Artificial intelligence in structural biology: Preface

INTRODUCTION

Structural biology has undergone a remarkable transformation, evolving from early pioneers like von Laue¹ and the Braggs, father and son²—who first demonstrated and explained x-ray diffraction by crystals in the early 20th century—to modern investigations reinvigorated by machine learning (ML) and Artificial Intelligence techniques.

The crystallographic solutions to myoglobin and hemoglobin by Perutz et al. and Kendrew et al. in the 1950s ushered in an era of molecular architecture solved by experiment and computation.^3,4 The invention of the computer was an essential part in the development of crystallography as the Fourier transforms are numerically demanding to calculate. As computing power expanded, the field gradually embraced more and more computational tools, but with the advent of AlphaFold2 in 2020—and the later release of AlphaFold 3 in 2024—AI truly revolutionized protein structure determination by prediction.^5,6 While AlphaFold-generated structures have limitations on their accuracy, they have proved to be extremely valuable and are now available for many protein sequences.

Recently, machine learning (ML) has seen widespread adoption in experimental structural biology. For example, in NMR, super-resolution methods such as DiffNMR3 allow low-field instruments to match the resolution of high-field ones,⁷ while graph neural networks have improved the prediction of shielding tensors and advanced 2D NMR spectra prediction using solvent-aware deep learning.⁸

Cryo-EM has similarly benefited from ML. CryoPPP offers expert-curated micrographs for training deep particle pickers,^9,10 and ML-enhanced density modification workflows are now part of cryo-EM map interpretation.¹¹ AI has also been used to identify contaminating proteins in cryo-EM studies.¹² There are many other examples, including a recent one that impacts both cryo-EM and crystallographic refinements—a novel AI-enabled quantum refinement based on machine learned interatomic potentials mimicking quantum mechanical electron distributions at substantially lower computational costs and providing improved chemical geometry.¹³

In crystallography, the program Helcaraxe uses convolutional networks to identify and correct for ice-crystal artifacts in diffraction data,¹⁴ and PhAI represents a deep learning approach to solving the phase problem for small molecules.¹⁵ Other advanced examples include the use of experimental electron density maps to improve the multiple sequence alignment (MSA) phase in generative model building to avoid prediction errors (Doake Hekstra, unpublished), the use of dimension reduction methods to sort conformational variability in proteins,¹⁶ the use of neural networks to assign space groups of crystals,¹⁷ and even the prediction of crystal structures from atomic composition.¹⁸

The use of AI in design of proteins de novo is also a burgeoning field. One can now routinely design a protein from scratch to achieve a particular fold¹⁹ and good progress is being made on designing binding sites for small molecules within those folds.²⁰ The ability to design protein catalysts de novo now seems within reach with a bit more understanding of the role of dynamics or the lack thereof in positioning the substrates for the detailed enzymatic process.

This special issue of Structural Dynamics highlights recent progress in AI-assisted structural biology in a variety of fields. In the next section, we summarize key contributions from these featured authors.

OVERVIEW OF FEATURED ARTICLES

Ille et al. contribute “From sequence to protein structure and conformational dynamics with AI/ML,” an example of how modern AI models are being tailored to address unresolved challenges in macromolecular structure prediction and interpretation.²¹

Patra et al. present “Aminoacyl-tRNA synthetase urzymes optimized by deep learning behave as a quasispecies,” using ProteinMPNN and AlphaFold2 to redesign ancient aminoacyl-tRNA synthetase urzymes. Their deep learning-driven approach improves solubility and catalytic properties. At the same time, the rapid generation of virtual homologs enables determination of how function varies with specific changes. These insights are relevant for the high-resolution attribution of functional effects to fine modular details and open new paths into the study of early enzyme evolution.²²

Trujillo et al. contribute “Filling data analysis gaps in time-resolved crystallography by machine learning,” an example of how modern AI models are being tailored to address unresolved challenges in macromolecular structure prediction and interpretation.²³

The article “Integrated edge-to-exascale workflow for real-time steering in neutron scattering experiments” by Yin et al. describes a scalable workflow for integrating AI into real-time, exascale structural biology. Their framework supports on-the-fly analysis of time-resolved experiments, enabling dynamic structural insights.²⁴

Nguyen et al. contribute “ProteinReDiff: Complex-based ligand-binding proteins redesign by equivariant diffusion-based generative models,” an example of how modern AI models are being tailored to address unresolved challenges in macromolecular structure prediction and interpretation.²⁵

Pan et al. contribute “CrysFormer: Protein structure determination via Patterson maps, deep learning and partial structure attention,” an example of how modern AI models are being tailored to address challenges related to the phase problem in macromolecular structure determination and electron density map interpretation.²⁶

In summary, there seems to be no end to the application of machine learning and other AI tools in the advancement of structural biology. Most of the examples described here do not involve developing new large language models (LLMs) in advancing the field, but no doubt LLMs are being used to help write code, and to draft manuscripts and grant proposals. The authors look forward to the thoughtful and ethical use of AI in the future of structural biology as a whole.

DATA AVAILABILITY

The data that support the findings of this study are available from the corresponding authors upon reasonable request.