Question on reference conformers used and training protocol

[FranklinHu1]

Hi there,

Thanks for sharing your excellent paper and code! I just had two quick question about the reference conformers that you use as input for your model and your training protocol:

1. In the paper, it states that these local structure predicted at the amino acid level using rule-based cheminformatics methods. Does this mean part of the input to the model are the structures of each amino acid generated using something like RDKit? Or am I misunderstanding the nature of the reference position? Also, has any work been done to quantify how big of an impact this input has on the quality of the generated protein structures?
2. In your default config file for the trainer, the setting is 100000 training steps with a val check interval of 5000. However, the model in src/simplefold/simplefold.py has no logic in its validation_step() method. Does this mean validation was not performed during training and the model was trained for 100000 steps before evaluation?

Any clarification would be greatly appreciated, and thank you very much!

[yuyangw]

Hi @FranklinHu1, thanks for your interest in our work. To answer your question:

1. Yes, input to the model includes structures of single amino acids from rule based method, following the setting as AlphaFold. We didn't evaluate the effect of this impact, but my bold guess is it won't largely affect the performance.
2. The validation check interval is used for checkpoint saving only, namely we don't do validation on the fly. And the model is trained for ~470K iterations in total, which is described in our preprint paper.