NNSFMDA: Lightweight Transformer Model with Bounded Nuclear Norm Minimization for Microbe-Drug Association Prediction

Abstract

Identifying potential connections between microbe-drug pairs play an important role in drug discovery and clinical treatment. Techniques like graph neural networks effectively derive accurate node representations from sparse topologies,however, they struggle with over-smoothing and over-compression, and their interpretability is relatively poor. Conversely, mathematical methods with low-rank approximations are interpretable but often get trapped in local optima. To address these issues, we propose a new prediction model named NNSFMDA, in which, the bounded nuclear norm minimization and the simplified transformer were combined to infer possible drug-microbe associations. In NNSFMDA, we first constructed a heterogeneous microbe-drug network by integrating multiple microbe and drug similarity metrics, according to which, we subsequently transformed the prediction problem to a matrix filling problem, and then, iteratively approximated the matrix by minimizing the number of bounded nuclear norm. Finally, based on the newly-filled matrix, we introduced a simplified transformer to estimate possible scores of microbe-drug pairs. Results showed that NNSFMDA could achieve reliable AUC value of 0.98, which outperformed existing state-of-the-art competitive methods. In the experimental section, ablation experiments and modular analyses further demonstrate the superiority of the model, and case studies of microbe-drug associations confirm the validity of the model. These tests have all highlighted the potential of the NNSFMDA to predict latent microbe-drug associations in the future.