DrugGEN / smiles_cor.py
mgyigit's picture
Upload 7 files
4e04e76 verified
raw
history blame
48 kB
import torch
import os
import pandas as pd
import random
from chembl_structure_pipeline import standardizer
from rdkit.Chem import MolStandardize
from rdkit import Chem
import time
import torch
import torch.nn as nn
from torchtext.data import TabularDataset, Field, BucketIterator, Iterator
import random
import os
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
import random
from torch import optim
import numpy as np
import itertools
import time
import statistics
from rdkit.Chem import GraphDescriptors, Lipinski, AllChem
from rdkit.Chem.rdSLNParse import MolFromSLN
from rdkit.Chem.rdmolfiles import MolFromSmiles
import torch
import torch.nn as nn
import torch.optim as optim
import pandas as pd
import numpy as np
from rdkit import rdBase, Chem
import re
from rdkit import RDLogger
RDLogger.DisableLog('rdApp.*')
SEED = 42
random.seed(SEED)
torch.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
##################################################################################################
##################################################################################################
# #
#  THIS SCRIPT IS DIRECTLY ADAPTED FROM https://github.com/LindeSchoenmaker/SMILES-corrector #
# #
##################################################################################################
##################################################################################################
def is_smiles(array,
TRG,
reverse: bool,
return_output=False,
src=None,
src_field=None):
"""Turns predicted tokens within batch into smiles and evaluates their validity
Arguments:
array: Tensor with most probable token for each location for each sequence in batch
[trg len, batch size]
TRG: target field for getting tokens from vocab
reverse (bool): True if the target sequence is reversed
return_output (bool): True if output sequences and their validity should be saved
Returns:
df: dataframe with correct and incorrect sequences
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
smiless: list of the predicted smiles
"""
trg_field = TRG
valids = []
smiless = []
if return_output:
df = pd.DataFrame()
else:
df = None
batch_size = array.size(1)
# check if the first token should be removed, first token is zero because
# outputs initaliazed to all be zeros
if int((array[0, 0]).tolist()) == 0:
start = 1
else:
start = 0
# for each sequence in the batch
for i in range(0, batch_size):
# turns sequence from tensor to list skipps first row as this is not
# filled in in forward
sequence = (array[start:, i]).tolist()
# goes from embedded to tokens
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
# print(trg_tokens)
# takes all tokens untill eos token, model would be faster if did this
# one step earlier, but then changes in vocab order would disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
if reverse:
rev_tokens = rev_tokens[::-1]
smiles = "".join(rev_tokens)
# determine how many valid smiles are made
valid = True if MolFromSmiles(smiles) else False
valids.append(valid)
smiless.append(smiles)
if return_output:
if valid:
df.loc[i, "CORRECT"] = smiles
else:
df.loc[i, "INCORRECT"] = smiles
# add the original drugex outputs to the _de dataframe
if return_output and src is not None:
for i in range(0, batch_size):
# turns sequence from tensor to list skipps first row as this is
# <sos> for src
sequence = (src[1:, i]).tolist()
# goes from embedded to tokens
src_tokens = [src_field.vocab.itos[int(t)] for t in sequence]
# takes all tokens untill eos token, model would be faster if did
# this one step earlier, but then changes in vocab order would
# disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", src_tokens))
smiles = "".join(rev_tokens)
df.loc[i, "ORIGINAL"] = smiles
return df, valids, smiless
def is_unchanged(array,
TRG,
reverse: bool,
return_output=False,
src=None,
src_field=None):
"""Checks is output is different from input
Arguments:
array: Tensor with most probable token for each location for each sequence in batch
[trg len, batch size]
TRG: target field for getting tokens from vocab
reverse (bool): True if the target sequence is reversed
return_output (bool): True if output sequences and their validity should be saved
Returns:
df: dataframe with correct and incorrect sequences
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
smiless: list of the predicted smiles
"""
trg_field = TRG
sources = []
batch_size = array.size(1)
unchanged = 0
# check if the first token should be removed, first token is zero because
# outputs initaliazed to all be zeros
if int((array[0, 0]).tolist()) == 0:
start = 1
else:
start = 0
for i in range(0, batch_size):
# turns sequence from tensor to list skipps first row as this is <sos>
# for src
sequence = (src[1:, i]).tolist()
# goes from embedded to tokens
src_tokens = [src_field.vocab.itos[int(t)] for t in sequence]
# takes all tokens untill eos token, model would be faster if did this
# one step earlier, but then changes in vocab order would disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", src_tokens))
smiles = "".join(rev_tokens)
sources.append(smiles)
# for each sequence in the batch
for i in range(0, batch_size):
# turns sequence from tensor to list skipps first row as this is not
# filled in in forward
sequence = (array[start:, i]).tolist()
# goes from embedded to tokens
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
# print(trg_tokens)
# takes all tokens untill eos token, model would be faster if did this
# one step earlier, but then changes in vocab order would disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
if reverse:
rev_tokens = rev_tokens[::-1]
smiles = "".join(rev_tokens)
# determine how many valid smiles are made
valid = True if MolFromSmiles(smiles) else False
if not valid:
if smiles == sources[i]:
unchanged += 1
return unchanged
def molecule_reconstruction(array, TRG, reverse: bool, outputs):
"""Turns target tokens within batch into smiles and compares them to predicted output smiles
Arguments:
array: Tensor with target's token for each location for each sequence in batch
[trg len, batch size]
TRG: target field for getting tokens from vocab
reverse (bool): True if the target sequence is reversed
outputs: list of predicted SMILES sequences
Returns:
matches(int): number of total right molecules
"""
trg_field = TRG
matches = 0
targets = []
batch_size = array.size(1)
# for each sequence in the batch
for i in range(0, batch_size):
# turns sequence from tensor to list skipps first row as this is not
# filled in in forward
sequence = (array[1:, i]).tolist()
# goes from embedded to tokens
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
# takes all tokens untill eos token, model would be faster if did this
# one step earlier, but then changes in vocab order would disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
if reverse:
rev_tokens = rev_tokens[::-1]
smiles = "".join(rev_tokens)
targets.append(smiles)
for i in range(0, batch_size):
m = MolFromSmiles(targets[i])
p = MolFromSmiles(outputs[i])
if p is not None:
if m.HasSubstructMatch(p) and p.HasSubstructMatch(m):
matches += 1
return matches
def complexity_whitlock(mol: Chem.Mol, includeAllDescs=False):
"""
Complexity as defined in DOI:10.1021/jo9814546
S: complexity = 4*#rings + 2*#unsat + #hetatm + 2*#chiral
Other descriptors:
H: size = #bonds (Hydrogen atoms included)
G: S + H
Ratio: S / H
"""
mol_ = Chem.Mol(mol)
nrings = Lipinski.RingCount(mol_) - Lipinski.NumAromaticRings(mol_)
Chem.rdmolops.SetAromaticity(mol_)
unsat = sum(1 for bond in mol_.GetBonds()
if bond.GetBondTypeAsDouble() == 2)
hetatm = len(mol_.GetSubstructMatches(Chem.MolFromSmarts("[!#6]")))
AllChem.EmbedMolecule(mol_)
Chem.AssignAtomChiralTagsFromStructure(mol_)
chiral = len(Chem.FindMolChiralCenters(mol_))
S = 4 * nrings + 2 * unsat + hetatm + 2 * chiral
if not includeAllDescs:
return S
Chem.rdmolops.Kekulize(mol_)
mol_ = Chem.AddHs(mol_)
H = sum(bond.GetBondTypeAsDouble() for bond in mol_.GetBonds())
G = S + H
R = S / H
return {"WhitlockS": S, "WhitlockH": H, "WhitlockG": G, "WhitlockRatio": R}
def complexity_baronechanon(mol: Chem.Mol):
"""
Complexity as defined in DOI:10.1021/ci000145p
"""
mol_ = Chem.Mol(mol)
Chem.Kekulize(mol_)
Chem.RemoveStereochemistry(mol_)
mol_ = Chem.RemoveHs(mol_, updateExplicitCount=True)
degree, counts = 0, 0
for atom in mol_.GetAtoms():
degree += 3 * 2**(atom.GetExplicitValence() - atom.GetNumExplicitHs() -
1)
counts += 3 if atom.GetSymbol() == "C" else 6
ringterm = sum(map(lambda x: 6 * len(x), mol_.GetRingInfo().AtomRings()))
return degree + counts + ringterm
def calc_complexity(array,
TRG,
reverse,
valids,
complexity_function=GraphDescriptors.BertzCT):
"""Calculates the complexity of inputs that are not correct.
Arguments:
array: Tensor with target's token for each location for each sequence in batch
[trg len, batch size]
TRG: target field for getting tokens from vocab
reverse (bool): True if the target sequence is reversed
valids: list with booleans that show if prediction was a valid SMILES (True) or invalid one (False)
complexity_function: the type of complexity measure that will be used
GraphDescriptors.BertzCT
complexity_whitlock
complexity_baronechanon
Returns:
matches(int): mean of complexity values
"""
trg_field = TRG
sources = []
complexities = []
loc = torch.BoolTensor(valids)
# only keeps rows in batch size dimension where valid is false
array = array[:, loc == False]
# should check if this still works
# array = torch.transpose(array, 0, 1)
array_size = array.size(1)
for i in range(0, array_size):
# turns sequence from tensor to list skipps first row as this is not
# filled in in forward
sequence = (array[1:, i]).tolist()
# goes from embedded to tokens
trg_tokens = [trg_field.vocab.itos[int(t)] for t in sequence]
# takes all tokens untill eos token, model would be faster if did this
# one step earlier, but then changes in vocab order would disrupt.
rev_tokens = list(
itertools.takewhile(lambda x: x != "<eos>", trg_tokens))
if reverse:
rev_tokens = rev_tokens[::-1]
smiles = "".join(rev_tokens)
sources.append(smiles)
for source in sources:
try:
m = MolFromSmiles(source)
except BaseException:
m = MolFromSLN(source)
complexities.append(complexity_function(m))
if len(complexities) > 0:
mean = statistics.mean(complexities)
else:
mean = 0
return mean
def epoch_time(start_time, end_time):
elapsed_time = end_time - start_time
elapsed_mins = int(elapsed_time / 60)
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
return elapsed_mins, elapsed_secs
class Convo:
"""Class for training and evaluating transformer and convolutional neural network
Methods
-------
train_model()
train model for initialized number of epochs
evaluate(return_output)
use model with validation loader (& optionally drugex loader) to get test loss & other metrics
translate(loader)
translate inputs from loader (different from evaluate in that no target sequence is used)
"""
def train_model(self):
optimizer = optim.Adam(self.parameters(), lr=self.lr)
log = open(f"{self.out}.log", "a")
best_error = np.inf
for epoch in range(self.epochs):
self.train()
start_time = time.time()
loss_train = 0
for i, batch in enumerate(self.loader_train):
optimizer.zero_grad()
# changed src,trg call to match with bentrevett
# src, trg = batch['src'], batch['trg']
trg = batch.trg
src = batch.src
output, attention = self(src, trg[:, :-1])
# feed the source and target into def forward to get the output
# Xuhan uses forward for this, with istrain = true
output_dim = output.shape[-1]
# changed
output = output.contiguous().view(-1, output_dim)
trg = trg[:, 1:].contiguous().view(-1)
# output = output[:,:,0]#.view(-1)
# output = output[1:].view(-1, output.shape[-1])
# trg = trg[1:].view(-1)
loss = nn.CrossEntropyLoss(
ignore_index=self.TRG.vocab.stoi[self.TRG.pad_token])
a, b = output.view(-1), trg.to(self.device).view(-1)
# changed
# loss = loss(output.view(0), trg.view(0).to(device))
loss = loss(output, trg)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.parameters(), self.clip)
optimizer.step()
loss_train += loss.item()
# turned off for now, as not using voc so won't work, output is a tensor
# output = [(trg len - 1) * batch size, output dim]
# smiles, valid = is_valid_smiles(output, reversed)
# if valid:
# valids += 1
# smiless.append(smiles)
# added .dataset becaue len(iterator) gives len(self.dataset) /
# self.batch_size)
loss_train /= len(self.loader_train)
info = f"Epoch: {epoch+1:02} step: {i} loss_train: {loss_train:.4g}"
# model is used to generate trg based on src from the validation set to assess performance
# similar to Xuhan, although he doesn't use the if loop
if self.loader_valid is not None:
return_output = False
if epoch + 1 == self.epochs:
return_output = True
(
valids,
loss_valid,
valids_de,
df_output,
df_output_de,
right_molecules,
complexity,
unchanged,
unchanged_de,
) = self.evaluate(return_output)
reconstruction_error = 1 - right_molecules / len(
self.loader_valid.dataset)
error = 1 - valids / len(self.loader_valid.dataset)
complexity = complexity / len(self.loader_valid)
unchan = unchanged / (len(self.loader_valid.dataset) - valids)
info += f" loss_valid: {loss_valid:.4g} error_rate: {error:.4g} molecule_reconstruction_error_rate: {reconstruction_error:.4g} unchanged: {unchan:.4g} invalid_target_complexity: {complexity:.4g}"
if self.loader_drugex is not None:
error_de = 1 - valids_de / len(self.loader_drugex.dataset)
unchan_de = unchanged_de / (
len(self.loader_drugex.dataset) - valids_de)
info += f" error_rate_drugex: {error_de:.4g} unchanged_drugex: {unchan_de:.4g}"
if reconstruction_error < best_error:
torch.save(self.state_dict(), f"{self.out}.pkg")
best_error = reconstruction_error
last_save = epoch
else:
if epoch - last_save >= 10 and best_error != 1:
torch.save(self.state_dict(), f"{self.out}_last.pkg")
(
valids,
loss_valid,
valids_de,
df_output,
df_output_de,
right_molecules,
complexity,
unchanged,
unchanged_de,
) = self.evaluate(True)
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(
start_time, end_time)
info += f" Time: {epoch_mins}m {epoch_secs}s"
break
elif error < best_error:
torch.save(self.state_dict(), f"{self.out}.pkg")
best_error = error
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
info += f" Time: {epoch_mins}m {epoch_secs}s"
torch.save(self.state_dict(), f"{self.out}_last.pkg")
log.close()
self.load_state_dict(torch.load(f"{self.out}.pkg"))
df_output.to_csv(f"{self.out}.csv", index=False)
df_output_de.to_csv(f"{self.out}_de.csv", index=False)
def evaluate(self, return_output):
self.eval()
test_loss = 0
df_output = pd.DataFrame()
df_output_de = pd.DataFrame()
valids = 0
valids_de = 0
unchanged = 0
unchanged_de = 0
right_molecules = 0
complexity = 0
with torch.no_grad():
for _, batch in enumerate(self.loader_valid):
trg = batch.trg
src = batch.src
output, attention = self.forward(src, trg[:, :-1])
pred_token = output.argmax(2)
array = torch.transpose(pred_token, 0, 1)
trg_trans = torch.transpose(trg, 0, 1)
output_dim = output.shape[-1]
output = output.contiguous().view(-1, output_dim)
trg = trg[:, 1:].contiguous().view(-1)
src_trans = torch.transpose(src, 0, 1)
df_batch, valid, smiless = is_smiles(
array, self.TRG, reverse=True, return_output=return_output)
unchanged += is_unchanged(
array,
self.TRG,
reverse=True,
return_output=return_output,
src=src_trans,
src_field=self.SRC,
)
matches = molecule_reconstruction(trg_trans,
self.TRG,
reverse=True,
outputs=smiless)
complexity += calc_complexity(trg_trans,
self.TRG,
reverse=True,
valids=valid)
if df_batch is not None:
df_output = pd.concat([df_output, df_batch],
ignore_index=True)
right_molecules += matches
valids += sum(valid)
# trg = trg[1:].view(-1)
# output, trg = output[1:].view(-1, output.shape[-1]), trg[1:].view(-1)
loss = nn.CrossEntropyLoss(
ignore_index=self.TRG.vocab.stoi[self.TRG.pad_token])
loss = loss(output, trg)
test_loss += loss.item()
if self.loader_drugex is not None:
for _, batch in enumerate(self.loader_drugex):
src = batch.src
output = self.translate_sentence(src, self.TRG,
self.device)
# checks the number of valid smiles
pred_token = output.argmax(2)
array = torch.transpose(pred_token, 0, 1)
src_trans = torch.transpose(src, 0, 1)
df_batch, valid, smiless = is_smiles(
array,
self.TRG,
reverse=True,
return_output=return_output,
src=src_trans,
src_field=self.SRC,
)
unchanged_de += is_unchanged(
array,
self.TRG,
reverse=True,
return_output=return_output,
src=src_trans,
src_field=self.SRC,
)
if df_batch is not None:
df_output_de = pd.concat([df_output_de, df_batch],
ignore_index=True)
valids_de += sum(valid)
return (
valids,
test_loss / len(self.loader_valid),
valids_de,
df_output,
df_output_de,
right_molecules,
complexity,
unchanged,
unchanged_de,
)
def translate(self, loader):
self.eval()
df_output_de = pd.DataFrame()
valids_de = 0
with torch.no_grad():
for _, batch in enumerate(loader):
src = batch.src
output = self.translate_sentence(src, self.TRG, self.device)
# checks the number of valid smiles
pred_token = output.argmax(2)
array = torch.transpose(pred_token, 0, 1)
src_trans = torch.transpose(src, 0, 1)
df_batch, valid, smiless = is_smiles(
array,
self.TRG,
reverse=True,
return_output=True,
src=src_trans,
src_field=self.SRC,
)
if df_batch is not None:
df_output_de = pd.concat([df_output_de, df_batch],
ignore_index=True)
valids_de += sum(valid)
return valids_de, df_output_de
class Encoder(nn.Module):
def __init__(self, input_dim, hid_dim, n_layers, n_heads, pf_dim, dropout,
max_length, device):
super().__init__()
self.device = device
self.tok_embedding = nn.Embedding(input_dim, hid_dim)
self.pos_embedding = nn.Embedding(max_length, hid_dim)
self.layers = nn.ModuleList([
EncoderLayer(hid_dim, n_heads, pf_dim, dropout, device)
for _ in range(n_layers)
])
self.dropout = nn.Dropout(dropout)
self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
def forward(self, src, src_mask):
# src = [batch size, src len]
# src_mask = [batch size, src len]
batch_size = src.shape[0]
src_len = src.shape[1]
pos = (torch.arange(0, src_len).unsqueeze(0).repeat(batch_size,
1).to(self.device))
# pos = [batch size, src len]
src = self.dropout((self.tok_embedding(src) * self.scale) +
self.pos_embedding(pos))
# src = [batch size, src len, hid dim]
for layer in self.layers:
src = layer(src, src_mask)
# src = [batch size, src len, hid dim]
return src
class EncoderLayer(nn.Module):
def __init__(self, hid_dim, n_heads, pf_dim, dropout, device):
super().__init__()
self.self_attn_layer_norm = nn.LayerNorm(hid_dim)
self.ff_layer_norm = nn.LayerNorm(hid_dim)
self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads,
dropout, device)
self.positionwise_feedforward = PositionwiseFeedforwardLayer(
hid_dim, pf_dim, dropout)
self.dropout = nn.Dropout(dropout)
def forward(self, src, src_mask):
# src = [batch size, src len, hid dim]
# src_mask = [batch size, src len]
# self attention
_src, _ = self.self_attention(src, src, src, src_mask)
# dropout, residual connection and layer norm
src = self.self_attn_layer_norm(src + self.dropout(_src))
# src = [batch size, src len, hid dim]
# positionwise feedforward
_src = self.positionwise_feedforward(src)
# dropout, residual and layer norm
src = self.ff_layer_norm(src + self.dropout(_src))
# src = [batch size, src len, hid dim]
return src
class MultiHeadAttentionLayer(nn.Module):
def __init__(self, hid_dim, n_heads, dropout, device):
super().__init__()
assert hid_dim % n_heads == 0
self.hid_dim = hid_dim
self.n_heads = n_heads
self.head_dim = hid_dim // n_heads
self.fc_q = nn.Linear(hid_dim, hid_dim)
self.fc_k = nn.Linear(hid_dim, hid_dim)
self.fc_v = nn.Linear(hid_dim, hid_dim)
self.fc_o = nn.Linear(hid_dim, hid_dim)
self.dropout = nn.Dropout(dropout)
self.scale = torch.sqrt(torch.FloatTensor([self.head_dim])).to(device)
def forward(self, query, key, value, mask=None):
batch_size = query.shape[0]
# query = [batch size, query len, hid dim]
# key = [batch size, key len, hid dim]
# value = [batch size, value len, hid dim]
Q = self.fc_q(query)
K = self.fc_k(key)
V = self.fc_v(value)
# Q = [batch size, query len, hid dim]
# K = [batch size, key len, hid dim]
# V = [batch size, value len, hid dim]
Q = Q.view(batch_size, -1, self.n_heads,
self.head_dim).permute(0, 2, 1, 3)
K = K.view(batch_size, -1, self.n_heads,
self.head_dim).permute(0, 2, 1, 3)
V = V.view(batch_size, -1, self.n_heads,
self.head_dim).permute(0, 2, 1, 3)
# Q = [batch size, n heads, query len, head dim]
# K = [batch size, n heads, key len, head dim]
# V = [batch size, n heads, value len, head dim]
energy = torch.matmul(Q, K.permute(0, 1, 3, 2)) / self.scale
# energy = [batch size, n heads, query len, key len]
if mask is not None:
energy = energy.masked_fill(mask == 0, -1e10)
attention = torch.softmax(energy, dim=-1)
# attention = [batch size, n heads, query len, key len]
x = torch.matmul(self.dropout(attention), V)
# x = [batch size, n heads, query len, head dim]
x = x.permute(0, 2, 1, 3).contiguous()
# x = [batch size, query len, n heads, head dim]
x = x.view(batch_size, -1, self.hid_dim)
# x = [batch size, query len, hid dim]
x = self.fc_o(x)
# x = [batch size, query len, hid dim]
return x, attention
class PositionwiseFeedforwardLayer(nn.Module):
def __init__(self, hid_dim, pf_dim, dropout):
super().__init__()
self.fc_1 = nn.Linear(hid_dim, pf_dim)
self.fc_2 = nn.Linear(pf_dim, hid_dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x):
# x = [batch size, seq len, hid dim]
x = self.dropout(torch.relu(self.fc_1(x)))
# x = [batch size, seq len, pf dim]
x = self.fc_2(x)
# x = [batch size, seq len, hid dim]
return x
class Decoder(nn.Module):
def __init__(
self,
output_dim,
hid_dim,
n_layers,
n_heads,
pf_dim,
dropout,
max_length,
device,
):
super().__init__()
self.device = device
self.tok_embedding = nn.Embedding(output_dim, hid_dim)
self.pos_embedding = nn.Embedding(max_length, hid_dim)
self.layers = nn.ModuleList([
DecoderLayer(hid_dim, n_heads, pf_dim, dropout, device)
for _ in range(n_layers)
])
self.fc_out = nn.Linear(hid_dim, output_dim)
self.dropout = nn.Dropout(dropout)
self.scale = torch.sqrt(torch.FloatTensor([hid_dim])).to(device)
def forward(self, trg, enc_src, trg_mask, src_mask):
# trg = [batch size, trg len]
# enc_src = [batch size, src len, hid dim]
# trg_mask = [batch size, trg len]
# src_mask = [batch size, src len]
batch_size = trg.shape[0]
trg_len = trg.shape[1]
pos = (torch.arange(0, trg_len).unsqueeze(0).repeat(batch_size,
1).to(self.device))
# pos = [batch size, trg len]
trg = self.dropout((self.tok_embedding(trg) * self.scale) +
self.pos_embedding(pos))
# trg = [batch size, trg len, hid dim]
for layer in self.layers:
trg, attention = layer(trg, enc_src, trg_mask, src_mask)
# trg = [batch size, trg len, hid dim]
# attention = [batch size, n heads, trg len, src len]
output = self.fc_out(trg)
# output = [batch size, trg len, output dim]
return output, attention
class DecoderLayer(nn.Module):
def __init__(self, hid_dim, n_heads, pf_dim, dropout, device):
super().__init__()
self.self_attn_layer_norm = nn.LayerNorm(hid_dim)
self.enc_attn_layer_norm = nn.LayerNorm(hid_dim)
self.ff_layer_norm = nn.LayerNorm(hid_dim)
self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads,
dropout, device)
self.encoder_attention = MultiHeadAttentionLayer(
hid_dim, n_heads, dropout, device)
self.positionwise_feedforward = PositionwiseFeedforwardLayer(
hid_dim, pf_dim, dropout)
self.dropout = nn.Dropout(dropout)
def forward(self, trg, enc_src, trg_mask, src_mask):
# trg = [batch size, trg len, hid dim]
# enc_src = [batch size, src len, hid dim]
# trg_mask = [batch size, trg len]
# src_mask = [batch size, src len]
# self attention
_trg, _ = self.self_attention(trg, trg, trg, trg_mask)
# dropout, residual connection and layer norm
trg = self.self_attn_layer_norm(trg + self.dropout(_trg))
# trg = [batch size, trg len, hid dim]
# encoder attention
_trg, attention = self.encoder_attention(trg, enc_src, enc_src,
src_mask)
# dropout, residual connection and layer norm
trg = self.enc_attn_layer_norm(trg + self.dropout(_trg))
# trg = [batch size, trg len, hid dim]
# positionwise feedforward
_trg = self.positionwise_feedforward(trg)
# dropout, residual and layer norm
trg = self.ff_layer_norm(trg + self.dropout(_trg))
# trg = [batch size, trg len, hid dim]
# attention = [batch size, n heads, trg len, src len]
return trg, attention
class Seq2Seq(nn.Module, Convo):
def __init__(
self,
encoder,
decoder,
src_pad_idx,
trg_pad_idx,
device,
loader_train: DataLoader,
out: str,
loader_valid=None,
loader_drugex=None,
epochs=100,
lr=0.0005,
clip=0.1,
reverse=True,
TRG=None,
SRC=None,
):
super().__init__()
self.encoder = encoder
self.decoder = decoder
self.src_pad_idx = src_pad_idx
self.trg_pad_idx = trg_pad_idx
self.device = device
self.loader_train = loader_train
self.out = out
self.loader_valid = loader_valid
self.loader_drugex = loader_drugex
self.epochs = epochs
self.lr = lr
self.clip = clip
self.reverse = reverse
self.TRG = TRG
self.SRC = SRC
def make_src_mask(self, src):
# src = [batch size, src len]
src_mask = (src != self.src_pad_idx).unsqueeze(1).unsqueeze(2)
# src_mask = [batch size, 1, 1, src len]
return src_mask
def make_trg_mask(self, trg):
# trg = [batch size, trg len]
trg_pad_mask = (trg != self.trg_pad_idx).unsqueeze(1).unsqueeze(2)
# trg_pad_mask = [batch size, 1, 1, trg len]
trg_len = trg.shape[1]
trg_sub_mask = torch.tril(
torch.ones((trg_len, trg_len), device=self.device)).bool()
# trg_sub_mask = [trg len, trg len]
trg_mask = trg_pad_mask & trg_sub_mask
# trg_mask = [batch size, 1, trg len, trg len]
return trg_mask
def forward(self, src, trg):
# src = [batch size, src len]
# trg = [batch size, trg len]
src_mask = self.make_src_mask(src)
trg_mask = self.make_trg_mask(trg)
# src_mask = [batch size, 1, 1, src len]
# trg_mask = [batch size, 1, trg len, trg len]
enc_src = self.encoder(src, src_mask)
# enc_src = [batch size, src len, hid dim]
output, attention = self.decoder(trg, enc_src, trg_mask, src_mask)
# output = [batch size, trg len, output dim]
# attention = [batch size, n heads, trg len, src len]
return output, attention
def translate_sentence(self, src, trg_field, device, max_len=202):
self.eval()
src_mask = self.make_src_mask(src)
with torch.no_grad():
enc_src = self.encoder(src, src_mask)
trg_indexes = [trg_field.vocab.stoi[trg_field.init_token]]
batch_size = src.shape[0]
trg = torch.LongTensor(trg_indexes).unsqueeze(0).to(device)
trg = trg.repeat(batch_size, 1)
for i in range(max_len):
# turned model into self.
trg_mask = self.make_trg_mask(trg)
with torch.no_grad():
output, attention = self.decoder(trg, enc_src, trg_mask,
src_mask)
pred_tokens = output.argmax(2)[:, -1].unsqueeze(1)
trg = torch.cat((trg, pred_tokens), 1)
return output
def remove_floats(df: pd.DataFrame, subset: str):
"""Preprocessing step to remove any entries that are not strings"""
df_subset = df[subset]
df[subset] = df[subset].astype(str)
# only keep entries that stayed the same after applying astype str
df = df[df[subset] == df_subset].copy()
return df
def smi_tokenizer(smi: str, reverse=False) -> list:
"""
Tokenize a SMILES molecule
"""
pattern = r"(\[[^\]]+]|Br?|Cl?|N|O|S|P|F|I|b|c|n|o|s|p|\(|\)|\.|=|#|-|\+|\\\\|\\|\/|:|~|@|\?|>|\*|\$|\%[0-9]{2}|[0-9])"
regex = re.compile(pattern)
# tokens = ['<sos>'] + [token for token in regex.findall(smi)] + ['<eos>']
tokens = [token for token in regex.findall(smi)]
# assert smi == ''.join(tokens[1:-1])
assert smi == "".join(tokens[:])
# try:
# assert smi == "".join(tokens[:])
# except:
# print(smi)
# print("".join(tokens[:]))
if reverse:
return tokens[::-1]
return tokens
def init_weights(m: nn.Module):
if hasattr(m, "weight") and m.weight.dim() > 1:
nn.init.xavier_uniform_(m.weight.data)
def count_parameters(model: nn.Module):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
def epoch_time(start_time, end_time):
elapsed_time = end_time - start_time
elapsed_mins = int(elapsed_time / 60)
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
return elapsed_mins, elapsed_secs
def initialize_model(folder_out: str,
data_source: str,
error_source: str,
device: torch.device,
threshold: int,
epochs: int,
layers: int = 3,
batch_size: int = 16,
invalid_type: str = "all",
num_errors: int = 1,
validation_step=False):
"""Create encoder decoder models for specified model (currently only translator) & type of invalid SMILES
param data: collection of invalid, valid SMILES pairs
param invalid_smiles_path: path to previously generated invalid SMILES
param invalid_type: type of errors introduced into invalid SMILES
return:
"""
# set fields
SRC = Field(
tokenize=lambda x: smi_tokenizer(x),
init_token="<sos>",
eos_token="<eos>",
batch_first=True,
)
TRG = Field(
tokenize=lambda x: smi_tokenizer(x, reverse=True),
init_token="<sos>",
eos_token="<eos>",
batch_first=True,
)
if validation_step:
train, val = TabularDataset.splits(
path=f'{folder_out}errors/split/',
train=f"{data_source}_{invalid_type}_{num_errors}_errors_train.csv",
validation=
f"{data_source}_{invalid_type}_{num_errors}_errors_dev.csv",
format="CSV",
skip_header=False,
fields={
"ERROR": ("src", SRC),
"STD_SMILES": ("trg", TRG)
},
)
SRC.build_vocab(train, val, max_size=1000)
TRG.build_vocab(train, val, max_size=1000)
else:
train = TabularDataset(
path=
f'{folder_out}{data_source}_{invalid_type}_{num_errors}_errors.csv',
format="CSV",
skip_header=False,
fields={
"ERROR": ("src", SRC),
"STD_SMILES": ("trg", TRG)
},
)
SRC.build_vocab(train, max_size=1000)
TRG.build_vocab(train, max_size=1000)
drugex = TabularDataset(
path=error_source,
format="csv",
skip_header=False,
fields={
"SMILES": ("src", SRC),
"SMILES_TARGET": ("trg", TRG)
},
)
#SRC.vocab = torch.load('vocab_src.pth')
#TRG.vocab = torch.load('vocab_trg.pth')
# model parameters
EPOCHS = epochs
BATCH_SIZE = batch_size
INPUT_DIM = len(SRC.vocab)
OUTPUT_DIM = len(TRG.vocab)
HID_DIM = 256
ENC_LAYERS = layers
DEC_LAYERS = layers
ENC_HEADS = 8
DEC_HEADS = 8
ENC_PF_DIM = 512
DEC_PF_DIM = 512
ENC_DROPOUT = 0.1
DEC_DROPOUT = 0.1
SRC_PAD_IDX = SRC.vocab.stoi[SRC.pad_token]
TRG_PAD_IDX = TRG.vocab.stoi[TRG.pad_token]
# add 2 to length for start and stop tokens
MAX_LENGTH = threshold + 2
# model name
MODEL_OUT_FOLDER = f"{folder_out}"
MODEL_NAME = "transformer_%s_%s_%s_%s_%s" % (
invalid_type, num_errors, data_source, BATCH_SIZE, layers)
if not os.path.exists(MODEL_OUT_FOLDER):
os.mkdir(MODEL_OUT_FOLDER)
out = os.path.join(MODEL_OUT_FOLDER, MODEL_NAME)
torch.save(SRC.vocab, f'{out}_vocab_src.pth')
torch.save(TRG.vocab, f'{out}_vocab_trg.pth')
# iterator is a dataloader
# iterator to pass to the same length and create batches in which the
# amount of padding is minimized
if validation_step:
train_iter, val_iter = BucketIterator.splits(
(train, val),
batch_sizes=(BATCH_SIZE, 256),
sort_within_batch=True,
shuffle=True,
# the BucketIterator needs to be told what function it should use to
# group the data.
sort_key=lambda x: len(x.src),
device=device,
)
else:
train_iter = BucketIterator(
train,
batch_size=BATCH_SIZE,
sort_within_batch=True,
shuffle=True,
# the BucketIterator needs to be told what function it should use to
# group the data.
sort_key=lambda x: len(x.src),
device=device,
)
val_iter = None
drugex_iter = Iterator(
drugex,
batch_size=64,
device=device,
sort=False,
sort_within_batch=True,
sort_key=lambda x: len(x.src),
repeat=False,
)
# model initialization
enc = Encoder(
INPUT_DIM,
HID_DIM,
ENC_LAYERS,
ENC_HEADS,
ENC_PF_DIM,
ENC_DROPOUT,
MAX_LENGTH,
device,
)
dec = Decoder(
OUTPUT_DIM,
HID_DIM,
DEC_LAYERS,
DEC_HEADS,
DEC_PF_DIM,
DEC_DROPOUT,
MAX_LENGTH,
device,
)
model = Seq2Seq(
enc,
dec,
SRC_PAD_IDX,
TRG_PAD_IDX,
device,
train_iter,
out=out,
loader_valid=val_iter,
loader_drugex=drugex_iter,
epochs=EPOCHS,
TRG=TRG,
SRC=SRC,
).to(device)
return model, out, SRC
def train_model(model, out, assess):
"""Apply given weights (& assess performance or train further) or start training new model
Args:
model: initialized model
out: .pkg file with model parameters
asses: bool
Returns:
model with (new) weights
"""
if os.path.exists(f"{out}.pkg") and assess:
model.load_state_dict(torch.load(f=out + ".pkg"))
(
valids,
loss_valid,
valids_de,
df_output,
df_output_de,
right_molecules,
complexity,
unchanged,
unchanged_de,
) = model.evaluate(True)
# log = open('unchanged.log', 'a')
# info = f'type: comb unchanged: {unchan:.4g} unchanged_drugex: {unchan_de:.4g}'
# print(info, file=log, flush = True)
# print(valids_de)
# print(unchanged_de)
# print(unchan)
# print(unchan_de)
# df_output_de.to_csv(f'{out}_de_new.csv', index = False)
# error_de = 1 - valids_de / len(drugex_iter.dataset)
# print(error_de)
# df_output.to_csv(f'{out}_par.csv', index = False)
elif os.path.exists(f"{out}.pkg"):
# starts from the model after the last epoch, not the best epoch
model.load_state_dict(torch.load(f=out + "_last.pkg"))
# need to change how log file names epochs
model.train_model()
else:
model = model.apply(init_weights)
model.train_model()
return model
def correct_SMILES(model, out, error_source, device, SRC):
"""Model that is given corrects SMILES and return number of correct ouputs and dataframe containing all outputs
Args:
model: initialized model
out: .pkg file with model parameters
asses: bool
Returns:
valids: number of fixed outputs
df_output: dataframe containing output (either correct or incorrect) & original input
"""
## account for tokens that are not yet in SRC without changing existing SRC token embeddings
errors = TabularDataset(
path=error_source,
format="csv",
skip_header=False,
fields={"SMILES": ("src", SRC)},
)
errors_loader = Iterator(
errors,
batch_size=64,
device=device,
sort=False,
sort_within_batch=True,
sort_key=lambda x: len(x.src),
repeat=False,
)
model.load_state_dict(torch.load(f=out + ".pkg",map_location=torch.device('cpu')))
# add option to use different iterator maybe?
valids, df_output = model.translate(errors_loader)
#df_output.to_csv(f"{error_source}_fixed.csv", index=False)
return valids, df_output
class smi_correct(object):
def __init__(self, model_name, trans_file_path):
# set random seed, used for error generation & initiation transformer
self.SEED = 42
random.seed(self.SEED)
self.model_name = model_name
self.folder_out = "DrugGEN/data/"
self.trans_file_path = trans_file_path
if not os.path.exists(self.folder_out):
os.makedirs(self.folder_out)
self.invalid_type = 'multiple'
self.num_errors = 12
self.threshold = 200
self.data_source = f"PAPYRUS_{self.threshold}"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
self.initialize_source = 'DrugGEN/data/papyrus_rnn_S.csv' # change this path
def standardization_pipeline(self, smile):
desalter = MolStandardize.fragment.LargestFragmentChooser()
std_smile = None
if not isinstance(smile, str): return None
m = Chem.MolFromSmiles(smile)
# skips smiles for which no mol file could be generated
if m is not None:
# standardizes
std_m = standardizer.standardize_mol(m)
# strips salts
std_m_p, exclude = standardizer.get_parent_mol(std_m)
if not exclude:
# choose largest fragment for rare cases where chembl structure
# pipeline leaves 2 fragments
std_m_p_d = desalter.choose(std_m_p)
std_smile = Chem.MolToSmiles(std_m_p_d)
return std_smile
def remove_smiles_duplicates(self, dataframe: pd.DataFrame,
subset: str) -> pd.DataFrame:
return dataframe.drop_duplicates(subset=subset)
def correct(self, smi):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, out, SRC = initialize_model(self.folder_out,
self.data_source,
error_source=self.initialize_source,
device=device,
threshold=self.threshold,
epochs=30,
layers=3,
batch_size=16,
invalid_type=self.invalid_type,
num_errors=self.num_errors)
valids, df_output = correct_SMILES(model, out, smi, device,
SRC)
df_output["SMILES"] = df_output.apply(lambda row: self.standardization_pipeline(row["CORRECT"]), axis=1)
df_output = self.remove_smiles_duplicates(df_output, subset="SMILES").drop(columns=["CORRECT", "INCORRECT", "ORIGINAL"]).dropna()
return df_output