pdbbind.py

from mpi4py import MPI
from mpi4py.futures import MPICommExecutor

import warnings
from Bio.PDB import PDBParser, PPBuilder, CaPPBuilder
from Bio.PDB.NeighborSearch import NeighborSearch
from Bio.PDB.Selection import unfold_entities

import numpy as np
import dask.array as da

from rdkit import Chem

from spyrmsd import molecule
from spyrmsd import graph
import networkx as nx

import os
import re
import sys

# all punctuation
punctuation_regex  = r"""(\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|[0-9])"""

# tokenization regex (Schwaller)
molecule_regex = r"""(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|[0-9])"""

max_seq = 2046 # = 2048 - 2 (accounting for [CLS] and [SEP])
max_smiles = 510 # = 512 - 2
chunk_size = '1G'

def rot_from_two_vecs(e0_unnormalized, e1_unnormalized):
    """Create rotation matrices from unnormalized vectors for the x and y-axes.
    This creates a rotation matrix from two vectors using Gram-Schmidt
    orthogonalization.
    Args:
    e0_unnormalized: vectors lying along x-axis of resulting rotation
    e1_unnormalized: vectors lying in xy-plane of resulting rotation
    Returns:
    Rotations resulting from Gram-Schmidt procedure.
    """
    # Normalize the unit vector for the x-axis, e0.
    e0 = e0_unnormalized / np.linalg.norm(e0_unnormalized)

    # make e1 perpendicular to e0.
    c = np.dot(e1_unnormalized, e0)
    e1 = e1_unnormalized - c * e0
    e1 = e1 / np.linalg.norm(e1)

    # Compute e2 as cross product of e0 and e1.
    e2 = np.cross(e0, e1)

    # local to space frame
    return np.stack([e0,e1,e2]).T

def get_local_frames(mol):
    # get the two nearest neighbors of every atom on the molecular graph
    # ties are broken using canonical ordering
    g = molecule.Molecule.from_rdkit(mol).to_graph()

    R = []
    for node in g:
        length = nx.single_source_shortest_path_length(g, node)

        neighbor_a = [n for n,l in length.items() if l==1][0]

        try:
            neighbor_b = [n for n,l in length.items() if l==1][1]
        except:
            # get next nearest neighbor
            neighbor_b = [n for n,l in length.items() if l==2][0]

        xyz = np.array(mol.GetConformer().GetAtomPosition(node))
        xyz_a = np.array(mol.GetConformer().GetAtomPosition(neighbor_a))
        xyz_b = np.array(mol.GetConformer().GetAtomPosition(neighbor_b))

        R.append(rot_from_two_vecs(xyz_a-xyz, xyz_b-xyz))

    return R

def parse_complex(fn):
    try:
        name = os.path.basename(fn)

        # parse protein sequence and coordinates
        parser = PDBParser()
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            structure = parser.get_structure('protein',fn+'/'+name+'_protein.pdb')

        res_frames = []

        # extract sequence, Calpha positions and local coordinate frames using the AF2 convention
        ppb = CaPPBuilder()
        seq = []
        xyz_receptor = []
        R_receptor = []
        for pp in ppb.build_peptides(structure):
            seq.append(str(pp.get_sequence()))
            xyz_receptor += [tuple(a.get_vector()) for a in pp.get_ca_list()]

            for res in pp:
                N = np.array(tuple(res['N'].get_vector()))
                C = np.array(tuple(res['C'].get_vector()))
                CA = np.array(tuple(res['CA'].get_vector()))

                R_receptor.append(rot_from_two_vecs(N-CA,C-CA).flatten().tolist())

        seq = ''.join(seq)

        # parse ligand, convert to SMILES and map atoms
        suppl = Chem.SDMolSupplier(fn+'/'+name+'_ligand.sdf')
        mol = next(suppl)

        # bring molecule atoms in canonical order (to determine local frames uniquely)
        m_neworder = tuple(zip(*sorted([(j, i) for i, j in enumerate(Chem.CanonicalRankAtoms(mol))])))[1]
        mol = Chem.RenumberAtoms(mol, m_neworder)

        # position of atoms in SMILES (not counting punctuation)
        smi = Chem.MolToSmiles(mol)
        atom_order = [int(s) for s in list(filter(None,re.sub(r'[\[\]]','',mol.GetProp("_smilesAtomOutputOrder")).split(',')))]

        # tokenize the SMILES
        tokens = list(filter(None, re.split(molecule_regex, smi)))

        # remove punctuation
        masked_tokens = [re.sub(punctuation_regex,'',s) for s in tokens]

        k = 0
        token_pos = []
        token_rot = []

        frames = get_local_frames(mol)

        for i,token in enumerate(masked_tokens):
            if token != '':
                token_pos.append(tuple(mol.GetConformer().GetAtomPosition(atom_order[k])))
                token_rot.append(frames[atom_order[k]].flatten().tolist())
                k += 1
            else:
                token_pos.append((np.nan, np.nan, np.nan))
                token_rot.append(np.eye(3).flatten().tolist())

        return name, seq, smi, xyz_receptor, token_pos, token_rot, R_receptor

    except Exception as e:
        print(e)
        return None


if __name__ == '__main__':
    import glob

    filenames = glob.glob('data/pdbbind/v2020-other-PL/*')
    filenames.extend(glob.glob('data/pdbbind/refined-set/*'))
    filenames = sorted(filenames)
    comm = MPI.COMM_WORLD
    with MPICommExecutor(comm, root=0) as executor:
        if executor is not None:
            result = executor.map(parse_complex, filenames, chunksize=32)
            result = list(result)
            names = [r[0] for r in result if r is not None]
            seqs = [r[1] for r in result if r is not None]
            all_smiles = [r[2] for r in result if r is not None]
            all_xyz_receptor = [r[3] for r in result if r is not None]
            all_xyz_ligand = [r[4] for r in result if r is not None]
            all_rot_ligand = [r[5] for r in result if r is not None]
            all_rot_receptor = [r[6] for r in result if r is not None]

            import pandas as pd
            df = pd.DataFrame({'name': names, 'seq': seqs,
                               'smiles': all_smiles,
                               'receptor_xyz': all_xyz_receptor,
                               'ligand_xyz': all_xyz_ligand,
                               'ligand_rot': all_rot_ligand,
                               'receptor_rot': all_rot_receptor})
            df.to_parquet('data/pdbbind.parquet',index=False)