Incorporating Small Molecule-Protein Contact and Pocket Restraints in Chai-1 for Improved Complex Structure Accuracy

[gibhdyw]

If I want to use contact or pocket restraints between a small molecule (e.g., in SMILES format) and a protein to improve the accuracy of the complex structure, how does Chai-1 handle these restraints during execution? Does the model strictly adhere to these restraints during the inference process, and if so, how does it ensure that these constraints are integrated into the folding process?

[wukevin]

Chai-1 is trained to fold complexes given a set of input features including input sequence, MSA, protein language model embeddings, etc. Among this set of input features are the restraints that you are referring to, and thus they are handled like any other input feature -- passed through the model to influence the final predicted complex. Like any input feature to a complex model, it is difficult to have the model "strictly adhere to the restraint" so to speak.

That being said, we have observed in internal testing that restraints can have a tremendous impact on folding accuracy, though this is not a perfect guarantee.

[gibhdyw]

Chai-1 is trained to fold complexes given a set of input features including input sequence, MSA, protein language model embeddings, etc. Among this set of input features are the restraints that you are referring to, and thus they are handled like any other input feature -- passed through the model to influence the final predicted complex. Like any input feature to a complex model, it is difficult to have the model "strictly adhere to the restraint" so to speak.

That being said, we have observed in internal testing that restraints can have a tremendous impact on folding accuracy, though this is not a perfect guarantee.

Thank you for your reply. Do you have any teaching examples available? Also, could you please consider suggesting that Chai provides a bulk download feature in the server version? Currently, downloading from your official website is quite inconvenient