Raman Spectra of Amino Acids and Peptides from Machine Learning Polarizabilities.

Area of Science:

• Computational chemistry
• Spectroscopy
• Machine learning

Background:

• Raman spectroscopy analyzes molecular vibrations and composition.
• Simulating Raman spectra relies on electronic polarizability, often derived from first-principles data.
• Current machine learning (ML) models face challenges in transferring knowledge from small molecules to larger structures like peptides due to high computational costs.

Purpose of the Study:

• To develop and evaluate ML models for predicting amino acid polarizabilities.
• To assess the transferability of these ML models to larger peptide structures.
• To simulate and analyze Raman spectra of amino acids and small peptides.

Main Methods:

• Training two ML models (including a neural network) on first-principles data to predict polarizabilities of all 20 amino acids.
• Benchmarking ML models against density functional theory (DFT) calculations.
• Combining ML-predicted polarizabilities with classical force field molecular dynamics for Raman spectra simulation.

Main Results:

• A neural network model demonstrated superior transferability for predicting amino acid polarizabilities compared to another ML approach.
• Simulated Raman spectra for amino acids showed good agreement with experimental data.
• Incorporating peptide bond structures into the training set significantly improved prediction accuracy for peptides, even those not in the training data.

Conclusions:

• ML models, particularly neural networks, can efficiently predict polarizabilities for Raman spectroscopy of amino acids.
• The developed models show promising transferability to peptide structures, overcoming limitations of direct DFT calculations.
• This approach facilitates accurate computational Raman spectra analysis for biomolecules.