File size: 7,792 Bytes
8019be0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
#!/usr/bin/env
import os
import torch
from torch.utils.data import Dataset, DataLoader
from datasets import Dataset,load_from_disk
import sys
import pytorch_lightning as pl
from pep_scoring.tokenizer.my_tokenizers import SMILES_SPE_Tokenizer
from functools import partial
import re

# Directory containing this file; used to resolve the in-repo tokenizer files.
_THIS_DIR = os.path.dirname(os.path.abspath(__file__))


class DynamicBatchingDataset(Dataset):
    def __init__(self, dataset_dict, tokenizer):
        print('Initializing dataset...')
        self.dataset_dict = {
            'attention_mask': [torch.tensor(item) for item in dataset_dict['attention_mask']],
            'input_ids': [torch.tensor(item) for item in dataset_dict['input_ids']],
            'labels': dataset_dict['labels']
        }
        self.tokenizer = tokenizer

    def __len__(self):
        return len(self.dataset_dict['attention_mask'])

    def __getitem__(self, idx):
        if isinstance(idx, int):
            return {
                'input_ids': self.dataset_dict['input_ids'][idx],
                'attention_mask': self.dataset_dict['attention_mask'][idx],
                'labels': self.dataset_dict['labels'][idx]
            }
        elif isinstance(idx, list):
            return {
                'input_ids': [self.dataset_dict['input_ids'][i] for i in idx],
                'attention_mask': [self.dataset_dict['attention_mask'][i] for i in idx],
                'labels': [self.dataset_dict['labels'][i] for i in idx]
            }   
        else:
            raise ValueError(f"Expected idx to be int or list, but got {type(idx)}")   

class CustomDataModule(pl.LightningDataModule):
    def __init__(self, dataset_path, tokenizer):
        super().__init__()
        self.dataset = load_from_disk(dataset_path)
        self.tokenizer = tokenizer
        
    def peptide_bond_mask(self, smiles_list):
        """

        Returns a mask with shape (batch_size, seq_length) that has 1 at the locations

        of recognized bonds in the positions dictionary and 0 elsewhere.



        Args:

            smiles_list: List of peptide SMILES strings (batch of SMILES strings).



        Returns:

            np.ndarray: A mask of shape (batch_size, seq_length) with 1s at bond positions.

        """
        # Initialize the batch mask
        batch_size = len(smiles_list)
        max_seq_length = 1035 #max(len(smiles) for smiles in smiles_list)  # Find the longest SMILES
        mask = torch.zeros((batch_size, max_seq_length), dtype=torch.int)  # Mask filled with zeros

        bond_patterns = [
            (r'OC\(=O\)', 'ester'),
            (r'N\(C\)C\(=O\)', 'n_methyl'),
            (r'N[12]C\(=O\)', 'peptide'),  # Pro peptide bonds
            (r'NC\(=O\)', 'peptide'),  # Regular peptide bonds
            (r'C\(=O\)N\(C\)', 'n_methyl'),
            (r'C\(=O\)N[12]?', 'peptide')
        ]

        for batch_idx, smiles in enumerate(smiles_list):
            positions = []
            used = set()

            # Identify bonds
            for pattern, bond_type in bond_patterns:
                for match in re.finditer(pattern, smiles):
                    if not any(p in range(match.start(), match.end()) for p in used):
                        positions.append({
                            'start': match.start(),
                            'end': match.end(),
                            'type': bond_type,
                            'pattern': match.group()
                        })
                        used.update(range(match.start(), match.end()))

            # Update the mask for the current SMILES
            for pos in positions:
                mask[batch_idx, pos['start']:pos['end']] = 1

        return mask

    def peptide_token_mask(self, smiles_list, token_lists):
        """

        Returns a mask with shape (batch_size, num_tokens) that has 1 for tokens

        where any part of the token overlaps with a peptide bond, and 0 elsewhere.



        Args:

            smiles_list: List of peptide SMILES strings (batch of SMILES strings).

            token_lists: List of tokenized SMILES strings (split into tokens).



        Returns:

            np.ndarray: A mask of shape (batch_size, num_tokens) with 1s for peptide bond tokens.

        """
        # Initialize the batch mask
        batch_size = len(smiles_list)
        token_seq_length = max(len(tokens) for tokens in token_lists)  # Find the longest tokenized sequence
        tokenized_masks = torch.zeros((batch_size, token_seq_length), dtype=torch.int)  # Mask filled with zeros
        atomwise_masks = self.peptide_bond_mask(smiles_list)

    
        for batch_idx, atomwise_mask in enumerate(atomwise_masks):
            token_seq = token_lists[batch_idx]
            atom_idx = 0
            
            for token_idx, token in enumerate(token_seq):
                if token_idx != 0 and token_idx != len(token_seq) - 1:
                    if torch.sum(atomwise_mask[atom_idx:atom_idx+len(token)]) >= 1:
                        tokenized_masks[batch_idx][token_idx] = 1
                    atom_idx += len(token)
        
        return tokenized_masks
    
    def collate_fn(self, batch):
        item = batch[0]
            
        token_array = self.tokenizer.get_token_split(item['input_ids'])
        bond_mask = self.peptide_token_mask(item['labels'], token_array)

        return {
            'input_ids': item['input_ids'],
            'attention_mask': item['attention_mask'],
            'bond_mask': bond_mask
        } 

    def train_dataloader(self):
        train_dataset = DynamicBatchingDataset(self.dataset['train'], tokenizer=self.tokenizer)
        return DataLoader(
            train_dataset, 
            batch_size=1, 
            collate_fn=self.collate_fn,  # Use the instance method
            shuffle=True, 
            num_workers=12, 
            pin_memory=True
        )

    def val_dataloader(self):
        val_dataset = DynamicBatchingDataset(self.dataset['val'], tokenizer=self.tokenizer)
        return DataLoader(
            val_dataset, 
            batch_size=1, 
            collate_fn=self.collate_fn,  # Use the instance method
            num_workers=8, 
            pin_memory=True
        )
        

def setup_data_and_update_config(config):
    """

    Get the dataset and update the config with token information for text datasets.

    """
    # SMILES Pair Encoding tokenizer ships with the repo under pep_scoring/tokenizer/.
    tokenizer = SMILES_SPE_Tokenizer(
        os.path.join(_THIS_DIR, 'pep_scoring', 'tokenizer', 'new_vocab.txt'),
        os.path.join(_THIS_DIR, 'pep_scoring', 'tokenizer', 'new_splits.txt'),
    )

    config.interpolant.tokens = len(tokenizer)
    config.interpolant.pad_token = tokenizer.pad_token_id
    config.interpolant.mask_token = tokenizer.mask_token_id

    # Path to the preprocessed (arrow) pretraining dataset saved via `save_to_disk`.
    # Download instructions are in the README; override with `training.data_path` in the config.
    data_path = config.training.get('data_path', os.path.join('data', '11M_peptide_smiles'))
    if not os.path.isabs(data_path):
        data_path = os.path.join(_THIS_DIR, data_path)
    if not os.path.exists(data_path):
        raise FileNotFoundError(
            f"Pretraining dataset not found at '{data_path}'. Download it (see a2d2_pep/README.md, "
            "'Pretraining the Any-Length Model') and set `training.data_path` in config_pep.yaml."
        )
    data_module = CustomDataModule(data_path, tokenizer)

    return data_module