Closed PhungVanDuy closed 2 years ago
Unfortunately, the current model does not provide a score for ranking.
You could retrain a model with the 'random_vec_dim' parameter set to something higher than 0. This model would then be able to produce different suggestions. Also, you could run SMINA on top of the EquiBind pocket prediction to obtain a score and multiple suggestions.
@HannesStark Thank you for your information.
Do you have any suggestions about the value I need to set for random_vec_dim
?
I ususally used 10.
@HannesStark I just finished the training process with random_vec_dim = 5
but can you give me any command to run to get multiple suggestions. I ran python inference.py --config=configs_clean/inference.yml
by replacing the path to the new model checkpoint it still got one suggestion.
You would have to run the model multiple times. Also make sure that you have a different seed everytime when you are running that in e.g. a for loop.
A for loop like that is not currently implemented
@HannesStark I already tried something like that but the result looks still similar, seems it is still the same position to bind.
As I know that, current models only return one output for each pair ligand-receptor. Does the current model extend to support multiple suggestions binding sites output with the ranking?