Physicochemical fingerprinting reveals convergent evolutionary determinants of enterovirus A71 neurovirulence through integrative machine learning and structural analysis.

Enterovirus A71 (EV-A71) causes hand, foot, and mouth disease and can trigger life-threatening neurological complications, yet the sequence-level physicochemical correlates of CNS involvement across globally circulating lineages remain incompletely defined. Here we screened 15,247 EV-A71 genomic entries spanning 1998-2024, retaining 267 full-length sequences (≥7,000 bp) with confirmed clinical outcomes (7 central nervous system [CNS]-involved, 260 non-CNS). This extreme 7:260 class imbalance, reflecting the scarcity of publicly available full-length CNS-associated EV-A71 genomes, is the principal limitation and interpretive premise of the study. Each polyprotein position was encoded by three Z-scale descriptors-hydrophobicity (Z1), molecular volume (Z2), and electrostatic polarity (Z3)-converting discrete residue identities into a continuous biophysical feature space. A two-stage statistical pipeline (Mann-Whitney U screening followed by odds-ratio ranking) distilled 20 significant loci down to five core positions: P2124_Z1, P997_Z2, P1246_Z3, P1743_Z2, and P1711_Z1 (all P<0.001). Leave-one-out cross-validated logistic regression achieved the highest area under the receiver operating characteristic curve (AUC = 0.889) among eight algorithms benchmarked. Because this AUC is estimated from only seven positive samples, it should be regarded as an exploratory internal performance signal rather than definitive evidence of generalisable accuracy. SHapley Additive exPlanations (SHAP) assigned the largest model contribution to P2124_Z1 (OR = 4.28; 95% CI 1.47-12.51), while P1246_Z3 was statistically associated with lower CNS odds (OR = 0.50); these model-derived quantities do not establish causal mechanisms. Reference-strain mapping linked the five polyprotein coordinates to mature-protein residues in 3D RdRp, 3C protease, 2C helicase, and 2A, thereby providing structural context for cautious biochemical hypotheses rather than confirmed mechanisms. Phylogenetic dispersion of CNS-associated strains was compatible with convergent evolution, but this inference remains limited by the seven available CNS genomes. We therefore present the five-position physicochemical signature and nomogram as hypothesis-generating tools for prioritising candidate neurovirulence markers, requiring prospective validation in larger and more balanced independent cohorts before clinical or field deployment.