MicroRNAs (miRNAs) are crucial noncoding RNA molecules that play a key role in the progression of many diseases. Gaining knowledge of miRNA-disease associations (MDAs) helps uncover the underlying mechanisms of diseases. While traditional lab experiments are expensive and slow, computational models now enable large-scale MDA prediction, boosting research efficiency. In this work, we introduce the AMPCL model, which leverages automatically selected metapaths and contrastive learning. The model begins by constructing a heterogeneous network using miRNA similarities, disease similarities, and existing associations. Through automatic metapath selection, graph convolutional networks extract low-dimensional fusion features from nodes. Contrastive learning is then applied to enhance these features, making positive and negative samples more distinguishable. A multilayer perceptron is finally used to predict miRNA-disease scores. Extensive experiments show the model’s superiority over current methods. Ablation studies and visual analyses further validate its effectiveness, offering meaningful insights for health research. Case studies of human diseases showcase the method’s strong predictive capabilities and its effectiveness in real-world use.