DPCIPI: A pre-trained deep learning model for estimation of cross-immunity between drifted strains of Influenza A/H3N2

Motivation: This study aims to develop a novel model called DNA Pretrained Cross-Immunity Protection Inference Model (DPCIPI) to predict the cross-immunity of influenza virus strains. The traditional method for measuring this is through HI experiments, which are costly and time-consuming. The DPCIPI model uses a pre-trained neural network to vectorize the gene sequences of viruses and predicts the degree of cross-immunity between them. Method: The paper describes the steps taken to develop the DPCIPI model. First, the gene sequence of two viruses is converted into k-mers. Then, the k-mers sequences are aligned, and identical k-mers at the exact position are deleted. The human DNA pre-trained model (DNABERT) is used to vectorize each k-mer in the remaining k-mers. This is followed using a BiLSTM encoder to encode the two viruses into sequence representation and embeddings. An information fusion operation is then performed on the two embeddings to obtain a splicing vector, which is further input into a fully connected neural network for prediction. All parameters of the model are trained simultaneously. Result: Binary cross-immunity prediction predicts whether the HI titer between two viruses exceeds a certain threshold (in our case, an HI titer measurement value higher than 40). Compared with baseline methods such as Logistic Regression, Perceptron, Decision Tree, and CNN-based models, DPCIPI achieves better performance: F1 (88.14%), precision (90.40%), recall (89.69%), and accuracy (89.69%). Multi-level cross-immunity prediction predicts different HI titer intervals. Again, DPCIPI's performance surpasses baseline models. The study concludes that DPCIPI has enormous potential for predicting the cross-immunity between influenza virus strains.