Highly accurate protein structure ... | Article | H1 Connect

Using artificial intelligence (AI), DeepMind, a unit of Google have developed algorithms that can predict protein structure with unprecedented accuracy (AlphaFold). This development is a step change in our abilities to predict and solve protein structures, and is likely to revolutionise research in the area of protein structure.

It has long been a given that knowledge of a protein’s structure provides invaluable insight into its function. For many decades this has been the domain of X-ray crystallography, and to some extent NMR, and more recently, cryo-EM. These methods have been of tremendous utility in research into proteins and other macromolecules, but it has been a long-held dream that protein structures should be predicable by computational methods. This has been manifested by the Critical Assessment of Techniques for Protein Structure Prediction (CASP) exercise, which evaluates progress in this area. In CASP14, AlphaFold structures were found to be much more accurate than competing methods, with a median backbone accuracy of 0.96Å RMSD. This development is seen as a significant scientific milestone and has been widely feted in the popular media. AlphaFold v2.0 is open source, allowing users to predict the 3-D structure of arbitrary proteins with unprecedented accuracy.

This work has solved a 50-year-old open research problem in the field of protein structure prediction. The authors have conceptualised this problem as a graph inference problem in 3D space and have used a transformer based deep neural network architecture to arrive at the final, near-accurate 3D structure of a protein from its sequence in an iterative manner. This work can have a direct application in the discovery of a new drug for a disease from the correct 3D structure prediction of its optimal target. 

This article reports a powerful computational method that is capable of predicting high accuracy protein structures. This neural network-based method, AlphaFold, works well even for those proteins with no known similar structure. The method has been validated in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The predicted protein structures for the human proteome and a few other key organisms are made available to the community in the form of a database. This is a huge resource for the community and is likely to accelerate scientific research. 

The authors report AlphaFold, a machine learning method for protein structure prediction, documenting the high accuracy that has been achieved by incorporating new neural networks and training procedures into the system, including multiple protein sequence alignments. In addition to highly accurate protein backbone prediction, AlphaFold can also accurately predict side chains when the backbone is accurate; it can predict hydrogen bonds effectively, and can be used for molecular replacement and interpreting cryoEM maps.

Factors that affect accuracy or limit applicability include multiple sequence alignments of fewer than ~30 sequences and proteins with few intrachain contacts. Further, bound ligands are not predicted, although side chains can be accurately predicted, as noted above, even without interacting ligands.

By making accurately predicted protein structures widely available, AlphaFold is a large step forward for structural biology and potentially transformative for biological studies, especially those that have previously been limited by lack of information regarding molecular interactions.

The use of AI promises to accelerate discoveries in structural biology at a pace never seen before. In this paper, DeepMind presents AlphaFold: a prediction programme that is going to change the way we do structural biology. It seems that no more tedious model building may be needed. The possible implications in drug design are obvious. I predict that the combination of AI methods with experimental structural solutions focusing on dynamics is going to facilitate a quantum leap in our understanding of how biomolecules work.

In the scientific discipline of predicting the three-dimensional structure of proteins, this study represents a tour de force. The authors have used artificial intelligence algorithms (names AlphaFold) to increase the number of human proteins for which the 3-D structure is predicted from 17% to 98.5%. Significantly, to validate the machine learning algorithms, the authors applied this to proteins where the structure had already been established by other experimental methods such as X-ray crystallography and found a very accurate 3-D structure concordance that surpassed competing software. The AlphaFold database has been made freely available and will be invaluable to the wider scientific community in understanding the structure/function relationship of their favourite protein.

The ability to predict the 3D structure of a protein from its sequence has been attempted for a long time, but its accuracy has been limited. In recent years, the rapid development of machine learning has made it possible to outperform humans in games such as Go and Shogi. This paper describes the results of the most accurate structure prediction program currently available from the group that developed AlphaGO. The program in the cloud (AlphaFold2) is open to the public and can be used freely by anyone. It is highly accurate even at the side-chain level, and many structural biologists are amazed by its accuracy. On the other hand, there seems to be some errors in the predicted structures of membrane proteins and proteins that are difficult to structure. In the future, highly accurate structure prediction of complex proteins will be a challenge.

This is an outstanding contribution that describes a truly major advance in the field of protein structure prediction.

The authors have designed and executed a truly novel deep learning structure prediction approach that clearly defines a new state of the art.

This article by Jumper et al. presents AlphaFold, a new machine learning-based program for protein structure prediction. This tool represents an innovative advance for structure prediction, and it is, importantly, open source. While the prediction of structures attainable by intrinsically disordered regions still remains a difficult task, in part because their conformational profiles likely depend on binding partners, the significant advancement made by AlphaFold suggests that accessing this area of biology could be closer than it seems.

This article introduces AphaFold2: It is a must read for all bioinformaticians and structural biologists. The authors have developed a computational method to predict 3D structures of proteins from sequences with atomic accuracy, even when no similar structure is known to serve as a template for comparative modeling. AlphaFold2 won the 14th Critical Assessment of protein Structure Prediction (CASP14).

The description of AlphaFold, an open-source system, developed by Deepmind, to predict relevant protein folding and 3D structure using a novel deep learning approach -- a revolution in the field opening new areas for understanding diseases.

Accurately predicting the structure of a protein solely based on its primary amino acid sequence has been a dream for structural biologists and bioinformaticians for more than 50 years. Despite much progress in our knowledge about protein folding, this has been an evasive goal. Only the use of homology modeling approaches provided satisfactory results, but they required the availability of a related protein with a known structure. This article by Jumper and colleagues marks a breakthrough towards the realization of that old dream. The authors used an artificial intelligence approach that incorporates physical and biological knowledge about protein structure and exploits existing sequences and structures to learn folding rules. This knowledge is then applied to target protein sequences to provide a prediction of how they would fold. In a benchmark, this approach was shown to provide accurate predictions at atomic resolution, even in the absence of similar proteins with known structures, surpassing other existing approaches. This technical advance will certainly have an impact in other fields of biology, biotechnology, and medicine, given the importance of structural knowledge. Now, in principle, it is possible to have a good workable structural model for any protein of interest.

The three-dimensional (3D) structural analysis of a protein is vital for understanding its functions. Although thousands of proteins structures have been determined using huge experimental efforts, they represent only a small fraction of known protein sequences. This illustrates the need for novel approaches that could be used for a large scale protein 3D structural analysis. In this paper, Jumper et al. described a novel computational method, which they call AlphaFold, that predicts protein 3D structures with high accuracy even in cases in which no similar structure is known. This is because AlphaFold is an innovative machine learning algorithm that includes physical and biological knowledge about protein structure and multi-sequence alignments into the design. This deep learning algorithm is hoped to become an important tool in predicting the 3D structure of proteins. 

The prediction of the three-dimensional structures of proteins using the amino acid sequence has been an outstanding scientific challenge as briefly outlined in the introduction. This paper represents an innovative leap in the quality of such predictions. The authors incorporated a number of novel approaches, notably including a machine learning segment. The algorithm was validated using the Critical Assessment of Techniques for Protein Structure Prediction (CASP) assessment, which is a blind test using structures that have been determined but not yet publicly released. The availability of this program provides another approach for scientists studying the functional and structural properties of proteins.

For many decades, biologists have sought to define the structure of proteins through experimental approaches that were pushed to their limit, cryo-EM being the last evolution on this path. Computational methods able to predict the structure of proteins simply through the analysis of their amino acid sequence have been progressively developed, with limited success so far, even if the structure of homologous proteins is available.

In this article by the DeepMind group (Jumper et al.) in London, UK, scientists have developed the first computational method that can predict protein structures with atomic accuracy, even if no similar structure has been solved. This neural-network-based model, named AlphaFold, was validated in the 14th Critical Assessment of protein Structure Prediction (CASP14). Amazingly, the results demonstrate an accuracy that is competitive with experimental structures and able to greatly outperform other computational methods.

With tools like AlphaFold, this may be the beginning of a new era for structural studies. AlphaFold is nothing less than a technological leap forward that demonstrates the power of AI to solve highly complex problems and a better ability to use proteins to understand how diseases work, develop new drugs, etc.