DeepFrag is a machine learning model for fragment-based lead optimization. In this repository, you will find code to train the model and code to run inference using a pre-trained model.
If you use DeepFrag in your research, please cite as:
Green, H., Koes, D. R., & Durrant, J. D. (2021). DeepFrag: a deep convolutional neural network for fragment-based lead optimization. Chemical Science.
@article{green2021deepfrag,
title={DeepFrag: a deep convolutional neural network for fragment-based lead optimization},
author={Green, Harrison and Koes, David Ryan and Durrant, Jacob D},
journal={Chemical Science},
year={2021},
publisher={Royal Society of Chemistry}
}
There are three ways to use DeepFrag:
- DeepFrag Browser App: We have released a free, open-source browser app
for DeepFrag that requires no setup and does not transmit any structures to
a remote server.
- View the online version at durrantlab.pitt.edu/deepfrag
- See the code at git.durrantlab.pitt.edu/jdurrant/deepfrag-app
- DeepFrag CLI: In this repository we have included a
deepfrag.py
script that can perform common prediction tasks using the API.- See the
DeepFrag CLI
section below
- See the
- DeepFrag API: For custom tasks or fine-grained control over
predictions, you can invoke the DeepFrag API directly and interface with
the raw data structures and the PyTorch model. We have created an example
Google Colab (Jupyter notebook) that demonstrates how to perform manual
predictions.
- See the interactive Colab.
The DeepFrag CLI is invoked by running python3 deepfrag.py
in this
repository. The CLI requires a pre-trained model and the fragment library to
run. You will be prompted to download both when you first run the CLI and
these will be saved in the ./.store
directory.
The input structures are specified using either a manual receptor and ligand pdb or by specifying a pdb id and the ligand residue number.
--receptor <rec.pdb> --ligand <lig.pdb>
--pdb <pdbid> --resnum <resnum>
DeepFrag will predict new fragments that connect to the connection point via a single bond. You must specify the connection point atom using one of the following:
--cname <name>
: Specify the connection point by atom name (e.g.C3
,N5
,O2
, ...).--cx <x> --cy <y> --cz <z>
: Specify the connection point by atomic coordinate. DeepFrag will find the closest atom to this point.
If you are using DeepFrag for fragment replacement, you must first remove the original fragment from the ligand structure. You can either do this by hand, e.g. editing the PDB, or DeepFrag can do this for you by specifying which fragment should be removed.
Note: predicting fragments in place of hydrogen atoms (e.g. protons) does not require any fragment removal since hydrogen atoms are ignored by the model.
To remove a fragment, you specify a second atom that is contained in the fragment. Like the connection point, you can either use the atom name or the atom coordinate.
--rname <name>
: Specify the connection point by atom name (e.g.C3
,N5
,O2
, ...).--rx <x> --ry <y> --rz <z>
: Specify the connection point by atomic coordinate. DeepFrag will find the closest atom to this point.
By default, DeepFrag will print a list of fragment predictions to stdout similar to the Browser App.
--out <out.csv>
: Save predictions in CSV format toout.csv
. Each line contains the fragment rank, score and SMILES string.
--full
: Generate SMILES strings with the full ligand structure instead of just the fragment. (IMPORTANT NOTE: Bond orders are not assigned to the parent portion of the full ligand structure. These must be added manually.)--cpu/--gpu
: DeepFrag will attempt to infer if a Cuda GPU is available and fallback to the CPU if it is not. You can set either the--cpu
or--gpu
flag to explicitly specify the target device.--num_grids <num>
: Number of grid rotations to use. Using more will take longer but produce a more stable prediction. (Default: 4)--top_k <k>
: Number of predictions to print in stdout. Use -1 to display all. (Default: 25)
You can use the DeepFrag CLI to reproduce the highlighted results from the main manuscript:
To replace fragments, specify the connection point (cname
or cx/cy/cz
) and
specify a second atom that is contained in the fragment (rname
or
rx/ry/rz
).
# Fig. 3: (2XP9) H. sapiens peptidyl-prolyl cis-trans isomerase NIMA-interacting 1 (HsPin1p)
# Carboxylate A
$ python3 deepfrag.py --pdb 2xp9 --resnum 1165 --cname C10 --rname C12
# Phenyl B
$ python3 deepfrag.py --pdb 2xp9 --resnum 1165 --cname C1 --rname C2
# Phenyl C
$ python3 deepfrag.py --pdb 2xp9 --resnum 1165 --cname C18 --rname C19
# Fig. 4A: (6QZ8) Protein myeloid cell leukemia1 (Mcl-1)
# Carboxylate group interacting with R263
$ python3 deepfrag.py --pdb 6qz8 --resnum 401 --cname C12 --rname C14
# Ethyl group
$ python3 deepfrag.py --pdb 6qz8 --resnum 401 --cname C6 --rname C10
# Methyl group
$ python3 deepfrag.py --pdb 6qz8 --resnum 401 --cname C25 --rname C30
# Chlorine atom
$ python3 deepfrag.py --pdb 6qz8 --resnum 401 --cname C28 --rname CL
# Fig. 4B: (1X38) Family GH3 b-D-glucan glucohydrolase (barley)
# Hydroxyl group interacting with R158 and D285
$ python3 deepfrag.py --pdb 1x38 --resnum 1001 --cname C2B --rname O2B
# Phenyl group interacting with W286 and W434
$ python3 deepfrag.py --pdb 1x38 --resnum 1001 --cname C7B --rname C1
# Fig. 4C: (4FOW) NanB sialidase (Streptococcus pneumoniae)
# Amino group
$ python3 deepfrag.py --pdb 4fow --resnum 701 --cname CAE --rname NAA
For fragment addition, you only need to specify the atom connection point
(cname
or cx/cy/cz
). In this case, DeepFrag will implicitly replace a
valent hydrogen.
# Fig. 5: Ligands targeting the SARS-CoV-2 main protease (MPro)
# 5A: (5RGH) Extension on Z1619978933
$ python3 deepfrag.py --pdb 5rgh --resnum 404 --cname C09
# 5B: (5R81) Extension on Z1367324110
$ python3 deepfrag.py --pdb 5r81 --resnum 1001 --cname C07
config
: fixed configuration information (e.g., TRAIN/VAL/TEST partitions)configurations
: benchmark model configurations (seeconfigurations/README.md
)data
: training/inference data (seedata/README.md
)leadopt
: main module codemodels
: pytorch architecture definitionsdata_util.py
: utility code for reading packed fragment/fingerprint data filesgrid_util.py
: GPU-accelerated grid generation codemetrics.py
: pytorch implementations of several metricsmodel_conf.py
: contains code to configure and train modelsutil.py
: utility code for rdkit/openbabel processing
scripts
: data processing scripts (seescripts/README.md
)train.py
: CLI interface to launch training runs
You can build a virtualenv with the requirements:
$ python3 -m venv leadopt_env
$ source ./leadopt_env/bin/activate
$ pip install -r requirements.txt
$ pip install prody
$ pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --index-url https://download.pytorch.org/whl/cu118
$ sudo apt install nvidia-cuda-toolkit
Regarding the nvidia-cuda-toolkit, you may wish to ensure that the toolkit version matches cuda installed on your machine. You can check the version of cuda by running the following commands:
$ nvcc --version
$ nvidia-smi
Note: We used Cuda 10.1
for training.
To train a model, you can use the train.py
utility script. You can specify
model parameters as command line arguments or load parameters from a
configuration args.json file.
python train.py \
--save_path=/path/to/model \
--wandb_project=my_project \
{model_type} \
--model_arg1=x \
--model_arg2=y \
...
or
python train.py \
--save_path=/path/to/model \
--wandb_project=my_project \
--configuration=./configurations/args.json
save_path
is a directory to save the best model. The directory will be
created if it doesn't exist. If this is not provided, the model will not be
saved.
wandb_project
is an optional wandb project name. If provided, the run will
be logged to wandb.
See below for available models and model-specific parameters:
In this repository, trainable models are subclasses of
model_conf.LeadoptModel
. This class encapsulates model configuration
arguments and pytorch models and enables saving and loading multi-component
models.
from leadopt.model_conf import LeadoptModel, MODELS
model = MODELS['voxel']({args...})
model.train(save_path='./mymodel')
...
model2 = LeadoptModel.load('./mymodel')
Internally, model arguments are configured by setting up an argparse
parser
and passing around a dict
of configuration parameters in self._args
.
--no_partitions If set, disable the use of TRAIN/VAL partitions during
training.
-f FRAGMENTS, --fragments FRAGMENTS
Path to fragments file.
-fp FINGERPRINTS, --fingerprints FINGERPRINTS
Path to fingerprints file.
-lr LEARNING_RATE, --learning_rate LEARNING_RATE
--num_epochs NUM_EPOCHS
Number of epochs to train for.
--test_steps TEST_STEPS
Number of evaluation steps per epoch.
-b BATCH_SIZE, --batch_size BATCH_SIZE
--grid_width GRID_WIDTH
--grid_res GRID_RES
--fdist_min FDIST_MIN
Ignore fragments closer to the receptor than this
distance (Angstroms).
--fdist_max FDIST_MAX
Ignore fragments further from the receptor than this
distance (Angstroms).
--fmass_min FMASS_MIN
Ignore fragments smaller than this mass (Daltons).
--fmass_max FMASS_MAX
Ignore fragments larger than this mass (Daltons).
--ignore_receptor
--ignore_parent
-rec_typer {single,single_h,simple,simple_h,desc,desc_h}
-lig_typer {single,single_h,simple,simple_h,desc,desc_h}
-rec_channels REC_CHANNELS
-lig_channels LIG_CHANNELS
--in_channels IN_CHANNELS
--output_size OUTPUT_SIZE
--pad
--blocks BLOCKS [BLOCKS ...]
--fc FC [FC ...]
--use_all_labels
--dist_fn {mse,bce,cos,tanimoto}
--loss {direct,support_v1}