iHofman: a predictive model integrating high-order and low-order features with weighted attention mechanisms for circRNA-miRNA interactions

Background: Increasing research indicates that the complex interactions between circular RNAs (circRNAs) and microRNAs (miRNAs) are critical for diagnosing and treating various human diseases. Consequently, accurately predicting potential circRNA-miRNA interactions (CMIs) has become increasingly important and urgent. Traditional biological experiments, however, are often labor-intensive, time-consuming, and prone to external influences. Results: To tackle this challenge, we present a novel model, iHofman, designed to predict CMIs by integrating high-order and low-order features with weighted attention mechanisms. Specifically, we first extract sequence and structural information representations using FastText and GraRep, respectively, and capture high-order and low-order features from sequence information representations using stacked autoencoders. Subsequently, weighted attention mechanisms are applied for feature fusion, focusing on the most relevant information. Finally, multi-layer perceptron is employed to accurately infer potential CMIs. In the fivefold cross-validation (CV) experiment on the baseline dataset, iHofman achieved an accuracy of 82.49% with an AUC of 0.9092. iHofman also demonstrates solid performance on other CMI datasets. In case studies, 26 of the top 30 CMIs with the highest iHofman predictive scores were confirmed in relevant literature. Conclusions: The above experimental results indicate that iHofman can effectively predict potential CMIs and has achieved outstanding performance compared with existing methods. It provides a reliable supplementary approach for subsequent biological wet experiments.