Investigating genotype-phenotype associations is essential for elucidating disease mechanisms, as it enables the identification of molecular pathways underlying pathological processes and facilitates the development of personalized therapeutic strategies. Therefore, it is urgent to develop efficient computational methods for predicting potential genotype-phenotype associations in order to reduce the cost of biological experiments. Although current prediction methods demonstrate satisfactory performance, their capacity to effectively capture feature-feature and sample-sample relationships remains limited. Additionally, models relying exclusively on single-omics data fail to achieve holistic sample profiling. To fully utilize the advances in omics and achieve a more comprehensive understanding of human diseases, we propose a novel interpretable framework, termed Multi-Omics Graph ATtention network with Feature Fusion (MOGATFF), for phenotype classification and biomarker discovery. Specifically, we first leverage biological networks to reduce the dimensionality of genotypic data. Then we compute associations between features to enable feature fusion and achieve optimization. To better capture complex sample relationships, a feature enhancement module based on graph attention network is employed to aggregate features for phenotype classification. Finally, we utilize the Permutation Method to identify important feature, thereby enhancing interpretability. In the five-fold cross-validation, MOGATFF achieves a mean ACC of 0.9281 on Braak, 0.8613 on Cerad and 0.8716 on Cogdx, outperforming eight baseline approaches. Furthermore, case study provides validation of its efficacy in detecting disease-associated biomarker.