Supporting Information for Machine-Learned Extrapolation of Quantum Mechanical Energies in Implicit Solvent from Short to Long Oligopeptides

This repository contains a set of XYZ files, which are PeptideCS2 datasets and two zip files, which are Supporting Information for the parent paper. The XYZ files contain geometries of capped peptides of different lengths calculated with density functional theory (DFT). Energies were calculated at BP86/dgauss-dzvp level with D3-BJ dispersion parameters from https://rezacovi.cz/science/dft-d3.html (a1=0.7182 a2=3.8572 s8=3.2176) with COSMO solvation model (epsilon=80). Geometries were optimized at GFN2-xTB level with ALPB solvation (water) by xTB program (https://github.com/grimme-lab/xtb), except where otherwise noted. The second line in the XYZ entry contains name, charge and multiplicity of the system, followed by key=value pairs. The important keys are "energy", which is the E_DFT and "energy_ref_cosmo", which is E_f. Both energies are in atomic units (Hartree, which is 627.509474065384 kcal/mol). The file names are composed of three parts: "pepml2", dataset designation (below) and length of peptides (1,2,3,5,10).• d2a (_C) - structures with constrained dihedral angles • d2m (_M)- minima obtained by optimizing d2a structures• d3 (_H) - constrained structures + 800 K hot MD Files d3gas_t_length.xyz contain gas-phase energies and and gradients (without COSMO solvation) of the constrained (_H) peptides. File pepml2_d2m_10_dft_reopt.xyz contains geometries from d2m, which were reoptimized at DFT level and which were used to benchmark geometry optimizations. File apsci_dataset.xyz contains geometries/energies from the Atlas of Protein Side Chain interactions. Energies (corresponding key in the file is "energy=", "grads=") were recalculated at DFT level. Individual XYZ entries are named by their amino acid sequence and a numerical suffix. The files containing decapeptides contain both ~1000 random decapeptides (suffix "_0") and conformers of decapeptides. Specifically, for 10 decapeptides the file contains 99 additional conformers each, denoted by suffixes other "_1" - "_99". The test7_sp80_structures_nov2025_amberFeb26.xyz file contains the V31C_M dataset. The trp_cage_kormi.xyz file contains the V23C_M dataset. Structures with names starting with "folded_1RIJ" correspond to structures from MD runs of the folded structure (PDB: 1RIJ, 1st model). Structures with names starting with "unfolded_1RIJ_[1-3]" correspond to structures of MD runs of 3 separated unfolded simulations. The file MD_folded_1RIJ_0_deployed_pepml2_3_230_d3_EF_530.0K_0.5fs_VelocityVerlet_sample.pdb contains a sample of the frames from the MD simulation of the E6apn1 miniprotein. The pair_scans.xyz file contains geometries and energies of acetate/methylammonium/ethane pairs. The SI.zip contains the following items:• python_libraries contains modules of our custom dataloader. You should put the folder in PYTHONPATH.• individual_correlations contains energy predictions (Ef) of trained models on validation datasets• images_training_runs contains training curves of models• correlation_statistics contains figures of RMSE, RMSE_shift, R2 for different model-validation set combinations• inference_demonstration.ipynb contains example Jupyter notebook The nequip_models.zip file contains the following folders:• nequip_models contains yaml files and also trained model weights in "lubos" subdirectory Additionally, the file pepml2_3_230_d3rs_E.zip contains trained model weights for the T3(RS)_H_E model.