Umol - Universal molecular framework

Structure prediction of protein-ligand complexes from sequence information

The protein is represented with a multiple sequence alignment and the ligand as a SMILES string, allowing for unconstrained flexibility in the protein-ligand interface. At a high accuracy threshold, unseen protein-ligand complexes can be predicted more accurately than for RoseTTAFold-AA, and at medium accuracy even classical docking methods that use known protein structures as input are surpassed.

Read the paper here

Umol is available under the Apache License, Version 2.0.
The Umol parameters are made available under the terms of the CC BY 4.0 license.

Colab (run Umol in the browser)

Colab Notebook

Local installation

The entire installation takes <1 hour on a standard computer.
The runtime will depend on the GPU you have available and the size of the protein-ligand complex you are predicting.
On an NVIDIA A100 GPU, the prediction time is a few minutes on average.

Install Python packages (several minutes)

For the python environment, we recommend to install it with pip as described below.
You can do this in your virtual environment of choice.

pip install -U jaxlib==0.3.24+cuda11.cudnn82 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
pip install jax==0.3.24
pip install ml-collections==0.1.1
pip install dm-haiku==0.0.9
pip install pandas==1.3.5
pip install biopython==1.81
pip install chex==0.1.5
pip install dm-tree==0.1.8
pip install immutabledict==2.0.0
pip install numpy==1.21.6
pip install scipy==1.7.3
pip install tensorflow==2.12.0
pip install optax==0.1.4
pip install rdkit-pypi

You also have to install a conda environment for OpenMM used in the protein relaxation. To do this, first install miniconda, see: https://docs.conda.io/projects/miniconda/en/latest/miniconda-install.html

conda create -c conda-forge --name openmm openmm
conda activate openmm #Assumes you have conda in your path
conda install -c conda-forge openff-toolkit
conda install -c conda-forge pdbfixer
pip install --upgrade lxml

Get network parameters for Umol (a few minutes)

wget https://zenodo.org/records/10048543/files/params40000.npy
mkdir data/params
mv params40000.npy  data/params/

Get Uniclust30 (10-20 minutes depending on bandwidth)

25 Gb download, 87 Gb extracted

wget http://wwwuser.gwdg.de/~compbiol/uniclust/2018_08/uniclust30_2018_08_hhsuite.tar.gz --no-check-certificate
mkdir data/uniclust30
mv uniclust30_2018_08_hhsuite.tar.gz data/uniclust30
cd data/uniclust30
tar -zxvf uniclust30_2018_08_hhsuite.tar.gz
cd ../..

Install HHblits (a few minutes)

git clone https://github.com/soedinglab/hh-suite.git
mkdir -p hh-suite/build && cd hh-suite/build
cmake -DCMAKE_INSTALL_PREFIX=. ..
make -j 4 && make install
cd ../..
mv data/uniclust30/uniclust30_2018_08 data/
rm -r data/uniclust30

Run the test case

ID=7NB4
FASTA=./data/test_case/7NB4/7NB4.fasta
POCKET_INDICES=./data/test_case/7NB4/7NB4_pocket_indices.npy #Zero indexed numpy array of what residues are in the pocket
LIGAND_SMILES='CCc1sc2ncnc(N[C@H](Cc3ccccc3)C(=O)O)c2c1-c1cccc(Cl)c1C' #Make sure these are canonical as in RDKit
UNICLUST=./data/uniclust30_2018_08/uniclust30_2018_08
OUTDIR=./data/test_case/7NB4/

Search Uniclust30 with HHblits to generate an MSA (a few minutes)

HHBLITS=./hh-suite/build/bin/hhblits
$HHBLITS -i $FASTA -d $UNICLUST -E 0.001 -all -oa3m $OUTDIR/$ID'.a3m'

Generate input feats (seconds)

python3 ./src/make_msa_seq_feats.py --input_fasta_path $FASTA \
--input_msas $OUTDIR/$ID'.a3m' \
--outdir $OUTDIR

python3 ./src/make_ligand_feats.py --input_smiles $LIGAND_SMILES \
--outdir $OUTDIR

Predict (a few minutes)

MSA_FEATS=$OUTDIR/msa_features.pkl
LIGAND_FEATS=$OUTDIR/ligand_inp_features.pkl
PARAMS=data/params/params40000.npy
NUM_RECYCLES=3

python3 ./src/predict.py --msa_features  $MSA_FEATS \
--ligand_features $LIGAND_FEATS \
--id $ID \
--ckpt_params $PARAMS \
--target_pos $POCKET_INDICES \
--num_recycles $NUM_RECYCLES \
--outdir $OUTDIR

wait
RAW_PDB=$OUTDIR/$ID'_pred_raw.pdb'
python3 ./src/relax/align_ligand_conformer.py --pred_pdb $RAW_PDB \
--ligand_smiles $LIGAND_SMILES --outdir $OUTDIR

grep ATOM $OUTDIR/$ID'_pred_raw.pdb' > $OUTDIR/$ID'_pred_protein.pdb'
echo "The unrelaxed predicted protein can be found at $OUTDIR/$ID'_pred_protein.pdb' and the ligand at $OUTDIR/$ID'_pred_ligand.sdf'"

Relax the protein (a few minutes)

This fixes clashes mainly in the protein, but also in the protein-ligand interface.

conda activate openmm #Assumes you have conda in your path
PRED_PROTEIN=$OUTDIR/$ID'_pred_protein.pdb'
PRED_LIGAND=$OUTDIR/$ID'_pred_ligand.sdf'
RESTRAINTS="CA+ligand" # or "protein"

python3 ./src/relax/openmm_relax.py --input_pdb $PRED_PROTEIN \
                        --ligand_sdf $PRED_LIGAND \
                        --file_name $ID \
                        --restraint_type $RESTRAINTS \
                        --outdir $OUTDIR

#Deactivate conda - only for the relaxation
conda deactivate
RAW_COMPLEX=$OUTDIR/$ID'_pred_raw.pdb'
RELAXED_COMPLEX=$OUTDIR/$ID'_relaxed_complex.pdb'
python3 ./src/relax/add_plddt_to_relaxed.py  --raw_complex $RAW_COMPLEX \
--relaxed_complex $RELAXED_COMPLEX  \
--outdir $OUTDIR
echo "The final relaxed structure can be found at $OUTDIR/$ID'_relaxed_plddt.pdb'"

xinluzhu/Umol