ImmFinder: A Multiomics-Based Neural Network Approach for Predicting the Immune Genes in Livestock

Abstract

The classification of immune and nonimmune genes in cattle is crucial for understanding immune mechanisms and their link to disease resistance. Traditional methods rely on manual curation and conventional bioinformatics tools, which are often time-consuming and labor-intensive. We introduce ImmFinder, a multimodal fully connected neural network (FCNN) framework designed to classify immune genes by integrating genomic and transcriptomic datasets. ImmFinder achieved an accuracy of 85.67%, an F1-score of 0.85, a precision of 0.86, and a recall of 0.85, demonstrating strong predictive performance. Additionally, the area under the curve-receiver operating characteristic (AUC-ROC) curve scores of 0.9250 (test set) and 0.9264 (validation set) further validate its robustness. These findings highlight the potential of a multimodal deep learning approach for immune gene classification, advancing functional genomics in cattle. The limitations of ImmFinder include reliance on the available bovine genomic and transcriptomic datasets used for training and evaluation, which may constrain immediate generalization to other breeds or species; additional external validation and experimental follow-up will be required to confirm biological hypotheses derived from model predictions. Currently, ImmFinder demonstrates the value of multimodal data fusion for functional gene annotation and provides a scalable baseline for integrating data types, such as genomics and transcriptomics. In future work, we will expand the training cohorts, broaden the range of data modalities, and pursue experimental validation of high-confidence model predictions. ImmFinder is implemented in Python, and all datasets, training models, preprocessing, and model development scripts are available on GitHub.