Abstract
Identifying essential human microRNAs (miRNAs) remains challenging due to scarce experimental labels. Traditional domain adaptation methods for cross-species knowledge transfer have significant limitations in capturing intricate data features and handling high-dimensional input data. To address these limitations, we propose DeepHEM, a deep domain-adversarial learning framework for predicting human miRNA essentiality. During training stage, a multi-modal feature extractor is designed to capture comprehensive miRNA representations from sequence data, inherent properties, and miRNA-target gene interactions. A multi-loss optimization module is then employed to align feature distributions across two domains, learn domain-invariant features, and ensure robust classification performance. For prediction, the trained feature extractor processes target domain data to generate feature representations, which are then classified by a label predictor. DeepHEM's innovations lie in comprehensive multi-modal feature extraction and reinforced cross-domain feature alignment. Biological correlation analysis reveals significant correlation and consistent trends between predicted essentiality scores and biological indicators. Comparative evaluations against three domain adaptation methods, along with correlation analysis involving biological indicators, demonstrate DeepHEM's superiority. Ablation studies confirm the effectiveness of each individual component. Additionally, a case study confirms 8 of the top 10 predictions with experimental literature. In summary, DeepHEM provides an effective cross-species transfer framework for predicting human miRNA essentiality.