HighFold2 can accurately predict the structures of cyclic peptides containing unnatural amino acids and their complexes. It involves training a deep learning model based on AlphaFold-Multimer using linear peptide structures with unnatural amino acids, then modifying the model's relative position encoding matrix, enabling it to predict the cyclic peptide structures successfully. Then, relaxation is performed to refine the spatial structure further. We believe that this method will serve as a powerful tool for the development of cyclic peptide-based therapeutics.
HighFold2 is based on LocalColabFold, which is available at https://github.com/YoshitakaMo/localcolabfold, so you should install it first. Then, you should download our source code as follows:
git clone https://github.com/hongliangduan/HighFold2.git
cd HighFold2After, you should replace the original alphafold and colabfold folders in /localcolabfold/colabfold-conda/lib/python3.10/site-packages/ with the ones from the modify_code folder. For relaxation and training, you should also install some other packages. All packages we used can be seen in the requirement.txt.
For the training, you should be in the modified locallocalfold environment and the HighFold2 folder and obtain the feature firstly as follows:
python gen_feature.pyAfter getting the feature, you should change the directory to the training folder and copy the original AlphaFold-Multimer parameters to the colabfold folder, then run the training.py as follows:
cd training
python training.pyFor the prediction, you should be in the HighFold2 folder and copy the trained parameters to the colabfold/params folder. Then, prepare the fasta file for your sequence, all the unnatural residues are replaced with 'X' in fasta files. Finally, enter a command similar to the following in the terminal:
colabfold_batch --model-type alphafold2_multimer_v3 {fasta_path} {output_path} --unnatural_residue {unnatural_amino_acids} --flag-cyclic-peptide 1 --flag-nc 1 --amberAll the flags useful for prediction can be seen by typing the following command in the terminal:
colabfold_batch -h