[Briefings in Bioinformatics]* We propose a novel FusionGDA model, which uses a contrastive learning–based training strategy together with a fusion module to enrich the gene and disease semantic representations encoded by pre-trained language models.