Proposal for merging input schemas into this repo

[y1zhou]

Hi @hgbrian, I've been working on a data schema package (https://github.com/y1zhou/uniaf3) to convert between inputs for AF3 and AF3-like models. My next step was to come up with some unified MSA search process and sharing results between the models. Once that's done I was thinking of building some interface for actually running the models on modal, which seems to be already done here in your repo. Do you think there's possibility for combining the repos into a full-fledged structure modeling app that supports ligands, contact restraints, covalent bonds, etc. and can run all models with the same set of inputs? Cheers

[hgbrian]

Hi @y1zhou thanks for the note, and nice work on uniaf3! It looks very comprehensive.

I am all for reusing code! I'm not sure what the next step would be though?
foldism is pretty self-contained right now, and focused on one use-case for the non-power user (e.g., it's protein only)

I definitely think there could be a more complete tool that includes more options. I'm not sure how best to achieve that though

[y1zhou]

I'll try to go through foldism and understand your current architecture. UI-wise perhaps something like a "Advanced options" section that's hidden by default containing a GeoFlow style residue selector for specifying restraints, and maybe also a separate section for model parameters like number of samples, recycles, diffusion steps, seeds, etc.? Happy to discuss and contribute!

[hgbrian]

Sounds great! Thanks for taking the time to dig in.

I am still undecided if this should be a new power-user tool vs this "easy" tool.
I personally use constraints, or other such options, very rarely, even as a "power user".
IIRC, alphafoldserver also seems to have no real options.

The things that are important to me:

• simple extensible architecture: one main file, foldism.py, and one python file per backend
• no params/options required for general use

Ideally, the advanced options would be wholly segregated in the codebase, perhaps even a different app that could be run: e.g., foldism_advanced.py

The goal is to keep things simple, and reduce interdependencies. For example, adding a new backend should not trigger any significant edits in any other files in the system.

[y1zhou]

Makes sense! I also believe the community (especially people from a wetlab background) could benefit from a simple interface without the daunting knobs to tune. I guess my question then is:

1. If it would be good to replace the converter functions with Pydantic schema mappings (similar to what we are trying to acheive with UniAF3 adapters). Personally I think stricter schema definitions would be easier to maintain because you no longer have to guess what the entities are, and with your current setup a new backend would trigger edits in the common module and foldism.py helper functions anyways.

foldism/backends/common.py

Lines 314 to 320 in 77f7e06

	CONVERTERS = {
	"boltz_yaml": _fasta_to_boltz_yaml,
	"chai_fasta": _fasta_to_chai_fasta,
	"protenix_fasta": _normalize_fasta,
	"af2_fasta": _fasta_to_af2_fasta,
	"openfold3_fasta": _normalize_fasta,
	}

2. Do you prefer the paste-FASTA box, or something like the AlphaFold server form where you specify entity types, copies, sequences, etc.? The models seem to have in general shifted away from FASTA inputs because things like PTMs and restraints are difficult to express in FASTA, but in your case of keeping things simple the FASTA is definitely easier to construct and read. If you'd prefer foldism to stay as a pure protein-folding app (no ligands, restraints, PTMs, or other inference argument changes) then the FASTA input would be superior to anything else IMO.

[hgbrian]

Hey Yi,
thanks for your input!

1. That makes sense. I am definitely a minimalist when it comes to this stuff. As long as the code doesn't get too verbose I think it could be a nice improvement to consolidate. If there is a good way to unify, let me know!

2. I think pasting fasta is most familiar to people, and a lot of tools output fasta. alphafoldserver does a good job automatically converting from fasta, but I feel this is kind of fussy. I am not sure right now what the right solution is. I could imagine a fasta input (most people will use that I think), plus an advanced dropdown to paste in ligands, constraints, etc separately?

[hgbrian]

Of course, if you have a vision for a more feature-rich version, feel free to take whatever code! These days, Claude Code makes the burden of maintaining two separate apps pretty light.