Machine learning-based insight into the evolutionary divergence of spike protein models in SARS-CoV-2 omicron BA.1 and BA.2 sublineages

SARS-CoV-2 (SARS2) omicron variant was detected with two broad subvariants identified (BA.1 and BA.2) which subsequently evolved into multiple constituent sublineages. During both omicron BA.1 and BA.2 subvariant evolution into sublineages, it was interesting to find that only the BA.2 sublineages led to future generations of SARS2 subvariants while subsequent generations from the BA.1 sublineages never fully proliferated. These distinct epidemiological trajectories suggest differences in spike structural evolution between BA.1 and BA.2, although comparative sublineage analyses remain limited. Here, we utilized AI-based approach to generate spike protein models of the omicron BA.1 and BA.2 subvariants and sublineages. Subsequently, structural differentiation between spike protein models was done using RMSD values, TM scores, and electrostatic potential. Lastly, we performed machine learning (t-SNE, MDS, hierarchical clustering) and dendrogram analyses to establish the structural relationship among the spike protein models of BA.1 and BA.2 sublineages. We found that the spike protein models of the parental omicron BA.1 subvariant varied with all subsequent BA.1 sublineages, whereas, a majority of the spike protein models of the parental omicron BA.2 subvariant had structural similarities with subsequent BA.2 sublineages. In contrast to BA.1, BA.2 sublineages displayed repeated structural convergence toward the BA.2 parental spike model, accompanied by conserved electrostatic surface patterns. We propose that this convergence was operationally described as a reversion-like structural tendency which in-turn may explain why the BA.1 sublineages failed to proliferate into future subvariants, whereas, BA.2 sublineages proliferated into future subvariants.