GNN-NPI: A Novel Parameter Initialization Method of Graph Convolutional Neural Network for Anti-cancer Peptide Recognition

Researchers have found that initialization methods can accelerate convergence speed and reduce training loss in graph convolutional neural networks for identifying anticancer peptides (ACPs). However, existing initialization methods for odd activation functions lack a reasonable mathematical explanation. Meanwhile, some initialization methods require the activation function's derivative to be equal to 1 near the origin. In this study, we proposed a novel initialization method (NPI) with theoretical derivation to explain the stability of an odd activation function. Based on the defined non-odd activation function KPReLU, we proved through mathematical derivation that NPI maintains the stability of the mean and variance of weight matrices during forward and backward propagation in graph neural networks, even when the derivative of the activation function near the origin deviates from 1. By integrating NPI, we developed a model named GNN-NPI, which uses multi-layered structures to extract and integrate information at different levels while considering both sequential and graph features. Experimental results on the benchmark dataset show that GNN-NPI outperforms other algorithms with SN 94.87%, SP 92.65%, AVE 93.76%, ACC 93.83%, and MCC 0.88. Moreover, GNN-NPI significantly surpasses state-of-the-art models, with improvements of 11.49%, 2.56%, 2.05%, and 0.06 in SN, AVE, ACC, and MCC, respectively, on the independent dataset. GNN-NPI enhances training efficiency and generalization. Feature selection and architectural optimization ensure robust discriminative capability. The superior performance indicates that GNN-NPI is effective for ACP recognition. The powerful capability of NPI may contribute to research in biology and bioinformatics. The dataset and codes are available at https://github.com/Zlclab/GNN-NPI.