The submitted sequences are the top two candidates from an ESM2-650M-guided evolutionary pipeline. An initial pool of 120 variants was generated by iterative masked-token extension, using multiple masked contexts per step to select amino acids from the model’s predicted distribution. These sequences were refined over 15 generations with conservative, class-preserving mutations and repeated heuristic scoring that favors balanced charge patterns and enrichment of C-terminal compatible residues. Final candidates were selected by a weighted composite score and diversity filtering.