<p>Cumulative studies have revealed that RNA N6-methyladenosine (m⁶A) modification plays a critical role in various human complex diseases including cancers. Exploring the relationship between m⁶As and diseases can reveal the molecular mechanism of disease pathogenesis. However, the known associations between m⁶As and diseases are very rare due to the time-consuming and laborious biological experiments. Some computational approaches have been proposed to predict the potential m⁶A-disease associations. However, these methods are restricted to specific application scenarios. To solve this limitation, we propose a novel computational framework, named m⁶A-Disease Prediction using Graph Convolutional Networks and Positive-Unlabeled Learning with Self-Adaptive Sampling (m⁶ADP-GCNPUAS), to accurately predict m⁶A-disease associations. m⁶ADP-GCNPUAS effectively captures the embedded features of m⁶A sites or diseases using GCN. Given the very limited m⁶A-disease associations, m⁶ADP-GCNPUAS adopts the PUAS framework to augment the potential positive samples, which improves the predictive power for m⁶A-disease associations. The experimental results indicate that m⁶ADP-GCNPUAS achieves over 0.99 across six evaluation metrics and PUAS significantly improves the predicted performance. In addition, case studies demonstrate that the predicted cancer-associated m⁶A sites participate in the regulation of cancer development and some predicted m⁶A genes could be the hallmarks for different cancers.</p> Graphical Abstract <p></p>