GATDHAN

Dual hierarchical attention network combined with residual mechanism for predicting circRNA–miRNA associations

Data and code

CMI-9589

9589_circRNA_sequence.csv：This file stores the name and the sequence of circRNAs.

9589_miRNA_sequence.csv: This file stores the name and the sequence of miRNAs.

A_9589.csv: This file stores the association information between circRNAs and miRNAs.

merged_9589_circRNA_3.pkl: This file stores the features of circRNA which can be download from URL: https://github.com/ntitatip/A-Knowledge-Graph-Enhanced-Pre-trained-Large-Language-Model-for-Predicting-MicroRNA-circRNA-Interac.

merged_9589_micRNA_3.pkl: This file stores the features of miRNA which can be download from URL: https://github.com/ntitatip/A-Knowledge-Graph-Enhanced-Pre-trained-Large-Language-Model-for-Predicting-MicroRNA-circRNA-Interac.

CMI-9905

circSequence_9905.csv： This file stores the name and the sequence of circRNAs.

miSequence_9905.csv: This file stores the name and the sequence of miRNAs.

A.csv: This file stores the association information between circRNAs and drugs.

merged_9905_circRNA_3.pkl: This file stores the features of circRNA which can be download from URL: https://github.com/ntitatip/A-Knowledge-Graph-Enhanced-Pre-trained-Large-Language-Model-for-Predicting-MicroRNA-circRNA-Interac.

merged_9905_micRNA_3.pkl: This file stores the features of miRNA which can be download from URL: https://github.com/ntitatip/A-Knowledge-Graph-Enhanced-Pre-trained-Large-Language-Model-for-Predicting-MicroRNA-circRNA-Interac.

GATDHAN_CMI_9905.ipynb: GATDHAN is run on Jupyter Notebook, and all the code for CMI-9905 dataset is contained within this file.

GATDHAN_CMI_9589.ipynb: GATDHAN is run on Jupyter Notebook, and all the code for CMI-9589 dataset is contained within this file.

Environment

GATDHAN is implemented to work under Python 3.9 and Jupyter Notebook
torch >=2.5.0
numpy >=1.22.0
pandas >=1.5.0
torch_geometric >=2.5.0

Run steps

GATDHAN_CMI_9905.ipynb and GATDHAN_CMI_9589.ipynb can be upload into Jupyter Notebook and select ‘run all’ to train the model and obtain prediction scores for circRNA-miRNA associations.

Function descriptions for in files GATDHAN_CMI_9905.ipynb and GATDHAN_CMI_9589.ipynb

Integration of feature

def get_syn_sim(A, seq_sim, str_sim, mode)
PARAMETERS:
A(numpy matrix): The adjacency matrix of circRNAs between miRNAs.
seq_sim(numpy matrix): The similarity matrix of circRNAs.
str_sim(numpy matrix): The similarity matrix of miRNAs.
mode: 0 = GIP kernel sim.
return the new similarity of circRNAs and miRNAs.

Get the index and type of the edges

def get_edge_index(matrix, new_A, threshold)
PARAMETERS:
matrix(numpy matrix): The similarity matrix of circRNAs or miRNAs.
new_A(numpy matrix): The adjacency matrix of circRNAs between miRNAs.
threshold (int): Threshold for distinguishing between head and tail node. Here, we do not distinguish between the head and tail nodes.

Dual attention layer

class DGATConv (self, in_hid, out_hid, num_edge_types, negative_slope=0.2, dual=True, heads=1, mask=None, global_weight=True)
PARAMETERS:
in_hid (int): Dimension of input features.
out_hid (int): Dimension of output features.
num_edge_types(int): Number of edge types.
negative_slope(float, optional): Controls the angle of the negative slope (which is used for negative input values). (default: 0.2)
dual(bool, optional): Controls whether dual attention mechanisms are used. (default: True)
heads(int, optional): Number of multi-head-attentions. (default: 1)
mask(int, optional): Feature extraction by specific types of edges. (default: None)
global_weight (bool, optional): Controls the use of global weights. (default: True)

Hierarchical attention layer

class HetGATConv (self, in_hid, out_hid, negative_slope=0.2, norm=True, dual=True, global_weight=True)
PARAMETERS:
in_hid (int): Dimension of input features.
out_hid (int): Dimension of output features.
norm(bool, optional): Parameter reservation. (default: True)
dual(bool, optional): Controls whether dual attention mechanisms are used. (default: True)
global_weight (bool, optional): Controls the use of global weights. (default: True)

GAT layer

class GATConv(MessagePassing): PARAMETERS: in_channels (int): Dimension of input features out_channels (int): Dimension of output features heads (int): Number of heads of multi-attention mechanisms

Our model

class GATDHAN(self,n_hid_layers: int, hid_features: list, hid_features2: list, n_heads: list, n_dis: int, n_pi: int, dropout: float = 0.6, n_hid_layers2: int = 1, dual: bool = False)
PARAMETERS: n_hid_layers (int): The number of GAT layer. hid_features (list): Dimension of circRNA features.
hid_features2 (list): Dimension of miRNA features.
n_heads (int): Number of heads of multi-attention mechanisms
n_dis (int): Number of circRNAs.
n_pi (int): Number of miRNAs.
n_hid_layers2 (int): Number of Dual attention layers. (default: 1)
dropout(float, optional): Dropout probability of the normalized attention coefficients which exposes each node to a stochastically sampled neighborhood during raining. (default: 0.6)
dual (bool, optional): If “True” then the dual mechanism is used in the DHAN layer. (default: False)