deepscreen/models/predictors/graph_dta.py

import torch
import torch.nn as nn

from lightning import LightningModule


class GraphDTA(LightningModule):
    """
    From GraphDTA (Nguyen et al., 2020; https://doi.org/10.1093/bioinformatics/btaa921).
    """
    def __init__(
            self,
            gnn: nn.Module,
            num_features_protein: int,
            n_filters: int,
            embed_dim: int,
            output_dim: int,
            dropout: float
    ):
        super().__init__()
        self.gnn = gnn

        # protein sequence encoder (1d conv)
        self.embedding_xt = nn.Embedding(num_features_protein, embed_dim)
        self.conv_xt = nn.LazyConv1d(out_channels=n_filters, kernel_size=8)
        self.fc1_xt = nn.Linear(32 * 121, output_dim)

        # combined layers
        self.fc1 = nn.Linear(256, 1024)
        self.fc2 = nn.Linear(1024, 512)

        # activation and regularization
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(dropout)

    # protein input feedforward
    def conv_forward_xt(self, v_p):
        v_p = self.embedding_xt(v_p.long())
        v_p = self.conv_xt(v_p)
        # flatten
        v_p = v_p.view(-1, 32 * 121)
        v_p = self.fc1_xt(v_p)
        return v_p

    def forward(self, v_d, v_p):
        v_d = self.gnn(v_d)
        v_p = self.conv_forward_xt(v_p)

        # concat
        v_f = torch.cat((v_d, v_p), 1)
        # dense layers
        v_f = self.fc1(v_f)
        v_f = self.relu(v_f)
        v_f = self.dropout(v_f)
        v_f = self.fc2(v_f)
        v_f = self.relu(v_f)
        v_f = self.dropout(v_f)
        # v_f = self.out(v_f)
        return v_f