torch-molecule 0.1.0.dev2__py3-none-any.whl

torch_molecule/generator/lstm/smiles_char_dict.py

@@ -0,0 +1,123 @@
+class SmilesCharDictionary(object):
+    """
+    A fixed dictionary for druglike SMILES.
+    Enables smile<->token conversion.
+
+    With a space:0 for padding, Q:1 as the start token and end_of_line \n:2 as the stop token.
+    """
+
+    PAD = ' '
+    BEGIN = 'Q'
+    END = '\n'
+
+    def __init__(self) -> None:
+        self.forbidden_symbols = {'Ag', 'Al', 'Am', 'Ar', 'At', 'Au', 'D', 'E', 'Fe', 'G', 'K', 'L', 'M', 'Ra', 'Re',
+                                  'Rf', 'Rg', 'Rh', 'Ru', 'T', 'U', 'V', 'W', 'Xe',
+                                  'Y', 'Zr', 'a', 'd', 'f', 'g', 'h', 'k', 'm', 'si', 't', 'te', 'u', 'v', 'y'}
+
+        self.char_idx = {self.PAD: 0, self.BEGIN: 1, self.END: 2, '#': 20, '%': 22, '(': 25, ')': 24, '+': 26, '-': 27,
+                         '.': 30,
+                        #  '*': 30,
+                         '*': 47,
+                         '0': 32, '1': 31, '2': 34, '3': 33, '4': 36, '5': 35, '6': 38, '7': 37, '8': 40,
+                         '9': 39, '=': 41, 'A': 7, 'B': 11, 'C': 19, 'F': 4, 'H': 6, 'I': 5, 'N': 10,
+                         'O': 9, 'P': 12, 'S': 13, 'X': 15, 'Y': 14, 'Z': 3, '[': 16, ']': 18,
+                         'b': 21, 'c': 8, 'n': 17, 'o': 29, 'p': 23, 's': 28,
+                         "@": 42, "R": 43, '/': 44, "\\": 45, 'E': 46, 
+                         }
+
+        self.idx_char = {v: k for k, v in self.char_idx.items()}
+        self.encode_dict = {"Br": 'Y', "Cl": 'X', "Si": 'A', 'Se': 'Z', '@@': 'R', 'se': 'E'}
+        self.decode_dict = {v: k for k, v in self.encode_dict.items()}
+
+    def allowed(self, smiles) -> bool:
+        """
+        Determine if smiles string has illegal symbols
+
+        Args:
+            smiles: SMILES string
+
+        Returns:
+            True if all legal
+        """
+        for symbol in self.forbidden_symbols:
+            if symbol in smiles:
+                # print('Forbidden symbol {:<2}  in  {}'.format(symbol, smiles))
+                return False
+        return True
+
+    def encode(self, smiles: str) -> str:
+        """
+        Replace multi-char tokens with single tokens in SMILES string.
+
+        Args:
+            smiles: SMILES string
+
+        Returns:
+            sanitized SMILE string with only single-char tokens
+        """
+
+        temp_smiles = smiles
+        for symbol, token in self.encode_dict.items():
+            temp_smiles = temp_smiles.replace(symbol, token)
+        return temp_smiles
+
+    def decode(self, smiles):
+        """
+        Replace special tokens with their multi-character equivalents.
+
+        Args:
+            smiles: SMILES string
+
+        Returns:
+            SMILES string with possibly multi-char tokens
+        """
+        temp_smiles = smiles
+        for symbol, token in self.decode_dict.items():
+            temp_smiles = temp_smiles.replace(symbol, token)
+        return temp_smiles
+
+    def get_char_num(self) -> int:
+        """
+        Returns:
+            number of characters in the alphabet
+        """
+        return len(self.idx_char)
+
+    @property
+    def begin_idx(self) -> int:
+        return self.char_idx[self.BEGIN]
+
+    @property
+    def end_idx(self) -> int:
+        return self.char_idx[self.END]
+
+    @property
+    def pad_idx(self) -> int:
+        return self.char_idx[self.PAD]
+
+    def matrix_to_smiles(self, array):
+        """
+        Converts an matrix of indices into their SMILES representations
+
+        Args:
+            array: torch tensor of indices, one molecule per row
+
+        Returns: a list of SMILES, without the termination symbol
+        """
+        smiles_strings = []
+
+        for row in array:
+            predicted_chars = []
+
+            for j in row:
+                next_char = self.idx_char[j.item()]
+                if next_char == self.END:
+                    break
+                predicted_chars.append(next_char)
+
+            smi = ''.join(predicted_chars)
+            smi = self.decode(smi)
+            smiles_strings.append(smi)
+
+        return smiles_strings

torch_molecule/generator/lstm/utils.py

@@ -0,0 +1,62 @@
+from rdkit import Chem
+from typing import Set, Any, Optional
+import torch
+from .smiles_char_dict import SmilesCharDictionary
+
+def canonicalize(smiles: str, include_stereocenters=True) -> Optional[str]:
+    """
+    Canonicalize the SMILES strings with RDKit.
+    The algorithm is detailed under https://pubs.acs.org/doi/full/10.1021/acs.jcim.5b00543
+    Args:
+        smiles: SMILES string to canonicalize
+        include_stereocenters: whether to keep the stereochemical information in the canonical SMILES string
+    Returns:
+        Canonicalized SMILES string, None if the molecule is invalid.
+    """
+
+    if smiles is None:
+        return None
+
+    if len(smiles) > 0:
+        mol = Chem.MolFromSmiles(smiles)
+
+        if mol is not None:
+            return Chem.MolToSmiles(mol, isomericSmiles=include_stereocenters)
+        else:
+            return None
+    else:
+        return None
+
+def remove_duplicates(list_with_duplicates):
+    """
+    Removes the duplicates and keeps the ordering of the original list.
+    For duplicates, the first occurrence is kept and the later occurrences are ignored.
+    Args:
+        list_with_duplicates: list that possibly contains duplicates
+    Returns:
+        A list with no duplicates.
+    """
+    unique_set: Set[Any] = set()
+    unique_list = []
+    for element in list_with_duplicates:
+        if element not in unique_set:
+            unique_set.add(element)
+            unique_list.append(element)
+
+    return unique_list
+
+
+def rnn_start_token_vector(batch_size, device='cpu'):
+    """
+    Returns a vector of start tokens for LSTM.
+    This vector can be used to start sampling a batch of SMILES strings.
+
+    Args:
+        batch_size: how many SMILES will be generated at the same time in LSTM
+        device: cpu | cuda
+
+    Returns:
+        a tensor (batch_size x 1) containing the start token
+    """
+    sd = SmilesCharDictionary()
+    return torch.LongTensor(batch_size, 1).fill_(sd.begin_idx).to(device)

torch_molecule/generator/molgpt/__init__.py

@@ -0,0 +1,3 @@
+from .modeling_molgpt import MolGPTMolecularGenerator
+
+__all__ = ['MolGPTMolecularGenerator']

torch_molecule/generator/molgpt/dataset.py

@@ -0,0 +1,87 @@
+from torch.utils.data import Dataset
+import torch
+
+class SmilesDataset(Dataset):
+    """Dataset for SMILES strings for MolGPT training"""
+    
+    def __init__(self, data, regex, max_len, properties=None, 
+                 scaffolds=None, scaffold_maxlen=None):
+        """
+        Initialize the dataset.
+        
+        Parameters
+        ----------
+        data : List[str]
+            List of SMILES strings
+        regex : str
+            Regex for tokenization
+        max_len : int
+            Maximum sequence length
+        properties : Optional[List[List[float]]]
+            Property values for conditional generation
+        scaffolds : Optional[List[str]]
+            Scaffold SMILES for conditional generation
+        scaffold_maxlen : Optional[int]
+            Maximum length of scaffold SMILES
+        """
+        
+        # Generate content by decomposing all SMILES strings
+        content = set()
+        for smile in data:
+            tokens = regex.findall(smile.strip())
+            content.update(tokens)
+        
+        # Add padding token
+        content.add('<')
+        
+        chars = sorted(list(set(content)))
+        data_size, vocab_size = len(data), len(chars)
+        
+        self.stoi = {ch: i for i, ch in enumerate(chars)}
+        self.itos = {i: ch for i, ch in enumerate(chars)}
+        self.max_len = max_len
+        self.vocab_size = vocab_size
+        self.data = data
+        self.properties = properties
+        self.scaffolds = scaffolds
+        self.scaffold_maxlen = scaffold_maxlen
+        self.regex = regex
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def __getitem__(self, idx):
+        # Get SMILES and property
+        smiles = self.data[idx].strip()
+        prop = self.properties[idx] if self.properties is not None else [0.0]
+        
+        # Get scaffold if available
+        if self.scaffolds is not None:
+            scaffold = self.scaffolds[idx].strip()
+        else:
+            scaffold = ""
+        
+        # Tokenize SMILES
+        tokens = self.regex.findall(smiles)
+        # Pad with '<' if needed
+        tokens += ['<'] * (self.max_len - len(tokens))
+        # Convert to indices
+        indices = [self.stoi.get(t, 0) for t in tokens]
+        
+        # Create input and target tensors
+        x = torch.tensor(indices[:-1], dtype=torch.long)
+        y = torch.tensor(indices[1:], dtype=torch.long)
+        
+        # Create property tensor
+        prop_tensor = torch.tensor(prop, dtype=torch.float)
+        
+        # Create scaffold tensor if needed
+        if self.scaffolds is not None:
+            scaffold_tokens = self.regex.findall(scaffold)
+            scaffold_tokens += ['<'] * (self.scaffold_maxlen - len(scaffold_tokens))
+            scaffold_indices = [self.stoi.get(t, 0) for t in scaffold_tokens]
+            scaffold_tensor = torch.tensor(scaffold_indices, dtype=torch.long)
+        else:
+            scaffold_tensor = torch.tensor([], dtype=torch.long)
+        
+        return x, y, prop_tensor, scaffold_tensor

torch_molecule/generator/molgpt/gpt.py

@@ -0,0 +1,251 @@
+"""
+GPT model:
+- the initial stem consists of a combination of token encoding and a positional encoding
+- the meat of it is a uniform sequence of Transformer blocks
+    - each Transformer is a sequential combination of a 1-hidden-layer MLP block and a self-attention block
+    - all blocks feed into a central residual pathway similar to resnets
+- the final decoder is a linear projection into a vanilla Softmax classifier
+"""
+
+import math
+import logging
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class CausalSelfAttention(nn.Module):
+    """
+    A vanilla multi-head masked self-attention layer with a projection at the end.
+    It is possible to use torch.nn.MultiheadAttention here but I am including an
+    explicit implementation here to show that there is nothing too scary here.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.hidden_size % config.num_head == 0
+        # key, query, value projections for all heads
+        self.key = nn.Linear(config.hidden_size, config.hidden_size)
+        self.query = nn.Linear(config.hidden_size, config.hidden_size)
+        self.value = nn.Linear(config.hidden_size, config.hidden_size)
+        # regularization
+        self.attn_drop = nn.Dropout(config.attn_pdrop)
+        self.resid_drop = nn.Dropout(config.resid_pdrop)
+        # output projection
+        self.proj = nn.Linear(config.hidden_size, config.hidden_size)
+        # causal mask to ensure that attention is only applied to the left in the input sequence
+        num = int(bool(config.num_task)) + int(config.scaffold_maxlen)   #int(config.lstm_layers)    #  int(config.use_scaffold) 
+        # num = 1
+        self.register_buffer("mask", torch.tril(torch.ones(config.max_len + num, config.max_len + num))
+                                     .view(1, 1, config.max_len + num, config.max_len + num))
+
+        self.n_head = config.num_head
+
+    def forward(self, x, layer_past=None):
+        B, T, C = x.size()
+
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        k = self.key(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = self.query(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = self.value(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        attn_save = att
+        att = self.attn_drop(att)
+        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+
+        # output projection
+        y = self.resid_drop(self.proj(y))
+        return y, attn_save
+
+class Block(nn.Module):
+    """ an unassuming Transformer block """
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(config.hidden_size)
+        self.ln2 = nn.LayerNorm(config.hidden_size)
+        self.attn = CausalSelfAttention(config)
+        self.mlp = nn.Sequential(
+            nn.Linear(config.hidden_size, 4 * config.hidden_size),
+            nn.GELU(),
+            nn.Linear(4 * config.hidden_size, config.hidden_size),
+            nn.Dropout(config.resid_pdrop),
+        )
+
+    def forward(self, x):
+        y, attn = self.attn(self.ln1(x))
+        x = x + y
+        x = x + self.mlp(self.ln2(x))
+        return x, attn
+
+class GPTConfig:
+    """ base GPT config, params common to all GPT versions """
+    embd_pdrop = 0.1
+    resid_pdrop = 0.1
+    attn_pdrop = 0.1
+
+    def __init__(self, vocab_size, max_len, **kwargs):
+        self.vocab_size = vocab_size
+        self.max_len = max_len
+        for k,v in kwargs.items():
+            setattr(self, k, v)
+
+class GPT(nn.Module):
+    """  the full GPT language model, with a context size of max_len """
+
+    def __init__(self, vocab_size, max_len, num_task, num_layer, num_head, hidden_size, use_scaffold, scaffold_maxlen, use_lstm, lstm_layers):
+        super().__init__()
+        config = GPTConfig(vocab_size, max_len, num_task = num_task, num_layer = num_layer, num_head = num_head, hidden_size = hidden_size, use_scaffold = use_scaffold, scaffold_maxlen = scaffold_maxlen, use_lstm = use_lstm, lstm_layers = lstm_layers)
+
+        # input embedding stem
+        self.config = config
+        self.tok_emb = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.type_emb = nn.Embedding(2, config.hidden_size)
+        if config.num_task:
+            self.prop_nn = nn.Linear(config.num_task, config.hidden_size)
+     
+        self.pos_emb = nn.Parameter(torch.zeros(1, config.max_len, config.hidden_size))
+        self.drop = nn.Dropout(config.embd_pdrop)
+        # transformer
+        self.blocks = nn.Sequential(*[Block(config) for _ in range(config.num_layer)])
+        # decoder head
+        self.ln_f = nn.LayerNorm(config.hidden_size)
+        self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.max_len = config.max_len
+
+        if config.use_lstm:
+            self.lstm = nn.LSTM(input_size = config.hidden_size, hidden_size = config.hidden_size, num_layers = config.lstm_layers, dropout = 0.3, bidirectional = False)
+    
+    def get_max_len(self):
+        return self.max_len
+
+    def initialize_parameters(self, seed=None):
+        if seed is not None:
+            torch.manual_seed(seed)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+            if isinstance(module, nn.Linear) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def configure_optimizers(self, train_config):
+        """
+        This long function is unfortunately doing something very simple and is being very defensive:
+        We are separating out all parameters of the model into two buckets: those that will experience
+        weight decay for regularization and those that won't (biases, and layernorm/embedding weights).
+        We are then returning the PyTorch optimizer object.
+        """
+
+        # separate out all parameters to those that will and won't experience regularizing weight decay
+        decay = set()
+        no_decay = set()
+        whitelist_weight_modules = (torch.nn.Linear, torch.nn.LSTM)
+        blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
+        for mn, m in self.named_modules():
+            for pn, p in m.named_parameters():
+                fpn = '%s.%s' % (mn, pn) if mn else pn # full param name
+
+                if pn.endswith('bias') or ('bias' in pn):
+                    # all biases will not be decayed
+                    no_decay.add(fpn)
+                elif (pn.endswith('weight') or ('weight' in pn)) and isinstance(m, whitelist_weight_modules):
+                    # weights of whitelist modules will be weight decayed
+                    decay.add(fpn)
+                elif pn.endswith('weight') and isinstance(m, blacklist_weight_modules):
+                    # weights of blacklist modules will NOT be weight decayed
+                    no_decay.add(fpn)
+
+        # special case the position embedding parameter in the root GPT module as not decayed
+        no_decay.add('pos_emb')
+
+        # validate that we considered every parameter
+        param_dict = {pn: p for pn, p in self.named_parameters()}
+        inter_params = decay & no_decay
+        union_params = decay | no_decay
+        assert len(inter_params) == 0, "parameters %s made it into both decay/no_decay sets!" % (str(inter_params), )
+        assert len(param_dict.keys() - union_params) == 0, "parameters %s were not separated into either decay/no_decay set!" \
+                                                    % (str(param_dict.keys() - union_params), )
+
+        # create the pytorch optimizer object
+        optim_groups = [
+            {"params": [param_dict[pn] for pn in sorted(list(decay))], "weight_decay": train_config["weight_decay"]},
+            {"params": [param_dict[pn] for pn in sorted(list(no_decay))], "weight_decay": 0.0},
+        ]
+        optimizer = torch.optim.AdamW(optim_groups, lr=train_config["learning_rate"], betas=train_config["betas"])
+        return optimizer
+
+    def compute_loss(self, x, targets, prop = None, scaffold = None):
+        logits, attn_maps = self.forward(x, targets, prop, scaffold)
+        loss = F.cross_entropy(logits.reshape(-1, logits.size(-1)), targets.view(-1))
+        return loss
+
+    def forward(self, idx, targets=None, prop = None, scaffold = None):
+        b, t = idx.size()
+        assert t <= self.max_len, "Cannot forward, model max_len is exhausted."
+
+        if self.config.num_task:
+            assert prop.size(-1) == self.config.num_task, "num_task should be equal to last dim of property vector"           
+
+        # forward the GPT model
+        token_embeddings = self.tok_emb(idx) # each index maps to a (learnable) vector
+        position_embeddings = self.pos_emb[:, :t, :] # each position maps to a (learnable) vector
+        type_embeddings = self.type_emb(torch.ones((b,t), dtype = torch.long, device = idx.device))
+        x = self.drop(token_embeddings + position_embeddings + type_embeddings)
+
+        if self.config.num_task:
+            type_embd = self.type_emb(torch.zeros((b, 1), dtype = torch.long, device = idx.device))
+            if prop.ndim == 2:
+                p = self.prop_nn(prop.unsqueeze(1))    # for single property
+            else:
+                p = self.prop_nn(prop)    # for multiproperty
+            p += type_embd
+            x = torch.cat([p, x], 1)
+
+        if self.config.use_scaffold:
+            type_embd = self.type_emb(torch.zeros((b, 1), dtype = torch.long, device = idx.device))
+
+            scaffold_embeds = self.tok_emb(scaffold)     # .mean(1, keepdim = True)
+            if self.config.use_lstm:
+                scaffold_embeds = self.lstm(scaffold_embeds.permute(1,0,2))[1][0]
+                # scaffold_embeds = scaffold_embeds.reshape(scaffold_embeds.shape[1], scaffold_embeds.shape[0], 2, self.config.hidden_size).mean(2)
+                scaffold_embeds = scaffold_embeds.permute(1,0,2)   # mean(0, keepdim = True)
+                # scaffold_embeds = scaffold_embeds.reshape(self.config.lstm_layers, 1, -1, self.config.hidden_size)[-1].permute(1,0,2)
+                # scaffold_embeds = scaffold_embeds.reshape(scaffold_embeds.shape[1], scaffold_embeds.shape[0], self.config.hidden_size)
+            scaffold_embeds += type_embd
+            x = torch.cat([scaffold_embeds, x], 1)
+
+        # x = self.blocks(x)
+        attn_maps = []
+
+        for layer in self.blocks:
+            x, attn = layer(x)
+            attn_maps.append(attn)
+
+        x = self.ln_f(x)
+        logits = self.head(x)
+
+        # print(logits.shape)
+        if self.config.num_task and self.config.use_scaffold:
+            num = int(bool(self.config.num_task)) + int(self.config.scaffold_maxlen)
+        elif self.config.num_task:
+            num = int(bool(self.config.num_task))
+        elif self.config.use_scaffold:
+            num = int(self.config.scaffold_maxlen) 
+        else:
+            num = 0
+
+        logits = logits[:, num:, :]
+
+        return logits, attn_maps # (num_layers, batch_size, num_heads, max_seq_len, max_seq_len)