Possible input features.

The features are divided into four categories: (1) One-hot encoding of 21 variables representing 20 amino acids and 1 stop codon; (2) AAIndex features consisting of 19 numerical values based on physicochemical properties [3]; (3) Rosetta energetics with 20 energy terms related to protein stability and folding, including attractive forces (fa_atr), van der Waals repulsive forces (fa_rep, fa_intra_rep), solvation energy (fa_sol, Fa_intra_sol_xover4, lk_ball_wtd), dielectric electrostatics (fa_elec), Proline ring closing energy (pro_close), disulfide statistical energies (dslf_fa13), and hydrogen bonding contributions (hbond_sr_bb, hbond_lr_bb, hbond_bb_sc, hbond_sc); and (4) the Root mean square fluctuation (RMSF) derived from atomistic simulations, capturing protein flexibility and represented by a single feature. Note that only 8 of the 20 Rosetta energy terms are selected to be used in training final VEPs reported in this work (highlighted in italic). (XLSX)