EFM²BF: Extraction and fusion of multi-network multi-scale bio-topological features for gene prediction in brain...

The identification of brain disease genes holds significant importance in revealing brain disorder mechanisms and facilitating drug development. To enhance that identification of genes associate with brain diseases, most methods fused various biological network features to boost the performance of downstream task. However, existing methods are often faced with some challenges, as they limited feature extraction and fusion capabilities. Hence, this paper introduces a framework named EFM²BF for predicting brain disease genes. Firstly, we integrate the data of PPI networks and R-fMRI as inputs. Besides, to better facilitate feature extraction and remedy the lost information, we propose a novel approach that combines the RWR algorithm and dual-channel GCN with skip connections for multi-network multi-scale feature extraction. This strategy enriches feature representations. Following this,addressing information loss and suboptimal fusion in current models, we present an enhanced adaptive semi-supervised autoencoder (SSAE). This approach handles network features of varying scales and achieves complementary fusion of key-vectors from each network through the imposition of joint constraints. Lastly,the particle swarm optimization (PSO) algorithm is employed to optimize the SVM model, known as PSO-SVM, enhancing efficiency and global optimization ability. The fused global features are used for predicting disease genes related to Parkinson's disease and major depressive disorder.