deepscreen/models/predictors/bacpi.py

from collections import defaultdict

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from rdkit import Chem
from rdkit.Chem import AllChem


class BACPI(nn.Module):
    def __init__(
            self,
            n_atom,
            n_amino,
            comp_dim,
            prot_dim,
            gat_dim,
            num_head,
            dropout,
            alpha,
            window,
            layer_cnn,
            latent_dim,
    ):
        super().__init__()
        self.embedding_layer_atom = nn.Embedding(n_atom + 1, comp_dim)
        self.embedding_layer_amino = nn.Embedding(n_amino + 1, prot_dim)

        self.dropout = dropout
        self.alpha = alpha
        self.layer_cnn = layer_cnn

        self.gat_layers = [GATLayer(comp_dim, gat_dim, dropout=dropout, alpha=alpha, concat=True)
                           for _ in range(num_head)]
        for i, layer in enumerate(self.gat_layers):
            self.add_module('gat_layer_{}'.format(i), layer)
        self.gat_out = GATLayer(gat_dim * num_head, comp_dim, dropout=dropout, alpha=alpha, concat=False)
        self.W_comp = nn.Linear(comp_dim, latent_dim)

        self.conv_layers = nn.ModuleList([nn.Conv2d(in_channels=1, out_channels=1, kernel_size=2 * window + 1,
                                                    stride=1, padding=window) for _ in range(layer_cnn)])
        self.W_prot = nn.Linear(prot_dim, latent_dim)

        self.fp0 = nn.Parameter(torch.empty(size=(1024, latent_dim)))
        nn.init.xavier_uniform_(self.fp0, gain=1.414)
        self.fp1 = nn.Parameter(torch.empty(size=(latent_dim, latent_dim)))
        nn.init.xavier_uniform_(self.fp1, gain=1.414)

        self.bidat_num = 4

        self.U = nn.ParameterList([
            nn.Parameter(torch.empty(size=(latent_dim, latent_dim))) for _ in range(self.bidat_num)
        ])
        for i in range(self.bidat_num):
            nn.init.xavier_uniform_(self.U[i], gain=1.414)

        self.transform_c2p = nn.ModuleList([nn.Linear(latent_dim, latent_dim) for _ in range(self.bidat_num)])
        self.transform_p2c = nn.ModuleList([nn.Linear(latent_dim, latent_dim) for _ in range(self.bidat_num)])

        self.bihidden_c = nn.ModuleList([nn.Linear(latent_dim, latent_dim) for _ in range(self.bidat_num)])
        self.bihidden_p = nn.ModuleList([nn.Linear(latent_dim, latent_dim) for _ in range(self.bidat_num)])
        self.biatt_c = nn.ModuleList([nn.Linear(latent_dim * 2, 1) for _ in range(self.bidat_num)])
        self.biatt_p = nn.ModuleList([nn.Linear(latent_dim * 2, 1) for _ in range(self.bidat_num)])

        self.comb_c = nn.Linear(latent_dim * self.bidat_num, latent_dim)
        self.comb_p = nn.Linear(latent_dim * self.bidat_num, latent_dim)

    def comp_gat(self, atoms, atoms_mask, adj):
        atoms_vector = self.embedding_layer_atom(atoms)
        atoms_multi_head = torch.cat([gat(atoms_vector, adj) for gat in self.gat_layers], dim=2)
        atoms_vector = F.elu(self.gat_out(atoms_multi_head, adj))
        atoms_vector = F.leaky_relu(self.W_comp(atoms_vector), self.alpha)
        return atoms_vector

    def prot_cnn(self, amino, amino_mask):
        amino_vector = self.embedding_layer_amino(amino)
        amino_vector = torch.unsqueeze(amino_vector, 1)
        for i in range(self.layer_cnn):
            amino_vector = F.leaky_relu(self.conv_layers[i](amino_vector), self.alpha)
        amino_vector = torch.squeeze(amino_vector, 1)
        amino_vector = F.leaky_relu(self.W_prot(amino_vector), self.alpha)
        return amino_vector

    def mask_softmax(self, a, mask, dim=-1):
        a_max = torch.max(a, dim, keepdim=True)[0]
        a_exp = torch.exp(a - a_max)
        a_exp = a_exp * mask
        a_softmax = a_exp / (torch.sum(a_exp, dim, keepdim=True) + 1e-6)
        return a_softmax

    def bidirectional_attention_prediction(self, atoms_vector, atoms_mask, fps, amino_vector, amino_mask):
        b = atoms_vector.shape[0]
        for i in range(self.bidat_num):
            A = torch.tanh(torch.matmul(torch.matmul(atoms_vector, self.U[i]), amino_vector.transpose(1, 2)))
            A = A * torch.matmul(atoms_mask.view(b, -1, 1).float(), amino_mask.view(b, 1, -1).float())

            atoms_trans = torch.matmul(A, torch.tanh(self.transform_p2c[i](amino_vector)))
            amino_trans = torch.matmul(A.transpose(1, 2), torch.tanh(self.transform_c2p[i](atoms_vector)))

            atoms_tmp = torch.cat([torch.tanh(self.bihidden_c[i](atoms_vector)), atoms_trans], dim=2)
            amino_tmp = torch.cat([torch.tanh(self.bihidden_p[i](amino_vector)), amino_trans], dim=2)

            atoms_att = self.mask_softmax(self.biatt_c[i](atoms_tmp).view(b, -1), atoms_mask.view(b, -1).float())
            amino_att = self.mask_softmax(self.biatt_p[i](amino_tmp).view(b, -1), amino_mask.view(b, -1).float())

            cf = torch.sum(atoms_vector * atoms_att.view(b, -1, 1), dim=1)
            pf = torch.sum(amino_vector * amino_att.view(b, -1, 1), dim=1)

            if i == 0:
                cat_cf = cf
                cat_pf = pf
            else:
                cat_cf = torch.cat([cat_cf.view(b, -1), cf.view(b, -1)], dim=1)
                cat_pf = torch.cat([cat_pf.view(b, -1), pf.view(b, -1)], dim=1)

        cf_final = torch.cat([self.comb_c(cat_cf).view(b, -1), fps.view(b, -1)], dim=1)
        pf_final = self.comb_p(cat_pf)
        cf_pf = F.leaky_relu(
            torch.matmul(
                cf_final.view(b, -1, 1), pf_final.view(b, 1, -1)
            ).view(b, -1), 0.1
        )

        return cf_pf

    def forward(self, compound, protein):
        atom, adj, fp = compound

        atom, atom_lengths = atom
        adj, _ = adj
        fp, _ = fp
        amino, amino_lengths = protein

        atom_mask = torch.arange(atom.size(1), device=atom.device) >= atom_lengths.unsqueeze(1)
        amino_mask = torch.arange(amino.size(1), device=amino.device) >= amino_lengths.unsqueeze(1)

        atoms_vector = self.comp_gat(atom, atom_mask, adj)
        amino_vector = self.prot_cnn(amino, amino_mask)

        super_feature = F.leaky_relu(torch.matmul(fp.float(), self.fp0), 0.1)
        super_feature = F.leaky_relu(torch.matmul(super_feature, self.fp1), 0.1)

        prediction = self.bidirectional_attention_prediction(
            atoms_vector, atom_mask, super_feature, amino_vector, amino_mask)

        return prediction


class GATLayer(nn.Module):
    def __init__(self, in_features, out_features, dropout=0.5, alpha=0.2, concat=True):
        super().__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.dropout = dropout
        self.alpha = alpha
        self.concat = concat

        self.W = nn.Parameter(torch.empty(size=(in_features, out_features)))
        nn.init.xavier_uniform_(self.W.data, gain=1.414)
        self.a = nn.Parameter(torch.empty(size=(2 * out_features, 1)))
        nn.init.xavier_uniform_(self.a.data, gain=1.414)

    def forward(self, h, adj):
        Wh = torch.matmul(h, self.W)
        a_input = self._prepare_attentional_mechanism_input(Wh)
        e = F.leaky_relu(torch.matmul(a_input, self.a).squeeze(3), self.alpha)

        zero_vec = -9e15 * torch.ones_like(e)
        attention = torch.where(adj > 0, e, zero_vec)
        attention = F.softmax(attention, dim=2)
        # attention = F.dropout(attention, self.dropout, training=self.training)
        h_prime = torch.bmm(attention, Wh)

        return F.elu(h_prime) if self.concat else h_prime

    def _prepare_attentional_mechanism_input(self, Wh):
        b = Wh.size()[0]
        N = Wh.size()[1]

        Wh_repeated_in_chunks = Wh.repeat_interleave(N, dim=1)
        Wh_repeated_alternating = Wh.repeat_interleave(N, dim=0).view(b, N * N, self.out_features)
        all_combinations_matrix = torch.cat([Wh_repeated_in_chunks, Wh_repeated_alternating], dim=2)

        return all_combinations_matrix.view(b, N, N, 2 * self.out_features)


atom_dict = defaultdict(lambda: len(atom_dict))
bond_dict = defaultdict(lambda: len(bond_dict))
fingerprint_dict = defaultdict(lambda: len(fingerprint_dict))
edge_dict = defaultdict(lambda: len(edge_dict))
word_dict = defaultdict(lambda: len(word_dict))


def create_atoms(mol):
    atoms = [a.GetSymbol() for a in mol.GetAtoms()]
    for a in mol.GetAromaticAtoms():
        i = a.GetIdx()
        atoms[i] = (atoms[i], 'aromatic')
    atoms = [atom_dict[a] for a in atoms]
    return np.array(atoms)


def create_ijbonddict(mol):
    i_jbond_dict = defaultdict(lambda: [])
    for b in mol.GetBonds():
        i, j = b.GetBeginAtomIdx(), b.GetEndAtomIdx()
        bond = bond_dict[str(b.GetBondType())]
        i_jbond_dict[i].append((j, bond))
        i_jbond_dict[j].append((i, bond))

    atoms_set = set(range(mol.GetNumAtoms()))
    isolate_atoms = atoms_set - set(i_jbond_dict.keys())
    bond = bond_dict['nan']
    for a in isolate_atoms:
        i_jbond_dict[a].append((a, bond))

    return i_jbond_dict


def atom_features(atoms, i_jbond_dict, radius):
    if (len(atoms) == 1) or (radius == 0):
        fingerprints = [fingerprint_dict[a] for a in atoms]
    else:
        nodes = atoms
        i_jedge_dict = i_jbond_dict
        for _ in range(radius):
            fingerprints = []
            for i, j_edge in i_jedge_dict.items():
                neighbors = [(nodes[j], edge) for j, edge in j_edge]
                fingerprint = (nodes[i], tuple(sorted(neighbors)))
                fingerprints.append(fingerprint_dict[fingerprint])

            nodes = fingerprints
            _i_jedge_dict = defaultdict(lambda: [])
            for i, j_edge in i_jedge_dict.items():
                for j, edge in j_edge:
                    both_side = tuple(sorted((nodes[i], nodes[j])))
                    edge = edge_dict[(both_side, edge)]
                    _i_jedge_dict[i].append((j, edge))
            i_jedge_dict = _i_jedge_dict

    return np.array(fingerprints)


def create_adjacency(mol):
    adjacency = Chem.GetAdjacencyMatrix(mol)
    adjacency = np.array(adjacency)
    adjacency += np.eye(adjacency.shape[0], dtype=int)
    return adjacency


def get_fingerprints(mol):
    fp = AllChem.GetMorganFingerprintAsBitVect(mol, 2, nBits=1024, useChirality=True)
    return np.array(fp)


def split_sequence(sequence, ngram=3):
    sequence = '-' + sequence + '='
    words = [word_dict[sequence[i:i + ngram]]
             for i in range(len(sequence) - ngram + 1)]
    return np.array(words)


def drug_featurizer(smiles, radius=2):
    from deepscreen.utils import get_logger
    log = get_logger(__name__)
    try:
        mol = Chem.MolFromSmiles(smiles)
        if not mol:
            return None
        mol = Chem.AddHs(mol)
        atoms = create_atoms(mol)
        i_jbond_dict = create_ijbonddict(mol)

        compound = atom_features(atoms, i_jbond_dict, radius)
        adjacency = create_adjacency(mol)
        fp = get_fingerprints(mol)

        return compound, adjacency, fp

    except Exception as e:
        log.warning(f"Failed to featurize SMILES ({smiles}) to graph due to {str(e)}")
        return None