At the moment, relative MSA lookups seem to work relative to the current working directory. I would like to propose a change in behavior where the MSA file is looked up relative to the directory of the input file. In this way, I do not need to update the input fasta/yaml if I run the model from a different directory. For me, this would make data exchange much easier, as I only need to preserve my directory layout for the exchange.

I understand that this is a change in the user interface and that there might be good reasons for the current behavior. As the code changes were rather minimal, I decided to go ahead an just open a PR. At the same time, I understand if further discussions are necessary. I am also completely open to the conclusion that this change proposal should be disregarded.

Example

Directory structure

data
├── fastas
│   └── relative.fasta
├── msas
│   └── 1e73__1__2.A__2.I.a3m
└── yamls
    └── relative.yaml

fasta input

>A|protein|../msas/1e73__1__2.A__2.I.a3m
DEEITCQENLPFTCGNTDALNSSSFSSDFIFGVASSAYQIEGTIGRGLNIWDGFTHRYPNKSGPDHGNGDTTCDSFSYWQKDIDVLDELNATGYRFSIAWSRIIPRGKRSRGVNEKGIDYYHGLISGLIKKGITPFVTLFHWDLPQTLQDEYEGFLDPQIIDDFKDYADLCFEEFGDSVKYWLTINQLYSVPTRGYGSALDAPGRCSPTVDPSCYAGNSSTEPYIVAHHQLLAHAKVVDLYRKNYTHQGGKIGPTMITRWFLPYNDTDRHSIAATERMKEFFLGWFMGPLTNGTYPQIMIDTVGERLPSFSPEESNLVKGSYDFLGLNYYFTQYAQPSPNPVNSTNHTAMMDAGAKLTYINASGHYIGPLFEKDKADSTDNIYYYPKGIYSVMDYFKNKYYNPLIYVTENGISTPGDENRNQSMLDYTRIDYLCSHLCFLNKVIKEKDVNVKGYLAWALGDNYEFNKGFTVRFGLSYIDWNNVTDRDLKKSGQWYQSFISP
>B|smiles
CC(=O)N[C@H]1CO[C@H](CO)[C@@H](O)[C@@H]1O

yaml input

version: 1
sequences:
  - protein:
      id: A
      sequence: DEEITCQENLPFTCGNTDALNSSSFSSDFIFGVASSAYQIEGTIGRGLNIWDGFTHRYPNKSGPDHGNGDTTCDSFSYWQKDIDVLDELNATGYRFSIAWSRIIPRGKRSRGVNEKGIDYYHGLISGLIKKGITPFVTLFHWDLPQTLQDEYEGFLDPQIIDDFKDYADLCFEEFGDSVKYWLTINQLYSVPTRGYGSALDAPGRCSPTVDPSCYAGNSSTEPYIVAHHQLLAHAKVVDLYRKNYTHQGGKIGPTMITRWFLPYNDTDRHSIAATERMKEFFLGWFMGPLTNGTYPQIMIDTVGERLPSFSPEESNLVKGSYDFLGLNYYFTQYAQPSPNPVNSTNHTAMMDAGAKLTYINASGHYIGPLFEKDKADSTDNIYYYPKGIYSVMDYFKNKYYNPLIYVTENGISTPGDENRNQSMLDYTRIDYLCSHLCFLNKVIKEKDVNVKGYLAWALGDNYEFNKGFTVRFGLSYIDWNNVTDRDLKKSGQWYQSFISP
      msa: ../msas/1e73__1__2.A__2.I.a3m
  - ligand:
      id: B
      smiles: CC(=O)N[C@H]1CO[C@H](CO)[C@@H](O)[C@@H]1O

Run command

boltz predict /data/fastas/relative.fasta

or

boltz predict /data/yamls/relative.yaml

I am happy to provide the MSA file. It seems to be too large to attach to the post.

Expected behavior

The prediction should work.

Observed behavior

Traceback (most recent call last):
  File "/xxx/xxx/boltz/src/boltz/main.py", line 608, in process_input
    raise FileNotFoundError(msg)  # noqa: TRY301
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FileNotFoundError: MSA file ../msas/1e73__1__2.A__2.I.a3m not found.
Failed to process /data/boltz_msa_example/yamls/relative.yaml. Skipping. Error: MSA file ../msas/1e73__1__2.A__2.I.a3m not found..