Inquiry on GARD : Nucleotides vs aminoacid breakpoints differences.

[Sasa-won-keuns-lab]

Hello,

I am running analyses on the attachment glycoprotein gene from different Jeilongvirus species, which show extreme variation in genome length. In this case, I used GARD prior to constructing independent phylogenetic trees, with assumption of different evolutionary models across different regions of the gene.

Initially, I performed translate–align–backtranslate and used the resulting nucleotide alignment for the GARD analysis. However, this consistently resulted in an error on the Datamonkey website (i think because it's too big). When I ran GARD locally (HYPHY 2.5.94(MP)), the result was:
CHARSET span_1 = 1-460;
CHARSET span_2 = 461-8232;

However, because I thought something might be wrong with my nucleotide alignment, I also ran the amino acid alignment on the web and obtained the following result:
CHARSET span_1 = 1-781;
CHARSET span_2 = 782-2162;

The breakpoints were very different. Biologically, I would prefer the nucleotide breakpoints because they occur near the transmembrane domain and make more sense for my current assumption. However, I am wondering whether I should use the amino acid result instead.
I use general discreet and rate classes 4 for both run.

Are there any recommended criteria or guidelines for deciding which result to follow?

[spond]

Dear @Sasa-won-keuns-lab,

This is an interesting discrepancy. To better understand what might be happening with your GARD runs, could you please consider the following:

1. Are the sequences in frame? I would expect the length of the amino-acid alignment (2162 sites) to be exactly 1/3 of the codon alignment (8232 nucleotides / 3 = 2744), but it appears shorter (2162 vs 2744). This suggests that some regions or sites might have been removed during the protein alignment or processing. Could you clarify if the alignments represent the exact same genomic regions?
2. GARD uses Neighbor-Joining (NJ) trees to search for breakpoints. If your alignment is sufficiently divergent (as may be the case with Jeilongvirus glycoproteins), NJ may produce different trees for the nucleotide and amino acid data, which would lead to different breakpoint inferences.
3. Did you check the AIC-c improvements for both the nucleotide and amino-acid options? Comparing the magnitude of the AIC-c improvement over the single-tree model for each data type can help identify which one provides a stronger signal for recombination.
4. What were the dimensions of your alignment (specifically the number of sequences)? The minimum length of a segment between breakpoints is a function of the number of sequences in the alignment; knowing this could help determine if one of the segments is nearing the resolution limit.
5. If you can share the Datamonkey URL or the alignment files, I can take a closer look at the results and provide more specific details.

Best,
Sergei

[Sasa-won-keuns-lab]

Dear @spond ,

For the GARD run using the amino-acid sequence, I used Clustal for the alignment (https://www.datamonkey.org/gard/69982bc53c97ca41d7b5dbbc).
For the nucleotide analysis, the initial back-translation from the amino-acid Clustal alignment failed (i have several attempts, this is one of them https://www.datamonkey.org/gard/699e9af33c97ca41d7c486ed). Therefore, I re-aligned the amino-acid sequences using MUSCLE, performed the back-translation, and attempted to run GARD on the Datamonkey web server (which failed), and then ran it locally (which succeeded).

However, I also ran GARD locally using the nucleotide alignment back-translated from the amino-acid Clustal alignment, and the result was:
CHARSET span_1 = 1-395;
CHARSET span_2 = 396-6486;

The JSON output also mentioned that “some or all of the breakpoints may reflect rate variation instead of topological incongruence.” This result seemed unusual to me, and I initially suspected that there might be an issue with my local HyPhy installation.

Because of this, I proceeded with the latter approach (translated → MUSCLE → back-translate). However, I did not run GARD on the MUSCLE amino-acid alignment itself. And I'm wondering which Δ c-AIC I need to consider more, since the vs null and vs single tree multiple partition was higher in different way. At the moment, I prefer the nucleotide result because the partitioned trees appear more reasonable. Nevertheless, this virus is highly diverse, so my current assumptions about its evolution could be incorrect.

In terms on AIC:
Amino-acid :
Δ c-AIC vs the null model = 1993.75
Δ c-AIC vs the single tree multiple partition : 37.7718

Nucleotide (from amino-acid clustal alignment) :
Δ c-AIC vs the null model = 1069.80
Δ c-AIC vs the single tree multiple partition : -187.897

Nucleotide (from amino-acid muscle alignment) :
Δ c-AIC vs the null model = 1167.18
Δ c-AIC vs the single tree multiple partition : 850.424

For the amino acid, after remove the duplicate, the total sequence was 65. The nucleotide sequences was 67.

I attach the nexus files here :

nucleotide_clustal.nex.txt
nucleotide_from_aa_MUSCLE.nex.txt
protein_clustal.nex.txt

Thank you very much for the input. Really appreciate it.

Best,
Augustine

[Sasa-won-keuns-lab]

Dear @spond ,

I tried to run the aminoacid aligned by MUSCLE and it gave closer breakpoint to the its backtranslated alignment GARD result.

Therefore, I think the problem is solved for me, and I will rely on MUSCLE more than CLUSTAL based on this experience.

Thank you for the advice!