tokenizer_bbb.py

import torch
import torch.nn as nn
from transformers import PreTrainedTokenizer
from transformers.tokenization_utils_base import BatchEncoding
from transformers import AutoTokenizer, AutoModel
from rdkit import Chem
from rdkit.Chem import Descriptors, AllChem, MACCSkeys
from rdkit.ML.Descriptors import MoleculeDescriptors
from rdkit import RDLogger
from rdkit.Chem import Draw
import joblib
import numpy as np
import os
from huggingface_hub import snapshot_download
import warnings
from sklearn.exceptions import InconsistentVersionWarning
from torchvision import models, transforms
from PIL import Image
warnings.filterwarnings("ignore", category=InconsistentVersionWarning)
RDLogger.DisableLog('rdApp.*')

class BBBTokenizer(PreTrainedTokenizer):
  def __init__(self, **kwargs):
    super().__init__(**kwargs)
    
    self.calc = MoleculeDescriptors.MolecularDescriptorCalculator([i[0] for i in Descriptors.descList])

    self.tokenizer = AutoTokenizer.from_pretrained('DeepChem/ChemBERTa-100M-MLM')
    self.chemberta = AutoModel.from_pretrained('DeepChem/ChemBERTa-100M-MLM').eval()

    self.resnet50_backbone = models.resnet50(weights="IMAGENET1K_V1")
    self.resnet = nn.Sequential(*list(self.resnet50_backbone.children())[:-1]).eval()
    self.img_preprocess = transforms.Compose([
      transforms.Resize((224, 224)),
      transforms.ToTensor(),
      transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225],
      )
    ])

    self.feature_transformer_tab = None
    self.feature_transformer_img = None
    self.feature_transformer_txt = None
    self.task = None

  def generate_tab_features(self, smiles):
    mol = Chem.MolFromSmiles(smiles)
    
    if mol is None:
      return torch.tensor(self.feature_transformer_tab.n_features_in_, dtype=torch.float32)
    
    rdkit_2d = np.array(self.calc.CalcDescriptors(mol))
    rdkit_2d[np.isinf(rdkit_2d)] = np.nan
    rdkit_2d = np.nan_to_num(rdkit_2d, nan=0.0, posinf=0.0, neginf=0.0)
    maccs = np.array(list(MACCSkeys.GenMACCSKeys(mol).ToBitString()), dtype=int)
    tab_input = np.concatenate([rdkit_2d, maccs])
    tab_input = self.feature_transformer_tab.transform(tab_input.reshape(1, -1))[0]
    tab_input = np.clip(tab_input, -1e5, 1e5)
    return torch.tensor(tab_input, dtype=torch.float32)

  def generate_img_features(self, smiles):
    mol = Chem.MolFromSmiles(smiles)
    if mol is None:
      img = Image.new("RGB", (300,300), color=(0,0,0))
    else:
      img = Draw.MolToImage(mol, size=(300, 300))
    img = self.img_preprocess(img)
    with torch.no_grad():
      img_input = self.resnet(img.unsqueeze(0)).squeeze(-1).squeeze(-1)
    img_input = self.feature_transformer_img.transform(img_input.reshape(1, -1))[0]
    return torch.tensor(img_input, dtype=torch.float32)

  def generate_txt_features(self, smiles):
    encoded = self.tokenizer(smiles, return_tensors="pt")
    with torch.no_grad():
      outputs = self.chemberta(**encoded)
      hidden_states = outputs.last_hidden_state[0].mean(axis=0).numpy()
    txt_input = self.feature_transformer_txt.transform(hidden_states.reshape(1, -1))[0]
    return torch.tensor(txt_input, dtype=torch.float32)

  def _batch_encode_plus(
    self,
    batch_smiles: list[str],
    task: str = 'classification',
    return_tensors: str = "pt",
    **kwargs
  ):
    if self.task is None or self.task != task:
      if task == 'classification':
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_cls_tabular.joblib"])
        transformer_tab_path = os.path.join(model_dir, "normalize_cls_tabular.joblib")
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_cls_image.joblib"])
        transformer_img_path = os.path.join(model_dir, "normalize_cls_image.joblib")
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_cls_text.joblib"])
        transformer_txt_path = os.path.join(model_dir, "normalize_cls_text.joblib")
        self.task = task

      elif task == 'regression':
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_reg_tabular.joblib"])
        transformer_tab_path = os.path.join(model_dir, "normalize_reg_tabular.joblib")
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_reg_image.joblib"])
        transformer_img_path = os.path.join(model_dir, "normalize_reg_image.joblib")
        model_dir = snapshot_download("SaeedLab/TITAN-BBB", allow_patterns=["normalize_reg_text.joblib"])
        transformer_txt_path = os.path.join(model_dir, "normalize_reg_text.joblib")
        self.task = task

      else:
        raise ValueError('task not defined')
        return

      self.feature_transformer_tab = joblib.load(transformer_tab_path)
      self.feature_transformer_img = joblib.load(transformer_img_path)
      self.feature_transformer_txt = joblib.load(transformer_txt_path)
    
    data_list = []
    tab, img, txt = [], [], []

    for smiles in batch_smiles:
      tab.append(self.generate_tab_features(smiles))
      img.append(self.generate_img_features(smiles))
      txt.append(self.generate_txt_features(smiles))

    tab = torch.stack(tab)
    img = torch.stack(img)
    txt = torch.stack(txt)

    output = {}
    output["tab"] = tab
    output["img"] = img
    output["txt"] = txt
 
    return BatchEncoding(output, tensor_type=return_tensors)

  def encode(self, 
             batch_smiles: list[str],
             task: str = 'classification',
             return_tensors: str = "pt",
             **kwargs):
    return self._batch_encode_plus(batch_smiles, task, return_tensors, **kwargs)

  def __call__(self, 
               batch_smiles: list[str], 
               task: str = 'classification',
               return_tensors: str = "pt", 
               **kwargs):
    return self._batch_encode_plus(batch_smiles, task, return_tensors, **kwargs)
  
  def _tokenize(self, text, **kwargs):
    return []

  def save_vocabulary(self, save_directory, filename_prefix=None):
    return ()
    
  def get_vocab(self):
    return {"<pad>":0, "<bos>":1, "<eos>":2, "<unk>":3, "<mask>":4}

  @property
  def vocab_size(self):
    return len(self.get_vocab())