[Pangenomics] Protein clusters -> gene clusters?

[meren]

I propose to change all instances of 'protein clusters' in our pangenomic workflow with 'gene clusters'.

Summary of the discussions so far

I will keep this section up-to-date, the original issue starts at the "Why bother" section below.

Although there is a large diversity of opinions, it doesn't seem anyone has a major concern with "gene clusters" yet (as far as how they are generated is clearly described):

• Hervé Tettelin thinks 'gene clusters' is the way to go.
• Simon Roux suggests that we could stick with 'protein clusters' because most predicted genes become proteins anyway.
• Narendrakumar Chaudharri suggests that 'gene families' could be a better alternative.
• David Needham brings up 'orthogroup' as an alternative.
• Alon Shaiber suggests that 'clusters of homologous genes' would be most appropriate.
• Tom Delmont suggests 'sequence cluster'.
• Rika Anderson suggests 'gene cluster' and 'sequence cluster' are fine options.
• Ryan Bartelme finds either 'gene clusters' or 'protein clusters' OK.
• Julie Reveillaud anything is OK if they have proper descriptions and asks for a detailed blog post on the topic.
• Pat Schloss mentions on Twitter that they previously called these clusters 'operational protein families'.
• Titus Brown also on Twitter discusses the accuracy of the term 'pangenome' (although I personally think it is important to strive for an accurate terminology, whether the term pangenomics truly represents what we are doing by grouping genes from genomes based on homology, is beyond our scope here (my opinion on this is here)).

I feel that our major problem was the need for a term that communicates the fact that we are working with sequences that are coming from predicted open reading frames. Coding sequence, protein, gene, family, orthology, paralogy, all bring in assumptions that do no apply to what goes in our clusters in a pangenome.

Why bother?

In the pangenomic workflow we take translated DNA sequences from open reading frames that are often identified by a gene prediction software, and then organize them in such a way that sequences that show a certain degree of homology form distinct 'clusters'. We currently call these resulting units as 'protein clusters', and investigate their distribution across contributing genomes to infer relationships between them.

To me, a more accurate definition of the result of this process is 'clusters of translated DNA sequences from predicted open reading frames', which is not very helpful. But it is rather straightforward and safe to suggest 'DNA sequences from open reading frames' represent 'genes', regardless of whether they ever become 'proteins' or not. 'Gene' is a much more modest term than 'protein' when one considers the fact that most of the open reading frames will correspond to genes by definition, but the actual number of them that ever turn into proteins will be less than that. So I think 'gene clusters' is a more appropriate term to describe what we are actually generating.

Clearly gene clusters can be generated from DNA or AA sequences, and the term 'protein clusters' clarifies that it was the AA sequences that were used to generate them. But this can be clarified in any methods or results section:

We used translated DNA sequences of predicted open reading frames to identify 'gene clusters' in our pangenome.

Literature note

The first 'pangenome' paper from Tettelin et al. only uses 'paralog clusters' and 'homolog clusters' to describe gene clusters, and never protein clusters:

https://www.ncbi.nlm.nih.gov/pubmed/16172379

PGAP: pan-genomes analysis pipeline, from Zhao et al., also uses gene clusters, and never protein clusters:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3268234/

A Concern

The term 'gene cluster' has a very well-defined and established meaning elsewhere in life sciences:

https://en.wikipedia.org/wiki/Gene_cluster

Can there be a more specific and meaningful term for what we generate in pangenomes?

Or should we not care since the context will make it obvious the gene clusters in pangenomes have nothing to with gene clusters in eukaryotic genomes like Hox genes?

For instance the NCBI uses the term 'protein cluster' to describe "collection of related protein sequences consists of proteins derived from the annotations of whole genomes, organelles and plasmids" (the grouping is informed by the protein function, so it is different than what pangenomic workflows do).

Any input is most welcome.

Best,

[luuuuuuuke]

Hi Meren, These are great thoughts! I have a question for clarification--maybe it's helpful to others...

Is a gene cluster simply a group of homologous sequences (according to set parameters), which could belong to a single genome or be shared across multiple genomes? Followup question: if so, does the pangenomic analysis distinguish between one genome with 25 genes belonging to gene cluster X, and another genome with only 1 gene belonging to gene cluster X? Or will these look identical in the display?

-Luke

On Thu, Nov 16, 2017 at 1:58 PM, A. Murat Eren notifications@github.com wrote:

I propose to change all instances of 'protein clusters' in our pangenomic workflow with 'gene clusters'.

But briefly my thinking goes like this:

In the pangenomic workflow we take translated DNA sequences from open reading frames that are often identified by prodigal, and then organize them in such a way that sequences that show a certain degree of homology form distinct 'clusters'. We currently call these resulting units as 'protein clusters', and investigate their distribution across contributing genomes to infer relationships between them.

To me, a more accurate definition of the result of this process is 'clusters of translated DNA sequences from open reading frames', which is not very helpful. But it is rather straightforward and safe to suggest 'DNA sequences from open reading frames' represent 'genes', regardless of whether they ever become 'proteins' or not. 'Gene' is a much more modest term than 'protein' when one considers the fact that most of the open reading frames will correspond to genes by definition, but the actual number of them that ever turn into proteins will be less than that. So I think 'gene clusters' is a more appropriate term to describe what we are actually generating.

Clearly gene clusters can be generated from DNA or AA sequences, and the term 'protein clusters' clarifies that it was the AA sequences that were used to generate them. But this can be clarified in any methods or results section:

We used translated DNA sequences of gene calls to identify 'gene clusters' in our pangenome.

Literature note

The first 'pangenome' paper from Tettelin et al only uses paralog and homolog clusters to describe gene clusters, and never protein clusters:

https://www.ncbi.nlm.nih.gov/pubmed/16172379

PGAP: pan-genomes analysis pipeline, from Zhao et al., uses gene clusters, and never protein clusters:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3268234/ Concerns

The term gene cluster has a very well-defined and established meaning elsewhere in life sciences:

https://en.wikipedia.org/wiki/Gene_cluster

Can there be a more specific and meaningful term for what we generate in pangenomes?

(another note: for instance the NCBI use the term 'protein cluster' to describe "collection of related protein sequences consists of proteins derived from the annotations of whole genomes, organelles and plasmids https://www.ncbi.nlm.nih.gov/proteinclusters" (the grouping is informed by the protein function, so it is different than what pangenomic workflows do).

Any input is most welcome.

Best,

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/merenlab/anvio/issues/644, or mute the thread https://github.com/notifications/unsubscribe-auth/AY6jPNspcmKOG7zIIYtDTasas2d5jA5sks5s3KHbgaJpZM4QhIL7 .

-- Dr. Luke McKay

Postdoctoral Research Scientist Department of Land Resources and Environmental Sciences (815 LJH) Center for Biofilm Engineering (313 Barnard Hall) Montana State University

[tettelin]

Hi Meren et al.,

Here are some points that I hope you will find helpful:

• I think that using "gene clusters" is indeed the way to go and as you mentioned, the key is to clearly define what you mean by this and how the clusters were generated. This is especially important to clarify the fact that you are not making any assumption about orthology of function. These are based on sequence similarity with defined parameters.

• One potential technical issue is that the term "gene" normally includes sequences outside of the open reading frame (ORF). The DNA sequences of a gene that are translated are the coding sequences or CDSs. In prokaryotes it's just the one ORF. Most people often use gene when they mean CDS so it's fine, again, as long as you make sure to clarify your terminology up front.

• Luuuuuuuke's question about paralogs is relevant. A method we use that generates reciprocal best blast match (RBBM)-based clusters we call JOCs, Jaccard-filtered clusters of orthologs (https://www.ncbi.nlm.nih.gov/pubmed/18314579), first generates gene clusters within each genome (JACs), then allows for RBBMs to occur between any gene or member of a JAC across genomes. The JOCs method sometimes results in abnormally larger gene clusters and I was wondering if your method does the same. Additional pruning of edges within clusters can help. Other methods use synteny to separate paralogs into distinct clusters of "syntenic orthologs" across genomes.

In any case, it would be good to provide the user with paralog information and how you dealt with it.

Hervé.

[dmneedha]

Hi Meren,

How would the term "orthogroup" fit into this conversation? I like it: https://genomebiology.biomedcentral.com/articles/10.1186/s13059-015-0721-2

Best, David

[meren]

Dear Luke,

Thank you very much for the input.

Is a gene cluster simply a group of homologous sequences (according to set parameters), which could belong to a single genome or be shared across multiple genomes?

Yes. A group of homologous sequences based on sequence identity, and/or network decomposition parameters for MCL-like algorithms. These clusters could have been generated by tools such as MMseqs2, too.

does the pangenomic analysis distinguish between one genome with 25 genes belonging to gene cluster X, and another genome with only 1 gene belonging to gene cluster X? Or will these look identical in the display?

Yes, good point. We do track paralogs. At a given time we know how many genes from each genome is contributed to a given gene cluster. This allows us to identify and mark the single-copy gene clusters, and show how many genes in each gene cluster:

And show paralogs when the user wants to inspect a given gene cluster:

I hope these make sense.

[meren]

Hi Hervé,

Thanks for taking the time :)

I think that using "gene clusters" is indeed the way to go and as you mentioned, the key is to clearly define what you mean by this and how the clusters were generated. This is especially important to clarify the fact that you are not making any assumption about orthology of function. These are based on sequence similarity with defined parameters.

Yes, these are based on sequence similarity with defined parameters. It will also be useful to extend our efforts to use functional information, but currently we haven't done anything in that front.

One potential technical issue is that the term "gene" normally includes sequences outside of the open reading frame (ORF). The DNA sequences of a gene that are translated are the coding sequences or CDSs. In prokaryotes it's just the one ORF. Most people often use gene when they mean CDS so it's fine, again, as long as you make sure to clarify your terminology up front.

Yes. Clarifying the terminology the first time we mention gene clusters will be essential to avoid any misunderstandings.

Luuuuuuuke's question about paralogs is relevant. A method we use that generates reciprocal best blast match (RBBM)-based clusters we call JOCs, Jaccard-filtered clusters of orthologs (https://www.ncbi.nlm.nih.gov/pubmed/18314579), first generates gene clusters within each genome (JACs), then allows for RBBMs to occur between any gene or member of a JAC across genomes. The JOCs method sometimes results in abnormally larger gene clusters and I was wondering if your method does the same. Additional pruning of edges within clusters can help. Other methods use synteny to separate paralogs into distinct clusters of "syntenic orthologs" across genomes.

I see your point. In some cases we do get clusters with very large number of genes in them simply because they all have a conserved region in otherwise dissimilar sequences. There are many things that can be done to divide them into synthetic orthologs. Trimming the conserved region from the alignment and re-clustering the remaining sequences can be one of them to re-organize those paralogs better.

In any case, it would be good to provide the user with paralog information and how you dealt with it.

We do our best to not make this information hard to find, but what we are doing in the interface could certainly be improved. Your comment made me realize that we can add another layer to show 'extent of paralogy' in each cluster (i.e., the maximum number of genes coming from the same genome for every cluster). I entered an issue, and we will implement that.

I guess no one so far has any real problems with "gene clusters" as far as how they are generated is clearly described.

[simroux]

Hi Meren et al., If I get the question correctly, it's really about how to name these clusters (i.e. "gene" vs "protein" clusters), and in that case I would vote to keep "protein clusters" particularly if the input for the clustering are predicted ORFs translated into amino acid sequences. I agree that not all genes will become proteins eventually (although with a predictor like prodigal relying on ORFs, I would suspect most of the predicted genes are either forming proteins or not genes at all, but rarely transcribed-but-not-translated), however my feeling is that "protein clusters" more intuitively reflects what has been done (i.e. the predicted protein sequence space has been clustered). On the other hand, I am not sure that gene is "much" more modest than protein, i.e. it could be my virus bias, but I don't feel like the number of genes which would not be translated into proteins is high enough to be much concerned about calling all these predicted ORFs "protein" clusters. Adding the notion of orthology to these is another question but I would think this would have more to do with changing the second word "cluster" to "orthologs groups" or some other terminology rather than the first word gene vs protein. Anyway, just my two cents :-)

[bio-developer]

Hello, the term protein clusters is not appropriate as you mentioned. Also, gene clusters is the term overlapping the other scientific domains. So, like our pan genome pipeline which refers to these as GENE FAMILIES would be appropriate as per my belief.

[meren]

@dmneedha,

How would the term "orthogroup" fit into this conversation?

I think orthology explains only a subset of the clusters we are generating, and ignores paralogy, hence it is not maybe the best option. To quote people who have thought about these a lot, Gabaldón & Koonin has a nice paper that starts like this: "Orthologues and paralogues are types of homologous genes that are related by speciation or duplication, respectively". So orthogroup does not explain what our gene clusters contain. Homology does, for instance, but "homologous groups" would have been even less helpful since the term "group" or "cluster" already implies homology.

@simroux,

I would vote to keep "protein clusters" particularly if the input for the clustering are predicted ORFs translated into amino acid sequences.

You make valid points, and I especially tend to agree when you suggest most of these predicted genes are translated to proteins, so why bother. But I think there is a large consensus that 'protein clusters' is not working well. Maybe 'gene clusters' is not working well either, but it seems that is a step towards a bit of a better direction.

@9aren,

So, like our pan genome pipeline which refers to these as GENE FAMILIES would be appropriate as per my belief.

I'm afraid 'gene families' is not any better than 'gene clusters' with respect to existing literature. Gene families are also well defined elsewhere in life sciences, and mean "a set of several similar genes, formed by duplication of a single original gene, and generally with similar biochemical functions [within a single organism]". Additionally, 'clusters' is a more neutral term represent what they really are in a more direct way than 'families' in my opinion. So if both 'gene clusters' and 'gene families' is equally contaminated with previous literature, I think 'gene clusters' is a simpler term.

Thank you for all the input.

[meren]

I think the big problem is the fact that even when we say 'gene' we make assumptions about the nature of these predicted sequences and are given back to us by a software that solely operates on sequencing data.

The "cluster" part is OK. The rest is problematic, in my opinion.

"Gene clusters" simplifies this to a degree while minimizing the amount of assumptions we make. But for instance, Prodigal often identifies two putative genes within the 16S rRNA gene (so it doesn't identify 16S rRNA gene as a gene, but often makes two short gene calls from within that gene when I carefully inspect results).

Maybe 'predicted gene clusters' is much more appropriate, and will fit in most cases, but I am not sure if it would get enough support.

I guess in our small world in my group, we will continue Gene Clusters if no one else suggests anything else.

PS: Thanks, science, for this Saturday morning depression :/

[ShaiberAlon]

@meren

but "homologous groups" would have been even less helpful since the term "group" or "cluster" already implies homology.

True, but if we want to solve to issue of using an established term (i.e. "gene clusters"), then "clusters of homologous genes" is, in my opinion, an appropriate description. Even though, as you say, cluster implies homology, but the term homology is not used in a general manner here; homologous genes are well-defined in our field, and we use NCBI-BLAST (or diamond or a different tool) to try to identify homology, and then use MCL (or other algorithms) to cluster them into clusters of homologous genes.

[simroux]

@meren I agree, "protein cluster" seems to be an issue for a number of folks (way more than I expected), and the issues raised are all valid. If we are looking for a denomination with the least assumption possible, then "ORF cluster" might be it, although it doesn't cover 16S, and it's kind of ugly. Everything else can be "attacked", especially calling predicted ORFs "proteins" or "genes", or calling automatically defined clusters of predicted ORFs "homologs". I also would vote for Pat Schloss discussion of "operational protein families", and maybe (although it's coining a new term we might not need), you might want to take this a step further and call these "OPUs", "Operational Protein Units" ? At least it would convene the point that, as OTUs vs Species, these clusters are proxies for groups of homologous genes.

[tdelmont]

Hi,

Could someone tell me why "sequence cluster" is a bad option?

A quick google search brings me to this:

In bioinformatics, sequence clustering algorithms attempt to group biological sequences that are somehow related. The sequences can be either of genomic, "transcriptomic" (ESTs) or protein origin. (...) Sequence clusters are often synonymous with (but not identical to) protein families.

"Sequence cluster" is a very neutral terminology, which might be a good thing for the matter at hand.

[rikander]

I think the most important thing here is for the community to find a term that we can agree is the least problematic and stick to it-- kind of like the way we circled around the terms "bin" and "population genome" and settled on "metagenome assembled genome" (MAG). I am wary of anything using the word "family" because this has a very specific meaning in the literature (i.e. see Pfam), and I agree that anything using the word "protein" makes the assumption that the sequence in question will be translated into a protein. Too many words (i.e. "clusters of homologous genes") gets unwieldy. "Gene cluster" works, and I like Tom's "sequence cluster" suggestion.

[rbeinart]

I'm comfortable with either "gene clusters" or "protein clusters" as long as it's defined. I also would vote for "homologous gene/protein clusters" or similar (i.e., HPCs, or HGCs). My only problem with "sequence cluster" is that it doesn't indicate the unit being clustered, so it's a less intuitive term. But I agree with rikander that the main thing is to just find a term that's not already used, define it, and stick with it.

[meren]

@simroux,

I also would vote for Pat Schloss discussion of "operational protein families"

It reminds me of OTUs, and makes me think that maybe we don't need a yet-another-acronym-where-every-letter-brings-its-own-baggage. Also, the study (in which this term is defined) is not related to pangenomes since these units the authors generated from open reading frames are not associated with any genomic context. The study proposes that OPFs are analogous to 16S OTUs to make comparisons among communities.. Historically important study, but I think the term it defined does not have any relevance here.

[meren]

@ShaiberAlon,

"clusters of homologous genes" is, in my opinion, an appropriate description

I really like this and I will tell you why :)

We are using homology (currently it is at the sequence-level, but in the future it can be at the functional-level, or a combination of both as people tried before). We are doing it things that are very close to genes (because the information often comes from gene callers, etc). And we group them into clusters. All checks out. And your suggestion addresses a subtle point: we are truly clustering homologous genes. Here is a very basic example to elaborate: Genes A and A' are homologous (very similar at the DNA or AA sequence-level, and let's say they have the same evolutionary origin). They are identified as homologs during the process due to homology search based on whatever. Then let's say we have genes B and B'. Which are also very similar to each other and are identified as homologs. The thing is, in a pangenomic analysis, A, A', B, and B', can end up in the same cluster if they share a domain that makes them somewhat similar to each other depending on the parameters. So, the units we are working with in fact contains one or more homologous groups of genes. We do create clusters of homologous genes. A and B or A and A' may be paralogs or orthologs, and it would not change the fact that they are clusters of homologous genes. It sets a very accurate level of specificity in my opinion.

Now I will tell you why I don't like it :/

While it sets a very accurate level of specificity, it doesn't remove the need to find something better to replace the word 'gene', and it doesn't add something to bring the predicted nature of this information into the mix. So the additional value to go from 'gene clusters' to 'clusters of homologous genes' is not justified in the expense of adding two more words (as @rikander mentioned above). Why does this matter? Because if we call these 'gene clusters' (and define them clearly every time we first mention them), we don't have to use an acronym. We can say gene clusters every time (I've tried it while writing this paper, and it reads rather well). If we call them "clusters of homologous genes", then we can't spell it out every time (because it is 4 words and will impact the flow too often), and we will have to use an acronym, such as CHGs. I think we are all tired of acronyms (because, you know, we are not immunologists), and in my opinion it would be great if we end up don't introduce a new acronym.

My 1½ cents.

[jreveillaud]

Well, I think this is a very interesting discussion and that there might not be a single consensus solution in there. why? because science. Which suggests that different 'term' would work if it comes with a clear definition along (like @rbeinart 'I'm comfortable with either "gene clusters" or "protein clusters" as long as it's defined'). Including that these are 'predicted genes' and might contain errors, ORF translated into AA and that homologs are grouped together into clusters based on sequence comparison (not function here). I vote for a detailed blog post instead of a new acronym.

[meren]

Thanks for the input, @jreveillaud. I will do my best to summarize all these points in a blog post for future references.

While I agree that it is OK to not have a single consensus, sadly, at least in the context of anvi'o, we have to pick a single term due to practical reasons (which will impact interfaces, tutorials, and the code behind).

I am aware that the discussion we are having (and its result) will have no binding power outside of the context of the anvi'o codebase. But we (the anvi'o developers) would still be much more comfortable if we present people (who wish to do pangenomic analyses using this tool) with terms with which the community does not disagree :)