search

a-r-j / ProteinWorkshop

Benchmarking framework for protein representation learning. Includes a large number of pre-training and downstream task datasets, models and training/task utilities. (ICLR 2024)
https://proteins.sh/
MIT License
204 stars 16 forks source link

question: fit for interface featurisation? #98

Open paoslaos opened 5 days ago

paoslaos commented 5 days ago

Dear authors,

I stumbled upon this wonderful package and was wondering if it is possible to featurize interfaces between complexes? The goal would be to learn a regression task eventually on binding / DDG.

Thanks in advance.

a-r-j commented 4 days ago

Hi @paoslaos could you please elaborate on exactly how you would like to featurise the interface?

paoslaos commented 4 days ago

Dear @a-r-j , thank you for the question.

In the end it does not really matter but for example, in ProteinMPNN multiple chains can be featurized and the applied distance cuttoffs determine how the feauturisation will look like. It could also be a naive featurisation based on the structure without distance constraints / focus on the interface. I am still learning and exploring - sorry for the imprecise language. My aim is to build a deep learning graph-based model on complexes from mutabind.