Determination of spectroscopic properties of biological molecules using machine learning methods (Python code)
Loading...
Date
2026-02-20
Authors
Μινασιάν, Αζάτ
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis investigates the application of supervised machine learning methods for predicting vertical electronic excitation energies of a biologically relevant chromophore embedded in a protein environment. The Photoactive Yellow Protein (PYP) was selected as a model system, with reference excitation energies obtained from quantum mechanics/molecular mechanics (QM/MM) simulations combined with time-dependent density functional theory (TD-DFT). Molecular configurations were encoded using Coulomb matrix representations to provide physically meaningful descriptors suitable for regression models. A dataset comprising 171 chromophore configurations was constructed and used to train and evaluate three distinct machine learning approaches: a Multilayer Perceptron (MLP), a Convolutional Neural Network (CNN), and Kernel Ridge Regression (KRR). All models were trained to predict excitation energies expressed in wavelength units (nm), enabling direct comparison of predictive performance. The results demonstrate that all three models achieve comparable accuracy, with mean absolute errors in the range of approximately 5–6 nm. Kernel Ridge Regression exhibited the lowest overall prediction error and the most stable learning behaviour, highlighting its strong data efficiency in the small-dataset regime. The MLP provided competitive performance with rapid convergence and moderate computational cost, while the CNN did not yield a clear advantage, indicating that the available dataset size is insufficient to fully exploit convolutional architectures. Further analysis of learning curves and excitation-energy distributions revealed distinct generalization characteristics among the models, emphasizing the importance of matching model complexity to data availability. Overall, this work demonstrates that machine learning models can serve as efficient surrogate models for computationally demanding quantum-chemical calculations in biomolecular spectroscopy. Future work should focus on expanding the dataset, incorporating more advanced molecular representations, and exploring hybrid physics-informed learning strategies to further improve accuracy and transferability.
Description
Keywords
Machine learning, Molecular dynamics, QM/MM, TD-DFT