Goal: Sequence, Structure of Complex, Binding Affinity -> Sequence' -> Complex' -> Binding Affinity'
-
First issue: Binding Affinity Prediction is unreliable. Need to test a combination of filters that achieves good binding affinity prediction
-
Second issue: We must find a way to be fairly confident that the fold of sequence' and epitope' are not changed significantly. Current idea is to try to use phylogenetic epitope consensus, Rosetta relax
- Select fewer models (Maybe 3 is enough)
- USP-ddG
- some LM
- a Physics based method
- Separate train test split using TM score and seq identity
- Test the difference due to diversity and measure performance increase from perfect data
- Test whether we can phylogenetically + using rosetta relax predict whether the epitope changes
MSA code + Getting related PDBs code
Get mutations
3 - Get USP-ddG working, setup SKEMPI, Separate train and test split (start out using seq identity). Start the lit
4 - Get Rosetta Relax working, Get a ddG LM working. Create the 4 datasets and train USP-ddG on them
5 - Get OAS structures and check if phylogenetic consensus binding is working