AlphaFold, an attention-based neural network recently developed by Google AI, predicts 3D protein structure from its amino acid sequence with near-experimental accuracy. This work can reveal the role of proteins in disease and facilitate the development of novel drugs.
Determining protein structures
Responsible for most biological processes, proteins are essential building blocks of life. Since their functions are inextricably linked to their structures, understanding protein structure generates mechanistic insights into their role in health and disease.
Previously, protein structure was resolved by experimental approaches, such as X-ray crystallography and cryo-electron microscopy. These methods are highly accurate and are able to reach atomic-level resolution. However, these existing protein structure resolution approaches are also laborious, time-consuming and expensive. As a result, the structure of a vast majority of proteins remains elusive.
Scientists have long puzzled over how a protein’s amino acid sequence determines its final 3D shape. They have made attempts to develop rapid computational alternatives to predict protein structure from a primary amino acid sequence. However, these tools are far from achieving experimental-level accuracy, especially when there is no known, experimentally-resolved structural homologue.
AlphaFold
Recently, DeepMind, an offshoot of Google AI, developed AlphaFold. This machine learning method predicts a protein’s 3D structure from its amino acid residue sequence with near-experimental accuracy, even in the absence of known homologues.
The latest version, AlphaFold 2, significantly outperformed around 100 other software platforms when assessed during the 14th Critical Assessment of Structure Prediction (CASP) challenge in 2020. CASP is a biennial gold-standard assessment for analyzing the accuracy of protein structure prediction methods. One of the criteria is the Cα root-mean-square deviation at 95% residue coverage (RMSD95), the average distance between atoms of the predicted protein structure and the target structure when they are superimposed. A lower RMSD95 thus indicates a more accurate prediction.
In CASP14, AlphaFold structures had a median backbone accuracy of 0.96 Å RMSD95. The next best method had one of 2.8 Å RMSD95. In comparison, carbon atoms are ~1.4 Å wide. Meanwhile, AlphaFold’s all-atom accuracy was 1.5 Å RMSD95, compared to the second-best method at 3.5 Å RMSD95.
Notably, AlphaFold 2 accurately predicted the SARS-CoV-2 Orf3a protein structure before cryo-electron microscopy was able to resolve it. Orf3a is a protein functioning in viral replication and release and may induce the inflammatory response to infection.
Attention-based neural networks in AlphaFold
The source code for AlphaFold 2 is now openly accessible in Nature, involving novel neural network architectures to improve predictive accuracy from earlier versions (Figure 1).
The first module of the network, Evoformer, treats protein structure prediction as a graph inference problem in 3D space. Folded proteins are considered as a “spatial graph”, where the nodes represent amino acid residues, and the edges connect the residues in close proximity to each other. This helps to understand the physical interactions within proteins and their evolutionary history.
As raw inputs, Evoformer takes a primary amino acid sequence and aligned sequences of its evolutionary homologues, or multiple sequence alignments (MSAs). It then generates two arrays. The MSA representation array encodes the relationship between each amino acid position and the different input sequences, while the pair representation array encodes the relationship between two amino acid residues in a protein. The network continuously updates these arrays. Here, an updated MSA representation serves as input to update the pair representation, which in turn becomes the input to update the MSA representation further. Iterating this process means that the two arrays continuously exchange information to refine the spatial graph. This allows for the network to interpret and directly reason about spatial and evolutionary relationships between sequences. Ultimately, it makes strong predictions of the protein’s underlying structure.
Figure 1 – Overview of the AlphaFold architecture. The MSA representation array (pink) continuously exchanges information with the pair representation array (green) to refine the structural predictions. (Credit: DeepMind)
Structure prediction
Informed by the Evoformer output, the Structure Module then generates a 3D backbone structure. Geometry of the protein backbone is represented with independent rotations and translations and are updated iteratively. Notably, the module violates stereochemical constraints in the process, such as by including unphysical peptide bond angles or lengths. This allows for all parts of the protein structure to be simultaneously refined without having to solve complex loop problems. The number of stereochemical violations falls with the number of iterations, ultimately producing a highly accurate protein structure. The Structure Module also calculates an internal measure of confidence to assess which parts of each predicted structure are reliable.
DeepMind trained the network end-to-end with ~170,000 known protein structures from the Protein Data Bank and protein sequences with unknown structures from Uniprot databases.
Summary
AlphaFold solves a longstanding biological puzzle – predicting a protein’s 3D structure by its amino acid sequence. As protein structure dictates its function, the algorithm can reveal the elusive role of proteins in health and disease. This work demonstrates the value of AI in accelerating discovery and research in some of the most fundamental fields in science. With further improvements, AlphaFold may be able to determine the structure of multi-protein complexes. It may even aid with designing novel proteins, driving the development of new drugs.
Image credit: Melnicki – Wikimedia Commons
