MambaCAttnGCN+: a comprehensive framework integrating MambaTextCNN, cross-attention and graph convolution network for...

Area of Science:

• Biomedical informatics
• Computational biology
• Genomics

Background:

• Understanding piwi-interacting RNA (piRNA) and disease interactions is vital for diagnostics and therapeutics.
• Existing computational methods struggle with sparse data, limiting accurate piRNA-disease association prediction.

Purpose of the Study:

• To develop a more accurate computational model for predicting piRNA-disease associations.
• To leverage heterogeneous graph construction and advanced sequence embedding for improved feature extraction.

Main Methods:

• Constructed a heterogeneous graph integrating piRNA sequences, disease semantics, and known associations.
• Employed MambaTextCNN for piRNA sequence feature extraction and heterogeneous graph convolution with cross-attention.
• Utilized positive unlabeled learning to develop the MambaCAttnGCN+ prediction model.

Main Results:

• MambaCAttnGCN+ achieved high AUCs of 0.94 and 0.953 in 5-fold cross-validation on two datasets.
• The model outperformed seven other state-of-the-art methods in predicting piRNA-disease associations.
• Ablation studies confirmed MambaTextCNN's superior performance in extracting sequence node features.

Conclusions:

• MambaCAttnGCN+ demonstrates significant potential as a predictive tool for piRNA-disease associations.
• The findings highlight the effectiveness of integrating sequence information, graph networks, and advanced learning techniques.
• This approach advances the study of piRNA roles in various diseases.