REDDA/model.py

"""
Module: model.py
Description:
    This module defines the neural network architecture for the REDDA framework used in drug repositioning.
    It includes several components such as a link prediction decoder, heterogeneous graph convolutional layers,
    semantic attention mechanisms, and graph attention modules. Adapted for the new knowledge graph dataset.
"""

import torch
import torch.nn as nn
import dgl.nn as dglnn
import dgl


class InnerProductDecoder(nn.Module):
    """
    Decoder layer that computes the inner product between drug and disease embeddings for link prediction.
    Optionally applies a linear transformation to the disease embeddings.
    """

    def __init__(self, input_dim=None, dropout=0.4):
        """
        Parameters:
            input_dim (int, optional): Dimension of input features. If provided, a linear layer is used.
            dropout (float): Dropout rate.
        """
        super(InnerProductDecoder, self).__init__()
        self.dropout = nn.Dropout(dropout)
        if input_dim:
            self.weights = nn.Linear(input_dim, input_dim, bias=False)
            nn.init.xavier_uniform_(self.weights.weight)

    def forward(self, feature):
        """
        Forward pass to compute the similarity score matrix.

        Parameters:
            feature (dict): Dictionary containing node features for 'drug' and 'disease'.

        Returns:
            torch.Tensor: The computed score matrix with shape [num_drugs, num_diseases].
        """
        # Apply dropout to drug and disease features
        feature["drug"] = self.dropout(feature["drug"])
        feature["disease"] = self.dropout(feature["disease"])
        R = feature["drug"]
        D = self.weights(feature["disease"]) if hasattr(self, "weights") else feature["disease"]
        outputs = R @ D.T
        return outputs


class Node_Embedding(nn.Module):
    """
    Basic heterogeneous graph convolution layer to generate node embeddings for each relation type.
    """

    def __init__(self, in_feats, out_feats, dropout, rel_names):
        """
        Parameters:
            in_feats (int): Input feature dimension.
            out_feats (int): Output feature dimension.
            dropout (float): Dropout rate.
            rel_names (list): List of relation names to construct individual GraphConv layers.
        """
        super().__init__()
        # Create a dictionary of GraphConv layers for each relation
        hetero_conv_dict = {}
        for rel in rel_names:
            graphconv = dglnn.GraphConv(in_feats, out_feats)
            nn.init.xavier_normal_(graphconv.weight)
            hetero_conv_dict[rel] = graphconv
        self.dropout = nn.Dropout(p=dropout)
        self.embedding = dglnn.HeteroGraphConv(hetero_conv_dict, aggregate="sum")
        self.bn_layer = nn.BatchNorm1d(out_feats)
        self.prelu = nn.PReLU()

    def forward(self, graph, inputs, bn=False, dp=False):
        """
        Forward pass for the node embedding.

        Parameters:
            graph (dgl.DGLHeteroGraph): Input heterogeneous graph.
            inputs (dict): Dictionary of input node features keyed by node type.
            bn (bool): Whether to apply batch normalization.
            dp (bool): Whether to apply dropout.

        Returns:
            dict: Updated node features keyed by node type.
        """
        h = self.embedding(graph, inputs)
        # Apply batch normalization and dropout if specified
        if bn and dp:
            h = {k: self.prelu(self.dropout(self.bn_layer(v))) for k, v in h.items()}
        elif dp:
            h = {k: self.prelu(self.dropout(v)) for k, v in h.items()}
        elif bn:
            h = {k: self.prelu(self.bn_layer(v)) for k, v in h.items()}
        else:
            h = {k: self.prelu(v) for k, v in h.items()}
        return h


class SemanticAttention(nn.Module):
    """
    Attention mechanism to aggregate multiple embeddings.
    """

    def __init__(self, in_feats, hidden_size=128):
        """
        Parameters:
            in_feats (int): Dimension of input features.
            hidden_size (int): Hidden layer size for attention projection.
        """
        super(SemanticAttention, self).__init__()
        self.project = nn.Sequential(
            nn.Linear(in_feats, hidden_size),
            nn.Tanh(),
            nn.Linear(hidden_size, 1, bias=False),
        )

    def forward(self, z, is_print=False):
        """
        Apply semantic attention to a stack of embeddings.

        Parameters:
            z (torch.Tensor): Tensor of shape [batch, num_layers, in_feats].
            is_print (bool): If True, prints the attention weights.

        Returns:
            torch.Tensor: Aggregated embedding of shape [batch, in_feats].
        """
        w = self.project(z).mean(0)  # Average over the batch dimension
        beta = torch.softmax(w, dim=0)
        beta = beta.expand((z.shape[0],) + beta.shape)
        if is_print:
            print(beta)
        return (beta * z).sum(1)


class SubnetworkEncoder(nn.Module):
    """
    Module to compute topological subnetwork embeddings using multiple heterogeneous graph convolution blocks.
    Adapted for the new knowledge graph with multiple node types.
    """

    def __init__(self, ntypes, in_feats, out_feats, dropout):
        """
        Parameters:
            ntypes (list): List of node types present in the graph.
            in_feats (int): Input feature dimension.
            out_feats (int): Output feature dimension.
            dropout (float): Dropout rate.
        """
        super(SubnetworkEncoder, self).__init__()
        self.ntypes = ntypes

        # Define subnetworks based on the new knowledge graph structure
        # Drug-centric subnetworks
        self.drug_disease = Node_Embedding(in_feats, out_feats, dropout,
                                           ['drug_drug', 'drug_disease_indication', 'disease_disease'])
        
        self.drug_protein = Node_Embedding(in_feats, out_feats, dropout,
                                           ['drug_drug', 'drug_protein', 'protein_protein'])
        
        self.drug_effect = Node_Embedding(in_feats, out_feats, dropout,
                                          ['drug_drug', 'drug_effect'])
        
        # Protein-centric subnetworks
        self.protein_bioprocess = Node_Embedding(in_feats, out_feats, dropout,
                                                 ['protein_protein', 'protein_bioprocess', 'bioprocess_bioprocess'])
        
        self.protein_cellcomp = Node_Embedding(in_feats, out_feats, dropout,
                                               ['protein_protein', 'protein_cellcomp', 'cellcomp_cellcomp'])
        
        self.protein_molfunc = Node_Embedding(in_feats, out_feats, dropout,
                                              ['protein_protein', 'protein_molfunc', 'molfunc_molfunc'])
        
        self.protein_pathway = Node_Embedding(in_feats, out_feats, dropout,
                                              ['protein_protein', 'protein_pathway', 'pathway_pathway'])
        
        self.protein_disease = Node_Embedding(in_feats, out_feats, dropout,
                                              ['protein_protein', 'protein_disease', 'disease_disease'])
        
        # Disease-centric subnetworks
        self.disease_phenotype = Node_Embedding(in_feats, out_feats, dropout,
                                                ['disease_disease', 'disease_phenotype_positive', 'phenotype_phenotype'])
        
        self.disease_exposure = Node_Embedding(in_feats, out_feats, dropout,
                                               ['disease_disease', 'disease_exposure'])
        
        # Biological process subnetworks
        self.bioprocess_network = Node_Embedding(in_feats, out_feats, dropout,
                                                 ['bioprocess_bioprocess'])
        
        self.cellcomp_network = Node_Embedding(in_feats, out_feats, dropout,
                                               ['cellcomp_cellcomp'])
        
        self.molfunc_network = Node_Embedding(in_feats, out_feats, dropout,
                                              ['molfunc_molfunc'])
        
        self.pathway_network = Node_Embedding(in_feats, out_feats, dropout,
                                              ['pathway_pathway'])
        
        self.phenotype_network = Node_Embedding(in_feats, out_feats, dropout,
                                                ['phenotype_phenotype'])
        
        self.semantic_attention = SemanticAttention(in_feats=out_feats)

    def forward(self, g, h, bn=False, dp=False):
        """
        Compute subnetwork embeddings and aggregate them using semantic attention.

        Parameters:
            g (dgl.DGLHeteroGraph): Input heterogeneous graph.
            h (dict): Dictionary of node features.
            bn (bool): Whether to apply batch normalization.
            dp (bool): Whether to apply dropout.

        Returns:
            dict: Updated node features for each node type.
        """
        new_h = {ntype: [] for ntype in self.ntypes}

        # Drug-Disease subnet
        if 'drug' in self.ntypes and 'disease' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['drug_drug', 'drug_disease_indication', 'disease_disease'])
                h_sub = self.drug_disease(subgraph, {'drug': h['drug'], 'disease': h['disease']}, bn, dp)
                new_h['drug'].append(h_sub['drug'])
                new_h['disease'].append(h_sub['disease'])
            except:
                pass

        # Drug-Protein subnet
        if 'drug' in self.ntypes and 'protein' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['drug_drug', 'drug_protein', 'protein_protein'])
                h_sub = self.drug_protein(subgraph, {'drug': h['drug'], 'protein': h['protein']}, bn, dp)
                new_h['drug'].append(h_sub['drug'])
                new_h['protein'].append(h_sub['protein'])
            except:
                pass

        # Drug-Effect subnet
        if 'drug' in self.ntypes and 'effect' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['drug_drug', 'drug_effect'])
                h_sub = self.drug_effect(subgraph, {'drug': h['drug'], 'effect': h['effect']}, bn, dp)
                new_h['drug'].append(h_sub['drug'])
                new_h['effect'].append(h_sub['effect'])
            except:
                pass

        # Protein-Bioprocess subnet
        if 'protein' in self.ntypes and 'bioprocess' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['protein_protein', 'protein_bioprocess', 'bioprocess_bioprocess'])
                h_sub = self.protein_bioprocess(subgraph, {'protein': h['protein'], 'bioprocess': h['bioprocess']}, bn, dp)
                new_h['protein'].append(h_sub['protein'])
                new_h['bioprocess'].append(h_sub['bioprocess'])
            except:
                pass

        # Protein-Cellcomp subnet
        if 'protein' in self.ntypes and 'cellcomp' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['protein_protein', 'protein_cellcomp', 'cellcomp_cellcomp'])
                h_sub = self.protein_cellcomp(subgraph, {'protein': h['protein'], 'cellcomp': h['cellcomp']}, bn, dp)
                new_h['protein'].append(h_sub['protein'])
                new_h['cellcomp'].append(h_sub['cellcomp'])
            except:
                pass

        # Protein-Molfunc subnet
        if 'protein' in self.ntypes and 'molfunc' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['protein_protein', 'protein_molfunc', 'molfunc_molfunc'])
                h_sub = self.protein_molfunc(subgraph, {'protein': h['protein'], 'molfunc': h['molfunc']}, bn, dp)
                new_h['protein'].append(h_sub['protein'])
                new_h['molfunc'].append(h_sub['molfunc'])
            except:
                pass

        # Protein-Pathway subnet
        if 'protein' in self.ntypes and 'pathway' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['protein_protein', 'protein_pathway', 'pathway_pathway'])
                h_sub = self.protein_pathway(subgraph, {'protein': h['protein'], 'pathway': h['pathway']}, bn, dp)
                new_h['protein'].append(h_sub['protein'])
                new_h['pathway'].append(h_sub['pathway'])
            except:
                pass

        # Protein-Disease subnet
        if 'protein' in self.ntypes and 'disease' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['protein_protein', 'protein_disease', 'disease_disease'])
                h_sub = self.protein_disease(subgraph, {'protein': h['protein'], 'disease': h['disease']}, bn, dp)
                new_h['protein'].append(h_sub['protein'])
                new_h['disease'].append(h_sub['disease'])
            except:
                pass

        # Disease-Phenotype subnet
        if 'disease' in self.ntypes and 'phenotype' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['disease_disease', 'disease_phenotype_positive', 'phenotype_phenotype'])
                h_sub = self.disease_phenotype(subgraph, {'disease': h['disease'], 'phenotype': h['phenotype']}, bn, dp)
                new_h['disease'].append(h_sub['disease'])
                new_h['phenotype'].append(h_sub['phenotype'])
            except:
                pass

        # Disease-Exposure subnet
        if 'disease' in self.ntypes and 'exposure' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['disease_disease', 'disease_exposure'])
                h_sub = self.disease_exposure(subgraph, {'disease': h['disease'], 'exposure': h['exposure']}, bn, dp)
                new_h['disease'].append(h_sub['disease'])
                new_h['exposure'].append(h_sub['exposure'])
            except:
                pass

        # Self-loop subnetworks for biological entities
        if 'bioprocess' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['bioprocess_bioprocess'])
                h_sub = self.bioprocess_network(subgraph, {'bioprocess': h['bioprocess']}, bn, dp)
                new_h['bioprocess'].append(h_sub['bioprocess'])
            except:
                pass

        if 'cellcomp' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['cellcomp_cellcomp'])
                h_sub = self.cellcomp_network(subgraph, {'cellcomp': h['cellcomp']}, bn, dp)
                new_h['cellcomp'].append(h_sub['cellcomp'])
            except:
                pass

        if 'molfunc' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['molfunc_molfunc'])
                h_sub = self.molfunc_network(subgraph, {'molfunc': h['molfunc']}, bn, dp)
                new_h['molfunc'].append(h_sub['molfunc'])
            except:
                pass

        if 'pathway' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['pathway_pathway'])
                h_sub = self.pathway_network(subgraph, {'pathway': h['pathway']}, bn, dp)
                new_h['pathway'].append(h_sub['pathway'])
            except:
                pass

        if 'phenotype' in self.ntypes:
            try:
                subgraph = g.edge_type_subgraph(['phenotype_phenotype'])
                h_sub = self.phenotype_network(subgraph, {'phenotype': h['phenotype']}, bn, dp)
                new_h['phenotype'].append(h_sub['phenotype'])
            except:
                pass

        # Aggregate multiple embeddings with semantic attention
        for ntype in self.ntypes:
            if new_h[ntype]:  # Only process if there are embeddings
                h[ntype] = torch.stack(new_h[ntype], dim=1)
                h[ntype] = self.semantic_attention(h[ntype])
        return h


class Graph_attention(nn.Module):
    """
    Multi-omics graph attention block that aggregates information from all node types.
    """

    def __init__(self, in_feats, out_feats, num_heads, dropout):
        """
        Parameters:
            in_feats (int): Input feature dimension.
            out_feats (int): Output feature dimension.
            num_heads (int): Number of attention heads.
            dropout (float): Dropout rate.
        """
        super().__init__()
        self.gat = dglnn.GATConv(in_feats, out_feats, num_heads,
                                 dropout, dropout,
                                 activation=nn.PReLU(),
                                 allow_zero_in_degree=True)
        self.gat.reset_parameters()
        self.linear = nn.Linear(in_feats * num_heads, out_feats)
        self.prelu = nn.PReLU()
        self.bn_layer = nn.BatchNorm1d(out_feats)

    def forward(self, graph, inputs, bn=False):
        """
        Forward pass for graph attention.

        Parameters:
            graph (dgl.DGLHeteroGraph): Input heterogeneous graph.
            inputs (dict): Dictionary of node features.
            bn (bool): Whether to apply batch normalization.

        Returns:
            tuple: Aggregated disease and drug embeddings.
        """
        # Convert heterogeneous graph to homogeneous
        num_dis = graph.num_nodes("disease")
        num_drug = graph.num_nodes("drug")
        new_g = dgl.to_homogeneous(graph)
        new_h = torch.cat([feat for feat in inputs.values()], dim=0)
        new_h = self.gat(new_g, new_h)
        new_h = self.prelu(torch.mean(new_h, dim=1))
        if bn:
            return self.bn_layer(new_h[:num_dis]), self.bn_layer(new_h[num_dis:num_drug + num_dis])
        return new_h[:num_dis], new_h[num_dis:num_drug + num_dis]


class Model(nn.Module):
    """
    Overall REDDA architecture for drug repositioning, adapted for the new knowledge graph.
    """

    def __init__(self, etypes, ntypes, in_feats, hidden_feats, num_heads, dropout):
        """
        Parameters:
            etypes (list): List of edge types.
            ntypes (list): List of node types.
            in_feats (int): Input feature dimension.
            hidden_feats (int): Hidden layer dimension.
            num_heads (int): Number of attention heads.
            dropout (float): Dropout rate.
        """
        super(Model, self).__init__()
        self.ntypes = ntypes

        # Linear projection layers for each node type
        self.node_projections = nn.ModuleDict()
        for ntype in ntypes:
            self.node_projections[ntype] = nn.Linear(in_feats, hidden_feats)
            nn.init.xavier_normal_(self.node_projections[ntype].weight)

        # Two layers of node embedding generation
        self.feat_generate_layer1 = Node_Embedding(hidden_feats, hidden_feats, dropout, etypes)
        self.feat_generate_layer2 = Node_Embedding(hidden_feats, hidden_feats, dropout, etypes)
        # Subnetwork encoder to capture local topological features
        self.subnet_layer = SubnetworkEncoder(ntypes, hidden_feats, hidden_feats, dropout)
        # Graph attention layer to capture global information
        self.totalnet_layer = Graph_attention(hidden_feats, hidden_feats, num_heads, dropout)
        # Layer-level semantic attention for drug and disease embeddings
        self.layer_attention_layer_drug = SemanticAttention(hidden_feats)
        self.layer_attention_layer_dis = SemanticAttention(hidden_feats)
        # Final prediction decoder using inner product
        self.predict = InnerProductDecoder(hidden_feats)

    def forward(self, g, x):
        """
        Forward pass of the REDDA model.

        Parameters:
            g (dgl.DGLHeteroGraph): Input heterogeneous graph.
            x (dict): Dictionary of initial node features for each node type.

        Returns:
            torch.Tensor: Predicted score matrix between drugs and diseases.
        """
        drug_emb_list, dis_emb_list = [], []
        h = {ntype: x[ntype] for ntype in self.ntypes}

        # Apply linear transformation to initial features
        for ntype in self.ntypes:
            h[ntype] = self.node_projections[ntype](h[ntype])

        drug_emb_list.append(h["drug"])
        dis_emb_list.append(h["disease"])

        # Generate updated embeddings via heterogeneous GCN layers
        h = self.feat_generate_layer1(g, h, bn=True, dp=True)
        h = self.feat_generate_layer2(g, h, bn=True, dp=True)
        drug_emb_list.append(h["drug"])
        dis_emb_list.append(h["disease"])

        # Obtain subnetwork embeddings
        h = self.subnet_layer(g, h, bn=False, dp=True)
        drug_emb_list.append(h["drug"])
        dis_emb_list.append(h["disease"])

        # Apply graph attention for global aggregation
        h["disease"], h["drug"] = self.totalnet_layer(g, h, bn=False)
        drug_emb_list.append(h["drug"])
        dis_emb_list.append(h["disease"])

        # Apply layer-level semantic attention to fuse embeddings from different layers
        h["drug"] = self.layer_attention_layer_drug(torch.stack(drug_emb_list, dim=1))
        h["disease"] = self.layer_attention_layer_dis(torch.stack(dis_emb_list, dim=1))

        return self.predict(h)