[AtharvaTilewale]

Problem

A custom a3m for a complex that contains only paired sequences (every row has residues in every segment, no unpaired block) aborts prediction with

Could not generate input features ...: MSA 0 must contain at least one sequence.

both in colabfold_batch and in the notebook. Reported in #770.

Root cause

unserialize_msa routes every row that has residues in more than one segment to the paired chunks, so for a paired-only a3m the unpaired MSA comes back as a list of empty strings (['', '']), not as None.
generate_input_feature only falls back to the single-sequence MSA when unpaired_msa is None, so it passes '' to build_monomer_feature and pipeline.make_msa_features raises MSA 0 must contain at least one sequence.

What the fix does

Treat an empty MSA entry like None in generate_input_feature, i.e. use >{101 + sequence_index}\n{sequence} for that chain. The same guard is applied to the paired MSA in the multimer branch, which has the symmetric problem for an a3m with only unpaired sequences.

Validation

Added tests/test_paired_only_custom_msa in tests/test_msa.py: it builds a paired-only heterodimer a3m, runs it through unserialize_msa (asserting the unpaired MSA really is ['', '']) and then through generate_input_feature with a multimer model type.

Pre-patch (git stash push colabfold/batch.py, test kept):

$ pytest tests/test_msa.py::test_paired_only_custom_msa -q
    raise ValueError(f'MSA {msa_index} must contain at least one sequence.')
E   ValueError: MSA 0 must contain at least one sequence.
1 failed

Post-patch:

$ pytest tests/test_msa.py -q
2 passed

$ pytest -q
29 passed, 20 skipped, 7 errors

The 7 errors are tests/test_colabfold.py prediction tests, which need network access to download parameters/MSAs; they error identically on unmodified main in this environment (baseline there is 28 passed, 20 skipped, 7 errors).

Fixes #770

Disclosure: this change was prepared with AI assistance and validated locally as described above.