aptaESM_head.py

import torch
import torch.nn as nn
from torch.utils.data import Dataset
from transformers import EsmModel
import pandas as pd

class ProteinDataset(Dataset):
    def __init__(self, data, esm_tokenizer, apta_tokenizer):
        self.data = data
        self.esm_tokenizer = esm_tokenizer
        self.apta_tokenizer = apta_tokenizer

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        protein_seq = self.data.iloc[idx, 0]
        protein_apta_tokens = self.apta_tokenizer(protein_seq, return_tensors="pt", padding=True, truncation=True)
        protein_esm_tokens = self.esm_tokenizer(protein_seq, return_tensors="pt", padding=True, truncation=True)
        return protein_apta_tokens, protein_esm_tokens

class ContrastiveNetwork(nn.Module):
    def __init__(self, apta_encoder, esm_encoder, out_tokens=64, output_dim=128):
        super(ContrastiveNetwork, self).__init__()
        self.apta_encoder = apta_encoder
        self.esm_encoder = esm_encoder
        self.apta_proj = nn.Linear(128, out_tokens)
        self.esm_proj = nn.Linear(1280, out_tokens)
        self.final_proj = nn.Linear(1, output_dim)

    def forward(self, protein_apta_tokens, protein_esm_tokens):
        # Input sizes: protein_apta_tokens: (BS x #aptaTokens), protein_esm_tokens: (BS x #esmTokens)
        protein_apta_embedding = self.apta_encoder(protein_apta_tokens)
        protein_esm_embedding = self.esm_encoder(protein_esm_tokens, repr_layers=[33], return_contacts=False)['representations'][33]
        # Sizes: protein_apta_embedding: (BS x #aptaTokens x apta_embed_dim), protein_esm_embedding: (BS x #esmTokens x esm_embed_dim)

        # Apply mean pooling
        protein_apta_pooled = protein_apta_embedding.mean(dim=1)
        protein_esm_pooled = protein_esm_embedding.mean(dim=1)

        # Project to a common space
        protein_apta_proj = self.apta_proj(protein_apta_pooled)
        protein_esm_proj = self.esm_proj(protein_esm_pooled)

        # Concatenate the projected embeddings
        combined_embedding = torch.cat((protein_apta_proj, protein_esm_proj), dim=1)

        # Unsqueeze to get an embedding
        combined_embedding_reshape = torch.unsqueeze(combined_embedding, -1)

        # Final projection to match the original AptaTrans embedding size
        output_embedding = self.final_proj(combined_embedding_reshape)

        # Output embedding has shape (BS x embedding_dim x output_dim) - can construct interaction map w/ this
        return protein_apta_proj, protein_esm_proj, output_embedding

def load_dataset(file_path):
    data = pd.read_csv(file_path)
    return data

def contrastive_loss(embedding_a, embedding_b, margin=1.0):
    positive_distance = nn.functional.pairwise_distance(embedding_a, embedding_b)
    loss = positive_distance.mean()
    return loss