Transferability of small-molecule-trained polarizability ML models to larger structures

Determine the extent to which machine-learning models for predicting molecular electronic polarizabilities that are trained on small molecules transfer to larger biomolecular structures such as peptides and proteins for Raman spectra simulations, without relying on direct training on large structures.

Background

Machine-learning models can predict electronic polarizabilities efficiently and enable molecular dynamics-based Raman spectra simulations. However, training such models directly on large biomolecules is computationally prohibitive, making transferability from small-molecule training sets a critical issue.

The work investigates neural-network and symmetry-adapted Gaussian process regressor approaches for amino acids and small peptides, but the broader question of transferability to substantially larger structures remains explicitly highlighted as uncertain at the outset.