Chapter 2. Appendix: The sources read

Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known.

The CASP assessment is carried out biennially using recently solved structures that have not been deposited in the PDB or publicly disclosed so that it is a blind test for the participating methods, and has long served as the gold-standard assessment for the accuracy of structure prediction

Overall, these analyses validate that the high accuracy and reliability of AlphaFold on CASP14 proteins also transfers to an uncurated collection of recent PDB submissions, as would be expected

The other substantial limitation that we have observed is that AlphaFold is much weaker for proteins that have few intra-chain or homotypic contacts compared to the number of heterotypic contacts

The model uses MSAs and the accuracy decreases substantially when the median alignment depth is less than around 30 sequences

In general, AlphaFold is trained to produce the protein structure most likely to appear as part of a PDB structure.

We hope that AlphaFold—and computational approaches that apply its techniques for other biophysical problems—will become essential tools of modern biology.

To train, we use structures from the PDB with a maximum release date of 30 April 2018.

It will enable biologists to obtain structural models for almost any protein sequence, changing how they tackle research questions and accelerate their projects.

Model confidence can vary significantly along a chain, making it essential to analyse the confidence measures before interpreting structural features.

It was recently described that very low confidence pLDDT scores correlate with high propensities for intrinsic disorder

However, figuring out the exact structure of a protein remains an expensive and often time-consuming process.

AlphaFold DB will enable biomedical scientists to use 3D models of protein structures as a core tool, driving research and innovation across multiple fields by providing open access to an ever-growing number of predicted structures.

Why is there a need for databases for predicted structures when researchers can run the software on their proteins of interest?

Regions with pLDDT greater than 90 are generally modelled with high accuracy, making them suitable for high accuracy-dependent applications such as characterising binding sites.

On the other hand, regions with pLDDT between 50 and 70 have lower confidence and should be used cautiously.

Unphysical bond lengths and clashes usually do not appear in the confident structured regions, and any region with several of these should be disregarded.

The database provides access to over 214 million predicted structures, although some sequences might be outdated compared to UniProt due to less frequent data releases in the AlphaFold DB.

These limitations are under discussion.

Arguably, we have entered a new era in structural biology, when the abundance of available predicted protein structure data enables researchers to probe an unprecedented range of biological questions.

As with AF2, AF3 confidence measures are well calibrated with accuracy.

whereby the model may invent plausible-looking structure even in unstructured regions.

Traditional docking methods use the latter privileged information, even though that information would not be available in real-world use cases.

Despite this, we still observe a chirality violation rate of 4.4% in the benchmark.

To obtain the highest accuracy, it may be necessary to generate a large number of predictions and rank them, which incurs an extra computational cost.

For example, E3 ubiquitin ligases natively adopt an open conformation in an apo state and have been observed only in a closed state when bound to ligands, but AF3 exclusively predicts the closed state for both holo and apo systems

We expect that structural modelling will continue to improve not only due to advances in deep learning but also because continuing methodological advances in experimental structure determination, such as the substantial improvements in cryo-electron microscopy and tomography, will provide a wealth of new training data to further the improve the generalization ability of such models.

We suggest considering AlphaFold predictions as exceptionally useful hypotheses.

Of course, many of the proteins in the human proteome that have low-confidence AlphaFold predictions are likely to contain regions that are intrinsically disordered

The definition of very large errors will depend on the situation, but often atomic positions that deviate by more than 2 or 3 Å are of limited value.

In particular, our results show that AlphaFold predictions are not better representations of the contents of a crystal than the models deposited in the PDB, as the deposited models agree much more closely with experimental data where the predicted and deposited models differ.

Using these predictions as starting hypotheses can also greatly accelerate the process of experimental structure determination

The resulting predictions will be increasingly useful structural hypotheses that will form a solid foundation for experimental and theoretical analyses of biological systems.

Our test set of residues (from 102 PDB entries for most analyses, 215 for some) is a small fraction of those in the entire PDB, so it could be useful to analyze a larger, more representative set.

All the unique, protein-containing structures in a 6-month period (December 2021 to May 2022) were analyzed (215 structures).

Although this has been independently verified, the implementation of these methods across structural-biology applications remains to be tested.

Although there may be bias because high-confidence AF2 regions are more likely to have relevant deposited templates, we suggest that the mean pLDDT of predicted pockets can be used as an additional criterion for pocket selection in AF2 structures.

However, Although AF2 returns full protein predictions, these can often contain protein segments that are placed with uncertainty.

As AF2 is trained primarily on X-ray data, the relation between disorder and predicted confidence could be a by-product of using this definition of disorder.

Although the capacity of AF2 to predict the structures of difficult targets has been demonstrated, the extent to which AF2 generalizes to truly never before seen folds remains to be understood.

In summary, we find that AF2 models, when their uncertainty is taken into account, can be applied to existing structural biology challenges, and their quality is near that of experimental models.

Determining these structures experimentally is typically slow and expensive, and sometimes proves impossible

Our results provide guidelines as to how AF2 models should—and should not—be used for effective ligand-binding prediction.

Second, our study examined only GPCRs. The results might in principle be different for other classes of drug targets.

Also, in the relatively rare cases when no structural template is available for a drug target, approaches like AF2 may be the only realistic option short of solving a structure experimentally.

One might be able to redesign docking methods to yield better performance given computationally predicted models.

However, a protein’s biological function often depends on multiple conformational substates

Using nuclear magnetic resonance (NMR) spectroscopy, we could indeed verify our AF-Cluster prediction.

Overall, NMR confirmed our prediction that a triple mutation switches KaiBRS to the FS state.

Second, the number of models returned for each state from AF-Cluster will reflect the abundance of constructs reflecting different states across the protein family, which cannot be interpreted as that state’s Boltzmann weight.

Given that known metamorphic proteins often switch folds through cellular stimuli, it may in general be difficult to experimentally validate novel folds identified through computational methods if the stimulus—whether pH, redox reaction or a binding partner—is unknown.

However, conformational substates not present in the evolutionary signal may require alternative methods.