Virtual screening

Setup

from kdock.af3.json import *
from kdock.af3.docker import *
import pandas as pd

Run single protein default pipeline

Json:

data = get_protein_json('proteinA','AAA','data/proteinA.json',seeds=[1,2,3])

data

{'name': 'proteinA',
 'modelSeeds': [1, 2, 3],
 'sequences': [{'protein': {'id': 'A', 'sequence': 'AAA'}}],
 'bondedAtomPairs': [],
 'dialect': 'alphafold3',
 'version': 3}

Docker command

Move the generated proteinA.json to the af_input/project_name folder

project_name='sdf'

docker_single_full(json_path=f"af_input/{project_name}/proteinA.json",
                               output_dir=f"af_output/{project_name}")

docker run --rm \
    --volume "$HOME/af_input:/root/af_input" \
    --volume "$HOME/af_output/sdf:/root/af_output" \
    --volume "$HOME/af_model:/root/models" \
    --volume "$HOME/af_database:/root/public_databases" \
    --gpus "device=0" \
    sky1ove/alphafold3 \
    python run_alphafold.py \
    --json_path=/root/af_input/sdf/proteinA.json \
    --output_dir=/root/af_output \
    --model_dir=/root/models

Prepare protein-smiles files

Read output json

protein_json = read_json('../af3/data/seq_only_data.json')

str(protein_json)[:1000]

'{\'dialect\': \'alphafold3\', \'version\': 2, \'name\': \'PDCD1_seq_only\', \'sequences\': [{\'protein\': {\'id\': \'A\', \'sequence\': \'LDSPDRPWNPPTFSPALLVVTEGDNATFTCSFSNTSESFVLNWYRMSPSNQTDKLAAFPEDRSQPGQDCRFRVTQLPNGRDFHMSVVRARRNDSGTYLCGAISLAPKAQIKESLRAELRVTERRAEVPTAHPSPSPRPAGQFQTLV\', \'modifications\': [], \'unpairedMsa\': ">query\\nLDSPDRPWNPPTFSPALLVVTEGDNATFTCSFSNTSESFVLNWYRMSPSNQTDKLAAFPEDRSQPGQDCRFRVTQLPNGRDFHMSVVRARRNDSGTYLCGAISLAPKAQIKESLRAELRVTERRAEVPTAHPSPSPRPAGQFQTLV\\n>UniRef90_UPI0009801507/25-167 [subseq from] Programmed cell death protein 1 n=10 Tax=Homo sapiens TaxID=9606 RepID=UPI0009801507\\nLDSPDRPWNPPTFSPALLVVTEGDNATFTCSFSNTSESFVLNWYRMSPSNQTDKLAAFPEDRSQPGQDCRFRVTQLPNGRDFHMSVVRARRNDSGTYLCGAISLAPKAQIKESLRAELRVTERRAEVPTAHPSPSPRPAGQFQ---\\n>UniRef90_A0A5F7ZCX7/24-168 [subseq from] Programmed cell death 1 n=1 Tax=Macaca mulatta TaxID=9544 RepID=A0A5F7ZCX7_MACMU\\n-ESPDRPWNPPTFSPALLLVTEGDNATFTCSFSNASESFVLNWYRMSPSNQTDKLAAFPEDRSQPGRDCRFRVTQLPNGRDFHMSVVRARRNDSGTYLCGAISLAPKAQIKESLRAELRVTERRAEVPTAHPSPSPRP'

Single protein-smile pair

out = get_protein_smiles_json('smi_name','CCC',protein_json,'data/protein_smi.json')

Multiple protein-smile pairs in a df

df = pd.DataFrame({'idx':['a','b'],'smi':['CCC','OCO']})
df

	idx	smi
0	a	CCC
1	b	OCO

for idx, smi in df.values:
    _ = get_protein_smiles_json(idx,smi,protein_json,f'af_input/{project_name}/{idx}.json',seeds=[1,2,3])

This will generate many json files in the directory

Split file into multiple subfolder for multi-GPUs

split_nfolder(f'af_input/{project_name}')

Distributed 2 files into 4 folders.

Docker

docker pull sky1ove/alphafold3

for i in range(4):
    docker_multi_infer(input_dir=f"af_input/{project_name}/folder_{i}",
                               output_dir=f"af_output/{project_name}",
                               gpus=i)
# norun_data_pipeline means skip template search as we already did in the first step

docker run --rm \
    --volume "$HOME/af_input:/root/af_input" \
    --volume "$HOME/af_output/sdf:/root/af_output" \
    --volume "$HOME/af_model:/root/models" \
    --volume "$HOME/af_database:/root/public_databases" \
    --volume "$HOME/af_cache:/root/cache" \
    --gpus "device=0" \
    sky1ove/alphafold3 \
    python run_alphafold.py \
    --input_dir=/root/af_input/sdf/folder_0 \
    --output_dir=/root/af_output \
    --model_dir=/root/models \
    --jax_compilation_cache_dir=/root/cache \
    --norun_data_pipeline
docker run --rm \
    --volume "$HOME/af_input:/root/af_input" \
    --volume "$HOME/af_output/sdf:/root/af_output" \
    --volume "$HOME/af_model:/root/models" \
    --volume "$HOME/af_database:/root/public_databases" \
    --volume "$HOME/af_cache:/root/cache" \
    --gpus "device=1" \
    sky1ove/alphafold3 \
    python run_alphafold.py \
    --input_dir=/root/af_input/sdf/folder_1 \
    --output_dir=/root/af_output \
    --model_dir=/root/models \
    --jax_compilation_cache_dir=/root/cache \
    --norun_data_pipeline
docker run --rm \
    --volume "$HOME/af_input:/root/af_input" \
    --volume "$HOME/af_output/sdf:/root/af_output" \
    --volume "$HOME/af_model:/root/models" \
    --volume "$HOME/af_database:/root/public_databases" \
    --volume "$HOME/af_cache:/root/cache" \
    --gpus "device=2" \
    sky1ove/alphafold3 \
    python run_alphafold.py \
    --input_dir=/root/af_input/sdf/folder_2 \
    --output_dir=/root/af_output \
    --model_dir=/root/models \
    --jax_compilation_cache_dir=/root/cache \
    --norun_data_pipeline
docker run --rm \
    --volume "$HOME/af_input:/root/af_input" \
    --volume "$HOME/af_output/sdf:/root/af_output" \
    --volume "$HOME/af_model:/root/models" \
    --volume "$HOME/af_database:/root/public_databases" \
    --volume "$HOME/af_cache:/root/cache" \
    --gpus "device=3" \
    sky1ove/alphafold3 \
    python run_alphafold.py \
    --input_dir=/root/af_input/sdf/folder_3 \
    --output_dir=/root/af_output \
    --model_dir=/root/models \
    --jax_compilation_cache_dir=/root/cache \
    --norun_data_pipeline