Different length weights / msa

[hredestig]

On the example dataset, the list of sequence weights has a different length compared to MSA. This is presumably inherited from plmc dropping invalid sequences, but would need to think about how to allow mapping the weights back to individual sequences for these values to be of any real value.

[hredestig]

On the A0A1I9GEU1_NEIME_Kennouche_2019 example dataset, the list of sequence weights has a different length compared to MSA. This is presumably inherited from plmc dropping invalid sequences, but would need to think about how to allow mapping the weights back to individual sequences for these to be of any real value.

[karel-w]

I think the question should be more do we need to have some over arching validation between sequence, structure, msa and msa weights? Such that when we extract the information for the ML models it is aligned, or we know beforehand that we sent an unaligned dataset.

[hredestig]

We should only allow coherent datasets imo: use pydantic model validation to make sure manifest makes sense in its entirety.