Molecule of the Month: Celebrating 55 Years of the PDB

Since the golden anniversary of the Protein Data Bank in 2021, the biggest development in structural biology has been the growth of computed structure models from AlphaFold2, RoseTTAFold, and other protein structure prediction programs utilizing artificial intelligence/machine learning approaches. This article follows the evolution of protein prediction through the exploration of three structures that relied on the development of these new computational methods: the successful prediction of a single protein chain, modeling a flexible molecular complex, and a phage built from a synthetic genome.

Current protein prediction methods were made possible by the wealth of structural data already contained within the PDB. AlphaFold and other algorithms can be thought of as pattern recognition systems, wherein the patterns they learned, such as how an amino acid sequence maps to a fold, and how different chains can pack together, were derived from the over 150,000 experimentally determined structures in the PDB archive at the time they were trained.

First held in 1994, CASP (Critical Assessment of Structure Prediction) is a biennial competition challenging the scientific community to predict the three dimensional structure of proteins based on their sequence. The closer the predicted structures are to an experimentally derived structure, the higher the awarded score. For decades, the average CASP prediction precision hovered around 20%. With the growing use of machine learning and deep neural networks, scores increased significantly during CASP 12 and CASP 13 (held in 2016 and 2018), but still left significant room for improvement. In CASP 14, held in 2020, however, a clear front runner emerged: DeepMind's AlphaFold2 consistently produced structures with accuracy comparable to lower resolution experimental techniques.

One of the protein targets in CASP 14 was a protein known as ORF8 from SARS-CoV-2. ORF8 is a small, rapidly evolving accessory protein the virus uses to blunt the host immune response. ORF8 was a challenging structure to predict because it shares very little sequence similarity with other proteins that have experimentally determined structures. As shown in the figure to the right, AlphaFold2 was able to produce an accurate prediction of the ORF8 experimental structure (pdb_00007jtl ), determined later.

Within a couple of years of AlphaFold2's success, predicted protein structures are now commonly used to help determine the structures of multi-subunit assemblies. An example of this development is the augmin, or HAUS, complex. Augmin is an eight-subunit complex that helps build the mitotic spindle by binding to the sides of microtubules and nucleating new, branching microtubules through the recruitment of γ-tubulin. Its elongated shape and flexibility made it difficult to solve using crystallography or single-particle cryo-EM methods alone, however.

Recently, researchers were able to solve the structure of augmin (shown on the left, pdb_00007sqk) using an integrative or hybrid methods approach. Individual augmin subunits and small groups of proteins were modeled using AlphaFold2 and docked into cryo-EM maps with the benefit of additional biochemical crosslinking data. The resulting structure revealed a 45 nanometer-long, highly flexible complex with a V-shaped head, which is thought to mediate binding to microtubules, and a bifurcated tail that is likely involved in the recruitment of γ-tubulin.

The final highlighted example takes a very different approach to the use of computation and AI. Instead of predicting a structure from a known sequence, researchers computationally generated a new phage genome using genome language models and then studied the resulting structures. They used the phage ΦX174, a small icosahedral virus with a single-stranded DNA genome and eleven genes, as the template. Out of numerous candidate genomes that were generated, the team synthesized and tested nearly three hundred designs and recovered sixteen viable phages capable of infecting ΦX174-resistant bacteria. One of the generated phages, Evo-Φ36 (pdb_000036cr) is shown in the illustration on the right. Evo-Φ36 is very similar to the ΦX174 template (pdb_000036cq), but had significant changes in its version of protein J, a small internal protein that helps package the genome and support the capsid shell from within. As shown in the illustration, Evo-Φ36 had incorporated a shorter, evolutionarily distant version of protein J (shown in purple). On its own, this change would have led to an unviable phage. The generated genome, however, also carried a constellation of accompanying changes that made this change in protein J compatible with its capsid.

While the success of protein prediction methods is undeniable, there are still remaining challenges that have been highlighted by CASP17. These include predictions of protein-ligand complexes, nucleic acids and protein-nucleic acid complexes, and conformational ensembles that describe the range of states that a molecule or complex samples over time.