DMGAT: predicting ncRNA-drug resistance associations based on diffusion map and heterogeneous graph attention network

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Non-coding RNAs (ncRNAs) are vital in drug resistance and sensitivity, serving as potential biomarkers and therapeutic targets.
• Predicting ncRNA-drug associations is hindered by data imbalance, sparsity, and limitations in capturing sequence information by existing models.

Purpose of the Study:

• To develop a novel deep learning model, DMGAT (diffusion map and heterogeneous graph attention network), for accurate prediction of ncRNA-drug associations.
• To address challenges like dataset imbalance and enhance the capture of local and global sequence information for reliable predictions.

Main Methods:

• DMGAT integrates diffusion maps for sequence embedding, graph convolutional networks for feature extraction, and a heterogeneous graph attention network (GAT) for information fusion.
• The model utilizes word2vec for embedding ncRNA sequences and drug SMILES, and constructs a heterogeneous network using sequence and Gaussian Interaction Profile kernel similarity.
• Dataset imbalance is addressed by incorporating sensitivity associations and using a random forest classifier for negative sample selection.

Main Results:

• DMGAT achieved superior performance in five-fold cross-validation on a curated dataset, outperforming seven state-of-the-art methods.
• The model attained the highest area under the receiver operating characteristic curve (0.8964), area under the precision-recall curve (0.8984), recall (0.9576), and F1-score (0.8285).

Conclusions:

• DMGAT demonstrates significant potential for identifying ncRNA-drug associations, offering a robust approach for biomarker discovery.
• The model's ability to integrate diverse data types and capture complex sequence information enhances prediction reliability.