PepFlow

Description

information from Gitlab: https://gitlab.com/oabdin/pepflow

PepFlow is a deep-learning method for direct all-atom peptide conformational sampling. The model is a Boltzmann generator trained in a diffusion framework and subsequently used as a flow for training by energy and sampling. PepFlow contains a number of provisions that enable the modelling of all the degrees of freedom in a peptide conformation. The details of this approach can be found in the associated preprint:

Abdin O, Kim PM. Pepflow: direct conformational sampling from peptide energy landscapes through hypernetwork-conditioned diffusion. bioRxiv. 2023. https://doi.org/10.1101/2023.06.25.546443

Installation

To install PepFlow and the necessary dependancies run the following commands, installation should take less than 15 minutes:

1. Clone the repository git clone https://gitlab.com/oabdin/pepflow.git
2. Create the conda environment conda env create -f pepflow_env.yml
3. Activate the conda environment conda activate pepflow_env
4. Install remaining requirements:
   cuda-compiled version of torch-scatter pip install torch-scatter -f https://data.pyg.org/whl/torch-1.13.1+cu117.html
   ml-collections pip install ml-collections==0.1.1
5. Install the pepflow package pip install .

To run PepFlow, download and extract the model parameters, http://pepflow.ccbr.proteinsolver.org/params.tar.gz.

Peptide structure prediction and ensemble generation

To predict peptide structures use the generate_peptide_samples.py with the following options:

-fm full model parameters file, either this parameter needs to be specified or all of -bm, -rm and -hm need to be specified
-bm backbone model parameters file
-rm rotamer model parameters file
-hm hydrogen model parameters file
-o output directory

-n number of samples to generate
-c number of samples per chunk to use when generating the samples. This value can be reduced to decrease memory usage at the cost of slower sampling, recommended value is 10

-s the sequence for which samples should be generated

--s flag to generate a single structure prediction

--r flag to not generate side chain heavy atoms
--p flag to not protonate generated peptides
--f flag to not correct the chirality of generated peptides
--d flag to generate D-amino acid peptides (if --f is not set)

--l flag to output likelihoods of generated peptides
--e flag to output energies of generated peptides

Examples

The following command generates 100 conformations for the peptide LQTKLKKLLGLESVF using PepFlow after training by energy, takes 3-4 minutes on an NVIDIA GeForce RTX 2070 SUPER:

python generate_peptide_samples.py -s LQTKLKKLLGLESVF -o LQTKLKKLLGLESVF_samples --e -fm params/full_model.pth -n 100 -c 10

The following command generates 100 conformations and a single-structure prediction for the peptide LKKLWRFLKKL using PepFlow that finetuned on peptide structures, takes 2-3 minutes on an NVIDIA GeForce RTX 2070 SUPER:

python generate_peptide_samples.py -s LKKLWRFLKKL -o LKKLWRFLKKL_samples --s -bm params/params_backbone_finetuned.pth -hm params/params_hydrogen.pth -rm params/params_rotamer.pth -n 100 -c 10

Example outputs are provided in the examples/ folder.

Macrocyclic peptide conformation prediction

To predict macrocyclic peptide conformations use the generate_cyclic_peptide_samples.py script with the following options:

-fm full model parameters file, either this parameter needs to be specified or all of -bm, -rm and -hm need to be specified
-bm backbone model parameters file
-rm rotamer model parameters file
-hm hydrogen model parameters file
-o output directory

-n number of samples to generate
-c number of samples per chunk to use when generating the samples. This value can be reduced to decrease memory usage at the cost of slower sampling, recommended value is 5
-st number of MCMC steps, recommended value is 250

-s the sequence for which samples should be generated
-b a file with a list of cyclic bonds. Each line in the file should contain two atoms separated by a comma, with each atom expressed as the amino acid number and the atom name separated by an underscore. Example files can be found in the examples/ folder

--s flag to generate a single structure prediction

--f flag to not correct the chirality of generated peptides
--d flag to generate D-amino acid peptides (if --f is not set)

Examples

The following command generates 25 conformations and a single structure prediction of the peptide GRCTKSIPPRCFPD with head-to-tail cyclization, takes ~13 minutes on an NVIDIA GeForce RTX 2070 SUPER:

python generate_cyclic_peptide_samples.py -s GRCTKSIPPRCFPD -b examples/bonds_GRCTKSIPPRCFPD.txt -c 5 -n 25 --s -o GRCTKSIPPRCFPD_samples -fm params/full_model.pth -st 100

The following command generates 25 conformations and a single structure prediction of the peptide EADKWQS with cyclization between the C-beta of the A2 and NE2 of Q6, takes ~22 minutes on an NVIDIA GeForce RTX 2070 SUPER:

python generate_cyclic_peptide_samples.py -s EADKWQS -b examples/bonds_EADKWQS.txt -c 5 -n 25 --s -o EADKWQS_samples -fm params/full_model.pth

Example outputs are provided in the examples/ folder.

Datasets

Datasets and processed features for training an evaluation can be downloaded from http://pepflow.ccbr.proteinsolver.org/. If md_pdbs is used please cite DBAASP 3.0 (https://doi.org/10.1093/nar/gkaa991) and if PED-derived peptides are used please cite the PED (https://doi.org/10.1093/nar/gkaa1021). Scripts for training and evaluation are present in the scripts folder.