domeet webmaster

DeepMind: Previously defeated humans with AI, this time defeated the new crown virus with AI!

Editor’s note: This article comes from WeChat public account “Xin Zhi Yuan” (ID: AI_era) .

Source | deepmind

Edit | Bai Feng, Peng Fei, Zhang Jia

[New Zhiyuan Guide] This time, DeepMind, known for defeating humans with AI, wants to help humans defeat the new crown virus! DeepMind posted that using the deep learning system AlphaFold can predict the protein structure of the new crown virus, and also released six predicted structures, which are crucial for scientists to understand the virus and develop vaccines.

DeepMind, a star AI company owned by Google ’s parent company Alphabet, defeated the world-renowned humanity, defeated the human Go world champion with AlphaGo, and defeated 99.8% of human players with AlphaStar. This time, DeepMind wants to help humans defeat the new crown virus.

In order to detect viruses and develop vaccines, scientists must first understand the structure of the virus, and especially the proteins of the virus. This is a long process that takes months and is sometimes futile. In recent years, researchers have turned to computer prediction.

Coronaviruses are being studied in laboratories around the world, and DeepMind’s deep learning system is called “AlphaFold”.

DeepMind recently published an article saying”Publishing structural predictions of several under-researched proteins related to SARS-CoV-2, the virus that causes COVID-19,” was released to help with the study. Next, let’s share this article with DeepMind.

DeepMind: Computational prediction of COVID-19-related protein structure

“Human studies on coronaviruses have a history of decades, so using the previous database to respond to the COVID-19 epidemic quickly, we have developed a new virus detection method in just a few days.

The culprit of COVID-19 this time has not been able to determine the protein structure of SARS-CoV-2 virus. The traditional method may take months or even longer, which will help us understand the function and transmission mechanism of the virus. Created great obstacles.

Given that the traditional method takes too long, we use the new version of AlphaFold to predict the protein structure of SARS-CoV-2. The new system can obtain accurate predictions without the similar protein structure.

We have shared the viral protein structure predicted by several models, and hope to provide some help for researchers.

We believe the new system is more accurate than our earlier CASP13 system. Previously, we successfully predicted the experimentally verified SARS-CoV-2 spinous protein structure in a protein database, which gave us enough confidence that the new system is also possible to predict other protein structures.

We recently shared our findings with several colleagues at the Francis Crick Institute in the UK, including structural biologists and virologists, who encouraged us to publish our structure now. Our model will indicate which parts of the structure are more likely to be correct. Although these have not been studiedProteins are not the focus of current treatments, but they may increase researchers’ understanding of SARS-CoV-2.

Normally, we will wait for this work to be published after peer review. However, in view of the urgency of time and the severity of the epidemic, we have decided to release the predicted structure of six proteins related to SARS-CoV-2. These data files are licensed under open source and can now be used by anyone.

Interested researchers can download these data files via the link we provide, with many technical details attached. Finally, I want to emphasize that these are predictive structures that have not been experimentally verified. “

Xin Zhiyuan produced the predicted structure diagrams of six proteins based on the download file provided in the DeepMind article, for reference only (not guaranteed to be 100% accurate):

The original file download address:

https://storage.googleapis.com/deepmind-com-v3-datasets/alphafold-covid19/structures_4_3_2020.zip

DeepMind predicts the new weapon of the new crown virus “protein folding”: AlphaFold! Precision crushes friends and humans

It is important to predict the protein structure of the new coronavirus, which can enable scientists to gain more knowledge about protein shape and how it works through simulations and models, and it opens up new potential for new drug development and reduces experiments Costs, and greatly speed up the process of scientists finding more effective treatments, ultimately saving patients around the world.

And this time DeepMind is used to predictThe blockbuster weapon for testing the protein structure of the new crown virus is AlphaFold, which was launched at the end of 2018 and is popular among various media.

DeepMind brings together experts from the fields of structural biology, physics, and machine learning to apply cutting-edge technology to predict the 3D structure of a protein based only on its gene sequence.

It is reported that AlphaFold is not only far ahead of its friends in the CASP global competition known as the “Olympic Games” for protein structure prediction, but even its prediction accuracy has crushed human experts.

Determining the three-dimensional shape of a protein purely from a gene sequence is a complex task. The challenge is that DNA contains only information about the sequence of protein building blocks, called amino acid residues, which are arranged to form long chains. Predicting how these long chains fold into complex 3D structures of proteins is known as the “protein folding problem.”

The larger the protein, the more complicated and difficult the model is because the interactions between amino acids are more complicated. Some researchers estimate that, at the level of current conditions, it may take longer than the life of the universe to accurately determine the 3D structure of all proteins.

A new method for protein structure prediction based on deep learning

AlphaFold models the morphology of a protein from scratch without using the resolved protein as a template. The results achieved high accuracy in predicting the physical properties of the protein structure. Based on this, two different methods were used to construct the prediction of the complete protein structure.

Both methods rely on deep neural networks to predict the properties of proteins from their gene sequences. The network has two main predictive indicators: (1) the distance between amino acid pairs, and (2) the angle between the chemical bonds connecting these amino acids. This technique is used to estimate whether amino acid pairsThis approach.

The above figure has a distance matrix of three proteins. The brightness of each pixel represents the distance between amino acids in the protein sequence. The brighter the pixels, the closer the pairing. The top row shows the actual, experimentally determined distances, while the bottom row shows the average of Alphafold’s expected distance distribution. Importantly, these matches are good both globally and locally. The bottom panel uses a 3D model to represent the same comparison, which is characterized by Alpha Fold’s prediction (blue) and ground truth data (green) relative to the same three proteins.

The DeepMind team trained a neural network to predict the individual distribution of distances between each pair of residues in a protein. These probabilities are then combined to form an accuracy score for the corresponding protein structure prediction. In addition, a separate neural network was trained to aggregate all the predicted distances to estimate the closeness between the predicted structure and the actual structure.

Use these scoring functions to find protein structures that match your predictions. The first method is based on the techniques commonly used in structural biology, and iteratively replaces the original protein structural fragments with new protein structural fragments. To this end, the research team trained a generative neural network to invent new protein fragments and continuously improve the score of predicted protein structures.

The second method uses gradient descent to optimize the score. Gradient descent is a technique commonly used in machine learning and can be used to achieve small, incremental improvements that ultimately produce highly accurate structures. The researchers reduced the complexity of the prediction process by applying the technique to the entire long protein chain, rather than to fragments that must be folded separately before assembly.

Can’t wait for peer review, release forecast results as soon as possible, unknown accuracy

This is in sharp contrast to the previous GPT-2’s covert release. Due to the urgency of the current epidemic, DeepMind couldn’t wait to pass the peer review, and immediately released the prediction results of the protein structure, and stored it in a pdb file. After downloading, researchers can draw the image by using the data in the pdb file.

Although AlphaFold has achieved excellent results in the competition, its shortcomings in the stability of prediction are also very significant. It showed two extremes in the competition: of the 43 predictions, 25 were very accurate, while others were outrageous. Therefore, the accuracy of AlphaFold’s prediction of the new crown virus is still unknown, and we look forward to peer review results and verification of actual clinical treatment.

In any case, DeepMind’s move has opened up a new idea for the application of AI to practical scenarios, especially in the fight against epidemics. At the same time, we look forward to more domestic AI companies to find their own breakthroughs in this fight against epidemics!