Integrating large language models with epidemiological data for antigenic drift monitoring and influenza A/B forecasting

Timely forecasting of influenza-like illness (ILI) and early identification of antigenic drift are critical for informing vaccine strain selection and reducing the impact of seasonal outbreaks. This study proposes a multi-modal deep learning framework that integrates nucleotide-level large language model (LLM) embeddings with time-series forecasting and explainable AI for enhanced influenza surveillance. More than one million curated HA and NA sequences (2000–2024) were used to fine-tune DNABERT-2, generating 768-dimensional embeddings capable of capturing high-resolution genetic variation. These embeddings supported the construction of a weekly mutation intensity index and accurate viral clade classification. A Temporal Fusion Transformer (TFT) combined these genetic features with CDC FluView ILI and laboratory-confirmed case data to predict ILI incidence up to eight weeks in advance. The model achieved perfect sequence classification (macro-F1 = 1.00, MCC = 1.00) and strong forecasting performance (MAE = 0.668). Explainability analyses, including integrated gradients and SHAP, highlighted biologically meaningful features consistent with known antigenic sites. To further interpret drift dynamics, we introduce the Antigenic Change Risk Index (ACRI), a composite score that identifies periods of elevated drift risk. Overall, this work demonstrates that transformer-based nucleotide embeddings, when integrated with epidemiological data in an interpretable framework, can deliver accurate real-time ILI forecasting and antigenic change monitoring, providing a scalable and reproducible tool to strengthen influenza preparedness and vaccine policy.