Citation: Carugo O, Djinović-Carugo K (2023) Structural biology: A golden era. PLoS Biol 21(6): e3002187. https://doi.org/10.1371/journal.pbio.3002187

Published: June 29, 2023

Copyright: © 2023 Carugo, Djinović-Carugo. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: OC acknowledges support from the Ministero dell’Università e della Ricerca (MUR) and the University of Pavia through the program “Dipartimenti di Eccellenza 2023–2027". The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Competing interests: The authors have declared that no competing interests exist.

The first protein structures (for myoglobin) were published in 1958 [1]. This endeavor took more than 20 years and required the development of a new method for protein crystallography, the isomorphous replacement method to solve the phase problem. Each step was manual, laborious, and time-consuming and involved the isolation of a protein of interest from native sources, its crystallization, collection of diffraction data on photographic films, and computational analysis on computers with a fraction of the processing power of a typical modern mobile phone. In 1971, the Protein Structure Database (PDB) was established as the central archive for experimentally determined macromolecular structure data. PDB pioneered data sharing practices at a time when this was still foreign to the vast majority of other scientific disciplines. In 1991, 20 years later, PDB included 684 entries, but since then, there has been an exponential increase in the number of structures deposited.

Many technological advances have resulted from structural genomics consortia, which were formed at the turn of the millennium as a response to large-scale genome sequencing projects and aimed at determining the 3D structure of every protein encoded by a given genome. In the past decade, advances in instrumentation and software have sparked a “resolution revolution” in cryo-EM [2], making it possible not only to achieve resolutions comparable to X-ray crystallography but also to conduct structural studies in intact cells. Advances in experimental structural biology methodologies have transformed structural biology into a multiscale and multimodal approach, which covers a wide range of sizes, resolutions, and timescales, by combining complementary approaches in an integrative manner.

One of the most difficult problems in structural biology is protein structure prediction, also known as modelling. Computational approaches, such as homology modelling, fold recognition (threading), docking, molecular dynamics, and ab initio methods, have been developed in parallel with experimental structure determination techniques to predict the 3D structure of macromolecules, as well as their dynamics and interactions. There are 2 main approaches for modern protein modelling: template-based and neural network (deep learning)-based. Template-based modelling, including fold recognition and homology modelling, uses available resembling structures as templates to generate structure models and has been used for decades, as implemented in the software suites Modeller and Phyre2. Some deep learning methods have been used to aid template-based modelling to produce better results [3]. In 2021, AlphaFold2 and RoseTTAFold were the first to use deep learning approaches [4,5] for protein modelling. This brought about a sensational breakthrough in protein structure prediction and ignited a new revolution in structural biology.

The Protein Data Bank was the source of the training data for Alphafold2, which at the time consisted of more than 170,000 protein structures. DeepMind and the EMBL’s European Bioinformatics Institute partnered to create AlphaFold DB to make structure predictions freely available to the scientific community [6]. The most recent database release comprises approximately 200 million entries with extensive coverage of the UniProt database and has transformed the way experimental structural biology is performed and democratized access and use of information about protein 3D structure. Importantly, the reliability of models is estimated at the level of individual amino acids by the predicted local distance difference test. This enables the user to detect potentially uncertain structural regions; for example, those caused by conformational disorder. Excellent follow-up advances in the prediction of protein–protein interactions and automatization of pipelines have been made with AlphaFold-Multimer [7] and AlphaPullDown [8], respectively, as well as on generation of models loaded with their ligands with AlphaFill [9], which “transplants” small molecules and ions from experimentally determined structures to predicted protein models.

Not surprisingly, these predictions have been challenged on numerous fronts. Some consider them suboptimal because they are based on statistics rather than first physicochemical principles, while others say that, at least for the time being, the accuracy of experimental structures cannot be reached. Careful comparison of AlphaFold2 models with a recently redetermined set of crystal structures revealed that the differences are greater than those between high-resolution structure pairs containing the same components but determined in different space groups [10]. Due to the nature of the prediction and the available training dataset, AlphaFold2 does not accurately model protein aggregation, macromolecular complexes (in particular protein–nucleic acids interactions, macromolecule–ligand interactions, and structural rearrangements triggered by, e.g., ligand binding), or external factors (e.g., pH, temperature, or conformationally dynamic systems), and further developments will be required (Box 1). Furthermore, although the effect of missense mutations on the thermodynamic stability of proteins can be predicted equally well by using experimental structures or AlphaFold2 models [11], AlphaFold2 currently does not appear to be suitable for predicting the structure of mutated proteins [12].

Despite the aforementioned shortcomings, the availability of a realm of 3D models of proteins has sparked a revolution in the design and execution of experimental structural biology projects. In the absence of homologous experimental structures, AlphaFold2 models can be used to bootstrap structure determination using X-ray crystallography (as reliable molecular replacement search models), cryo-EM/ET (structural models fitting electron density maps and interpretation of low-resolution regions), or for modelling small-angle X-ray scattering-derived molecular envelopes. A fine example of the successful use of AlphaFold2 models is the integrative structural biology analysis of the nuclear pore complex [13]. Furthermore, AlphaFold-Multimer and AlphaPullDown are being used successfully for interaction discovery, which requires experimental validation as the subsequent step. AlphaFold2 predictions, therefore, provide a solid foundation for generating hypotheses about molecular mechanisms of action, enable the design of experiments with specific expected outcomes, and permit the investigation of complex systems that would be difficult or impossible to study through conventional methods.