Aligning new sequences to an existing HMM and updating it iteratively?

[pentamorfico]

Hi!

I've been using your tool quite a lot and it's been giving me great results — it's also incredibly powerful for large-scale data!

I was wondering if the following would be possible (or perhaps already supported):
If I have a reference HMM, would it be possible to use learnMSA to align new sequences against that HMM, and update the model in successive rounds?

My idea is to start from a high-quality HMM built from structural alignments (e.g. only including proteins with known PDB/Alphafold structures), and then progressively add sequences that don't have structural data available.

Does that sound too far-fetched? Would it be feasible to do something like this with learnMSA?

Thanks in advance for the amazing work and for any guidance you can provide!

[felbecker]

Hi!

thanks for reaching out. What you suggest sounds very reasonable. However, it would require initializing the HMM from a file format such as HMMER, which unfortunately is currently not supported due to some technical difficulties, but it's something I have on my to-do list.

Could you elaborate on how you’d ideally like to provide an existing HMM to learnMSA? I’ll look into ways of supporting this.

Would it be sufficient—or perhaps even preferable—to initialize with an existing (structural) MSA rather than an HMM?

Best,
Felix

[pentamorfico]

Hi Felix!

Thanks a lot for the fast reply!

The main two applications/utilities where I think this could be useful are like this:

• Use an already existing HMM (for example from PFAM) to guide an alignment. This could work to align very large protein families an update MSAs or HMMs in an incremental way in a fast and accurate way. It could be useful to improve sensitivity of HMMs when supplemented with more divergent sequences (that will further improve/re-train the existing HMM) and to align large protein families/domains that are well-known and available in dbs.

• Scale structure-guided MSAs to millions of proteins. Doing a structure-guided MSA of tons of structures is slow or just impossible, so a workaround would be to generate a MSA with a representative subset of the whole proteome of interst, and then use learnMSA from that starting MSA to align the rest of the proteins.

So I think for both cases it would work with the MSA! In the first case, perhaps hmmemit can be used to generate the MSA when the source MSA is not available?

Let me know if I can help with anything 👍

[felbecker]

Hi there,

yeah, initializing with an HMM or an MSA should be almost the same thing, as you can get one from the other rather easily. The problem with external HMMs is that learnMSA's HMM has additional parameters that need to be initialized properly without breaking the whole thing. Initializing with an MSA is preferred, because learnMSA will easily be able to fit all its parameters on that data. If you only have an HMM, but no MSA and no sequences, we could indeed use sampling, as you suggested.

That being said, I'll implement a mode for learnMSA that will allow you to start from an existing MSA and then learn on larger set of unaligned sequences. I'm not sure when I will be able to start working on this, but earliest in October.

One final remark: we do have (very early) plans for actual large-scale multiple structure alignment using learnMSA. That might address your problem more directly, though that’s still a longer-term goal.

Best,
Felix