General Intelligence framework to predict Virus Adaptation based on genome Language model

Current artificial intelligence (AI) solutions for assessing virus phenotypes are mostly limited to fixed tasks, with trained models. We aim to build General Intelligence to predict Virus Adaptation based on Language model (GIVAL), instantly for any input gene or gene segment from any virus. A gene embedder in GIVAL, named virus Bidirectional Encoder Representations from Transformers (vBERT), was pretrained with context-dependently segmented tokens of presently available viruses. Host adaptation of virus input was predicted based on its vBERT embedding, by the input-specified deep learning model, trained with input-specified training data and labels. GIVAL’s vBERT performed well (better than vBERT pretrained on fixed number of amino acids as tokens) on embedding virus genes, at both coarse-grained intact gene and fine-grained gene site levels. GIVAL interpretably predicted the high human adaptation of swine H3N2 and equine H3N8 influenza viruses, and the receptor binding variance of various types of coronaviruses, based on the input of segmented Hemagglutinin (HA) or Spike. GIVAL predicted a significant adaptation shift of the monkeypox viruses since 2022 based on multiple intact viral genes. Summarily, this study provides a general AI solution to assess virus risk, highlights the importance of adaptation shift on transmission risk of multiple viruses. Scripts and data are available on GitHub (https://github.com/Jamalijama/GIVAL) and Zenodo (10.5281/zenodo.14233092).