Modelling protein folding using machine learning methods

DOI: 10.31673/2412-9070.2025.029944

Abstract

The article highlights modern approaches to modelling protein folding using machine learning methods, a rapidly developing field at the intersection of bioinformatics, physics and artificial intelligence. The purpose of the article is to systematise existing approaches to modelling protein folding using machine learning, identify the advantages and limitations of modern techniques, and identify areas for further research in this area.
Predicting the three-dimensional structure of proteins based on amino acid sequences remains a challenging task, as the structure of a protein determines its function in the cell, and its misfolding often leads to severe diseases. This paper reviews the most successful deep learning models, including AlphaFold, MSA Transformer, and ultra-deep neural networks, which have demonstrated the ability to accurately predict protein structures based on the analysis of evolutionary relationships and contact maps.
Particular attention is paid to the limitations of such methods, in particular, their complexity in processing dynamic processes and failure to take into account the stochastic nature of protein interactions. In this context, the author proposes an innovative approach, which consists in the integration of quantum mechanical models, in particular the mechanism of wave function collapse, into classical machine learning algorithms. This approach allows to take into account the probabilistic transitions between protein conformational states and minimise the free energy of the system. Mathematical formalisations and examples of implementation based on the Monte Carlo method are presented.
The proposed integrated model demonstrates an increased prediction accuracy (up to 95%) compared to existing solutions. Its application is promising in personalised medicine (analysis of the effect of mutations on protein structure), pharmacology (improvement of drug design), industrial biotechnology (optimisation of enzymes), and in studies of complex protein complexes. The work forms the scientific basis for the creation of new intelligent tools that combine structural prediction with functional activity analysis, which opens up new horizons for the development of bioinformatics and related fields.

Keywords: proteins, protein folding, machine learning, quantum mechanics, deep neural networks, wave function collapse, bioinformatics, protein structure prediction, recurrent neural networks,
convolutional neural networks.