Timely-MDA: A Benchmark for Generalizable MiRNA-Disease Association Prediction

The identification of miRNA-disease associations (MDAs) holds significant value in the field of disease diagnosis and treatment. Recently, computational prediction methods have been increasingly proposed to detect potential MDAs, so as to assist experimental verifications. Despite the success of deep learning models, studies in the MDA prediction are still limited by the datasets and the evaluation framework employed. Concretely, existing datasets comprise only hundreds of diseases, and the random-split-based evaluation framework provides an overly optimistic estimate of the performance of MDA prediction methods. In this study, we propose a novel benchmark, Timely-MDA, for generalizable MDA prediction. First, we construct a comprehensive dataset comprising a broad scope of miRNA and disease entities and diverse semantic features. Second, four existing MDA prediction methods are implemented, and a new baseline is proposed based on our dataset. Third, the performance of these methods is analyzed using our timely-split evaluation framework. Overall, Timely-MDA provides a robust data foundation for MDA modeling, and enables quantitative estimation of the generalization ability of MDA prediction methods. Data and code are available at https://github.com/EchoChou990919/Timely-MDA.