The AlphaFold Database of Protein Structures: A Biologist’s Guide

1In July 2021, the predicted three-dimensional models for the whole human proteome generated using AlphaFold, the deep learning algorithm developed by DeepMind, were made available to the public, as recently reported in Nature. In the absence of an experimental structure, computational methods have been used for decades to predict three-dimensional protein models. Before the advent of the AlphaFold algorithm, the main approaches were homology modelling and ab-initio. In homology modelling (or template-based approach), which was the most successful and widely used approach, a model is built based on the experimental structure of a homologue, which serves as a structural template. In the ab-initio method (or template-free approach), the model is built by using physics-based and/or knowledge-based energy functions, combined with evolutionary information, which are used to generate distance (or contact) maps.

2The deep neural network of the AlphaFold algorithm, which combines features derived from homologous templates and from multiple sequence alignment to generate the predicted structure, has shown an outstanding accuracy in predicting the three-dimensional structure of proteins with otherwise unknown fold. In CASP14, which is a blind trial that critically assesses techniques for protein structure prediction, AlphaFold (which entered the blind trial under the name AlphaFold2, to distinguish this from an earlier version), markedly outperformed other protein structure modelling methods. When using the root mean square deviation (rmsd), a commonly used method to measure the similarity between two structures (the lower the score the more similar the structures), AlphaFold models had a median backbone accuracy of 0.96 Å rmsd compared to 2.80 Å rmsd of the next best performing method. AlphaFold models also had a high level of accuracy in predicting the position of residue side chains when the protein backbone prediction was accurate. The leading edge performance of AlphaFold is confirmed by the on-going Continuous Automated Model EvaluatiOn (CAMEO). In light of this remarkable achievement, DeepMind made the entire set of models for the human proteome freely available to the scientific community, available at https://alphafold.ebi.ac.uk/ and hosted by the European Bioinformatics Institute.

3From the perspective of the biologist and the non-expert in the structural biology field, what are the advantages, the limitations and the still unsolved challenges of the models generated by AlphaFold? Currently <10% of the proteins in the human proteome have at least some experimentally-obtained coordinates (protein-level coverage) and ∼17% of the residues in the human proteome can be mapped to an experimental structure (residue-level coverage) (4). In the AlphaFold database, the protein-level coverage for the human proteome is 98.5%. However, only 58% of residues are modelled with high confidence, defined as a predicted local distance difference test score [pLDDT] > 70. This 58% high confidence residue-level coverage is an overall improvement of <10% compared to the combined coverage of experimental structures and models generated using templates with sequence identity >30% and standard template modelling predictors (∼50% residue-level coverage).,, However, this increment of coverage will be transformative by providing models which would not be otherwise available to the community. Moreover, the improved accuracy of AlphaFold models compared to template-based ones will be important in several applications, including structure-based drug discovery, variant prediction and to assist experimental structure determination (e.g. molecular),,, (and extensively discussed in the JMB AlphaFold Special Issue, Volume 433, Issue 20, 1st October 2021). However, in cases where the predicted model of the holo form with its cognate ligand is important, a less accurate model which inherits the ligand coordinates from the template may provide more biological insights compared to a more accurate AlphaFold model of the apo form. At present, the models released by AlphaFold do not allow user selection of the appropriate ligand-bound template, which is facilitated by many of the traditional template-based methods.,

4This relatively small improvement in coverage is not surprising given that 37–50% of the human proteome is predicted to be structurally disordered. Disordered protein domains are often important for intracellular signalling and can transition from a disordered to an ordered state, e.g. upon binding to other proteins. Predicting how these amino acid sequences fold remains a challenge. In the AlphaFold models, these disordered regions are identified by a pLDDT < 50 and are often graphically presented as long filaments.

5Another major challenge in the field of structural biology and protein modelling is the identification of the correct placement of domains in a multi-domain protein, also known as inter-domain accuracy. AlphaFold provides full chain models for >98% of human proteins, many of which are multidomain. In CASP14, the AlphaFold inter-domain accuracy was good (formally 70% of models having a template modelling (TM) score > 0.7). Domains are often connected by short and flexible stretches of amino acids, known as linkers, which allow domains to undergo conformational changes in response to biological stimuli. In the AlphaFold models, these linkers are not always predicted at high confidence (pLDDT > 70). The implication of this is that the spatial placement, and in some cases, proximity of two ordered domains should be interpreted with caution. Here, we wish to highlight the need to inspect the heat map or “predicted aligned error” provided by AlphaFold that displays the model’s inter-domain accuracy, which should always be considered alongside the per-residue pLDDT score when interpreting model accuracy. Additionally, the relative position of domains should be explored using biological data, For example, by using experimental structures with lower resolution, structures of homogous proteins or of complexes with partial coverage of the protein sequence.

6We illustrate the challenge of positioning domains with two examples. Figure 1(A) shows the predicted structure for the growth hormone receptor, where the long disordered intracellular tail is placed next to the ordered extracellular domain. Figure 1(B) shows that the relative location of the domains in PIK3R1 is inconsistent with the experimental structure of the PIK3R1 / PIK3CD complex with major clashes between chains. In this example, the relative positions of the PIK3R1 domains may alter between the single chain and the complex. Hence, if the links between the domain are flexible, AlphaFold could be generating a correct model for the single chain or be generating one of an ensemble of domain conformations.

7Another challenge for protein structure predictions is that several proteins are very long. Currently the AlphaFold database on the EBI website does not include models for proteins longer than 2700 residues. Thus, no models are available for 207 large (residue range 2701–34350), biologically important human proteins, such as those encoded by Titin and Dystrophin, the main genes responsible for congenital cardiomyopathy and muscular dystrophy. However, AlphaFold has generated several overlapping model fragments for these proteins (available for download at https://alphafold.ebi.ac.uk/download). Inevitably, interpreting models for very long proteins will be difficult.

8The structural coverage of the human proteome is not uniform. A recent study showed that some classes of proteins, such as drug targets, have been studied better than others and their structural coverage at protein level is already very high. We explored the additional value of AlphaFold models compared to the coverage that can be obtained using standard homology modelling algorithms, such as Phyre2, on two sets of proteins that make a fundamental contribution to morbidity and mortality: the top 25 cancer proteins from the PanCan TumorPortal database and the top 5 proteins causing familial hypercholesterolemia, one of the main inherited causes of premature cardiovascular disease (https://panelapp.genomicsengland.co.uk/panels/772/). Of these 30 proteins, 8 are longer than 2700 residues and models are not provided for these on the EBI website. For the remaining 22 proteins, the additional coverage at residue level provided by AlphaFold models (pLDDT > 70) over standard homology methods, exemplified by Phyre2, was not substantial: 13,059 versus 13,214 (Table 1).

9In conclusion, the AlphaFold algorithm has rightly been called a “game changer” in the field of structural biology and has demonstrated one of the many applications of deep learning algorithms in biomedicine., However, AlphaFold has not completely solved the “protein folding problem” and many challenges remain, such as predicting the relative position of domains within a chain, how domains shift their relative conformation in response to stimuli, and how domains transition from disorder to order.

10AD and SI performed model analysis. AD wrote the first draft of the manuscript. All authors contributed to the interpretation of findings and manuscript preparation. All authors approved the final version of the manuscript.

11AD is supported by Wellcome Trust (grant 218242/Z/19/Z).

12ET is supported by a BBSRC grant to Imperial College London (BB/M011178/1).

13These Funders and DeepMind had no role in the conceptualization, design, data collection, analysis, decision to publish or preparation of the manuscript.

14This research was funded in whole, or in part, by the Wellcome Trust grant number 218242/Z/19/Z. For the purpose of open access, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission.

15DeepMind are proving funding for Master studentships at Imperial College London including potentially for a course of which MJES is the Director. The Authors declare no other competing interests.