Predicting the mutation direction of SARS-CoV-2 using exploratory computational methods presents a challenging, yet prospective, research avenue. However, existing research methods often ignore the effects of protein structure and multi-source viral information on mutation prediction, making it difficult to accurately predict the evolutionary trend of the SARS-CoV-2 S protein receptor-binding domain (RBD). To overcome this limitation, we proposed an interpretable language model combining structural, sequence and immune information. The dual utility of this model lies in its ability to predict SARS-CoV-2's affinity for the ACE2 receptor, and to assess its potential for immune evasion. Additionally, it explores the mutation trend of SARS-CoV-2 via a genetic algorithm-directed evolution. The model exhibits high accuracy in both regards and has displayed promising early warning capabilities, effectively identifying 13 out of 14 high-risk strains, marking a success rate of 93%.". This study provides a novel method for discerning the molecular evolutionary pattern, as well as predicting the evolutionary trend of SARS-CoV-2 which is of great significance for vaccine design and drug development of new coronaviruses. We further developed VirEvol, a unique platform designed to visualize the evolutionary trajectories of novel SARS-CoV- 2 strains, thereby facilitating real-time predictive analysis for researchers. The methodologies adopted in this work may inspire new strategies and offer technical support for addressing challenges posed by other highly mutable viruses.