Request for training about alphalink

[lhatsk]

Hi,

You should be able to train it using train_openfold.py. It's set up to do fine-tuning by initialising the network with the original weights: https://github.com/lhatsk/AlphaLink/blob/main/train_openfold.py#L135

[sehooni]

Thank you for reply my issue.

By the way, than how to set the initial dataset?

In the paper, you said that 'fine the network on 13,000 proteins from the trRosetta38 training set with simulated photo-AA crosslinking data.'.

And I already downloaded the trRosetta training set, but it consisted with 15,000 proteins which didn't match 13,000. Furthermore, how can I preprocess it with AA-crosslinking data?

Additionally, I want to ask how to consist of the data with crop size!
Thank you for reading.

[lhatsk]

I don't quite remember why I didn't take the full set. There are two possibilities that come to mind: I think I originally had a train/ validation split on the trRosetta training data from a previous project and only used the training split. Another possibility is that I didn't have crosslinks for all the targets. Just take as much as you can. You would need to simulate the photoAA crosslinking data. I did it on the fly during training.

Additionally, I want to ask how to consist of the data with crop size!
I used the default crop size for fine-tuning which is 384AA.

[sehooni]

thank you for explaining for my question.

Additionally, I have some more questions.

I used the default crop size for fine-tuning which is 384AA.

1. In this case, I want to know the way to crop methodolgy.
   Did you choose overlap strategy for cropping AA? or using consequent cropping AA?

[lhatsk]

I didn't alter the cropping strategy. It's a random crop of size 384AA.