the reduced depth of coverage from running so many samples/lane.
The main issue however, is the combined 350 dataset produces 2 genetic clusters when using program Structure and using a Principal Coordinates Analysis (PCoA) in Genalex. One cluster is the 300 and the other cluster is the 50. This is biologically impossible that either of these clusters are true as there are field sites in the 50 group that are very close to fieldsites in the 300 group. What could be causing the genetic "clusters" to align by sequencing run?
Kindly go through this paper. Hope you get the answers to variuos questions regarding the same.
"Insight into biases and sequencing errors for amplicon sequencing with the Illumina MiSeq platform"
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4381044/
Barring any bioinformatic causes (demultiplexing problems or something??), what you would be seeing are likely lane/library specific biases ("lane effects" or "batch effects")- variation in PCR amplification, specificity of size selection, depth of coverage, coverage of loci across individuals (which can be a consequence of the above), variation in recovered loci across preps, etc, which ultimately create biases representation of loci in one batch vs the other.
I am not sure what bioinformatic solutions to suggest. Perhaps increase the minimum depth of coverage per sample, and minimum coverage of individuals within a locus to retain it for analysis. You may also try only retaining loci which are present in BOTH the 50 and 300 sample multiplex runs and see if this corrects the problem.
How are you currently processing the data? What parameters for filtering?
Article Amplification Biases and Consistent Recovery of Loci in a Do...
Article Identifying and mitigating batch effects in whole genome seq...
If the two samples (50 and 300) are really biologically similar, then your are looking for artifacts from the sequencing itself.
So this would certainly include any variables from Tyler's suggestions of " lane/library specific biases ("lane effects" or "batch effects")- variation in PCR amplification, specificity of size selection, depth of coverage, coverage of loci across individuals (which can be a consequence of the above), variation in recovered loci across preps, etc"
Again, if you measured any of these variables in the two analyses you should be able to determine which ones are responsible for the different clusters by looking at the first and second principal component coefficients.
Suppose the largest absolute values of the coefficients of your first principal component are those for the PCR amplification and lane effects variables and the depth of coverage and batch effects are relative small. Then the interpretation of the 1st principal component should be that differences in PCR amplification and lane effects are responsible for the differences in the two clusters (if they are separated along the 1st principal component axis).
If the largest coefficients have opposite signs (+ or -) then you may interpret them as "the PCR amplification is negatively correlated with lane effects".
Similarly for the second largest PC which should have a completely different set of most influential variable, notice which variables have the largest absolute values of among all variables.
Now to specifically decide why they cluster, you need to determine which of the PCs is responsible for the gap between the clusters. It may be just the 1st component is which case you might have a plot that does not vary much on the PC2 axis. If the second PC is responsible for the difference then you may see little variation on the PC1 axis but lots on the PC2 axis. If the clusters are separated on both axes then the largest variables associated with the PCS are most responsible for the differences.
If you intend to publish the data, then you will need to discuss the reasons for the clustering.
- Badges
- Science method