Decoding SARS-CoV-2 Spike Protein Evolution via AI-Enhanced Sequence Entropy Learning and Structure-Based Modeling

SARS-CoV-2 has left a lasting mark on our world through the emergence of multiple Variants of Interest (VOIs) and Variants of Concern (VOCs). These evolutionary trajectories encode valuable information in both genetic/amino acid sequences and three-dimensional structural changes, revealing how the virus adapts and alters its functions¹. Building on the sequence–structure–function paradigm, this presentation asks a central question: can we identify a distinct signal—a “VOC alarm”—capable of detecting critical evolutionary transitions associated with viral adaptation? We first explore the structural evolution of the SARS-CoV-2 spike glycoprotein controlling virus entry to the host cell; therefore a major therapeutic target for neutralizing antibodies and T-cell responses. To investigate its large-scale structural dynamics and mutational effects, we developed a symmetry-information-loaded coarse-grained structure-based model² (SymSBM) using cryo-EM structural data³. Our results uncover a conserved allosteric “cervical collar,” formed by the 630 loop and fusion peptide proximal region (FPPR), that regulates receptor-binding domain dynamics across major variants including Alpha, Delta, and Omicron⁴. Furthermore, Seq-dCor, a phylogeny-inspired residue-level distance-correlation framework, identifies coordinated mutational partners embedded within conserved allosteric pathways, revealing how epistatic interactions rewire conformational landscapes to generate adaptive phenotypes during viral evolution. To investigate sequence-space evolution and collective mutational patterns underlying transitions from variants to VOCs, we introduce the “Mutational Response Function” (MRF), a statistical-mechanics-inspired quantity that quantifies domain-wise entropy fluctuations in spike protein sequences⁵. Remarkably, transitions to VOCs are accompanied by sharp changes in MRF, consistent with signatures of dynamic phase transitions. Leveraging this entropic information, we developed EVOLVE, a statistical-mechanics-guided machine learning framework and webtool that predicts future domain-specific mutations and mutational hotspots¹. Validation across diverse viral proteins demonstrates its predictive capability for identifying mutations and their potential functional consequences.