Mismatched Keys in the Validation Step in Boltz1

[SophieSarceau]

Dear Authors,

Thanks for your insightful work.

When I use your training code to train Boltz1 from scratch, I found there are some errors in the training code.

1. Missing max_parallel_samples for the input of self.forward.

2. Key mismatch when updating the LDDT score

"/home/Datasets/biology/boltz/boltz/src/boltz/model/models/boltz1.py",
line 873, in validation_step self.lddt[m].update(best_lddt_dict[m], best_total_dict[m])
KeyError: 'modified'

There is modified key in const.out_types, but there is no such key in best_lddt_dict. Is the training code the latest version? Thanks a lot!

[chmnk]

max_parallel_samples are also missing in the training_step.

I filled the "modified" lddt fields with zeros -- did you manage to find out what is it for? My first guess was modified residues (or whatever else where "modified" token feature is True), but not sure if it is protein-modified or intra-modified lddt. Or something completely different.

[SophieSarceau]

No, I just skipped modified as this indicator seems to have nothing to do with my training target (protein complex).

[alexbui91]

Key 'modified' is added in the commit "afe4aa657e80bc2974c27b7095bc0ff1091d71500". In boltz1, 'modified' key wasn't existed. I think we can just skip this key.