Google’s DeepMind effort for COVID-19 coronavirus is based on the shoulders of giants – Mashviral News – Mash Viral

Posted: March 8, 2020 at 10:47 am


without comments

Coronavirus could make remote work the norm, something companies need to know COVID-19 coronavirus outbreak could be the catalyst for a dramatic increase in telecommuting. Businesses should be preparing for the rise of remote work and the long-term effects on marketing budgets, corporate travel and commercial real estate.

Sixty years ago, research was underway to understand the structure of proteins, since Nobel Laureates Max Perutz and John Kendrew in the 1950s gave the world the first glimpse into what a protein looks like.

It was that pioneering work and decades of research that followed, which made Googles DeepMind announcement Thursday that an idea of the structure of a handful of proteins associated with respiratory disease known as COVID-19. which is spreading all over the world.

Proteins do a great deal of work for organisms, and understanding the three-dimensional shape of proteins in COVID-19 could possibly conceive of a type of virus behind the disease, which could be a vaccine. Efforts are being made around the world to determine the structure of these viral proteins, of which DeepMinds is merely an effort.

There is always a little self-promotion about DeepMinds AI accomplishments, so it helps to remember the context in which science was created. The DeepMind Protein Polling Program reflects decades of work by chemists, physicists and biologists, computer scientists and data scientists, and would not be possible without this intense global effort.

Since the 1960s, scientists have been fascinated by the difficult problem of protein structure. Proteins are amino acids, and the forces that pull them in a certain way are fairly straightforward because some amino acids are attracted or repelled by positive or negative charges, and some amino acids are hydrophobic that is, they hold further away. away from water molecules.

However, these forces, so basic and so easy to understand, lead to amazing protein forms that are difficult to predict only from the acids themselves. And so decades have passed, trying to guess what a given amino acid sequence will look like, usually developing increasingly sophisticated computer models to simulate the process of folding a protein, the interaction of forces that make a protein take it. whatever shape it ends up taking.

An illustration of the possible structure of a coronavirus-associated membrane protein, according to a model created by DeepMinds AlphaFold program.

DeepMind

Twenty-six years ago, a bi-annual competition, called Critical Evaluation of Predicting Protein Structure, or CASP, was held. Scientists are challenged to submit their best computer simulated predictions of a given protein after telling them only the amino acid sequence. The judges know the structure, which is determined by a lab experiment, so its a test of how you can guess what is in a lab.

DeepMind honored the latest CASP, CASP13, which took place throughout 2018.To grab gold, DeepMind developed a computer model, AlphaFold, which shares a naming convention with the DeepMind model that won. chess and Gos game. AlphaZero. In one of those trophy moments similar to other DeepMind headlines, the company found its closest competitor to the CASP13 competition in 2018, producing high-precision structures for 24 of the 43 domains of proteins, with the highest single effort. producing 14 models of this type.

Writing in Nature this January, Mohammed AlQuraishi with the Systems Pharmacology Lab at Harvard Medical School, called the development of AlphaFold a watershed moment for the science of protein folding. His essay accompanies DeepMinds formal AlphaFold scientific work in this issue, entitled Predicting Enhanced Protein Structure with Deep Learning Potentials.

AlphaFold is a union of AIs work with DeepMind, a product of decades of machine learning progress, but also decades of publicly-acquired protein knowledge. The deep neural network developed by DeepMind consists of a mechanism for measuring the local set of atoms in a convolutional filter-like protein perfected by Turing Yann LeCun winner and used in ubiquitous convolutive neural networks to determine structure local of an image. To that, DeepMind added the so-called waste blocks of the type developed a few years ago by Kaiming He and his colleagues at Microsoft.

DeepMind calls the resulting structure a deep two-dimensional diluted convolutive residual network. The purpose of this mouth is to predict the amino acid pairs distance given their sequence. AlphaFold does this by optimizing their convolutions and residual connections using the stochastic gradient descent learning rule developed in the 1980s, which powers all deep learning today.

This AlphaFold network would not be possible without decades of knowledge of proteins built into publicly accessible databases. The deep network takes in the known amino acid sequence, in a form called multiple sequence alignment, or MSA. These are the pixel equivalent of an image operated by a CNN when image recognition. These MSAs are only available for decades because scientists have been mounting them in databases, in particular the UniProt or Universal Protein Resource database, which is maintained by a consortium of research centers around the world. funded by a group of governments. offices, including the National Institutes of Health and the National Science Foundation. The six DeepMind protein structures published this week for COVID-19 began by taking the freely available amino acid sequences at UniProt, making UniProt the raw material for DeepMinds science.

In addition, on the road to his impressive results, AlphaFold had to be trained. The deep web of convolutions and residual blocks had to take their form, giving examples of structures known as labeled examples. This was made possible by another 49-year-old organization called NSF-funded Protein Data Bank, the U.S. Department of Energy and others. The basic PDB database is managed by a consortium of Rutgers University, the San Diego Supercomputer Center / University of California San Diego, and the National Institute of Standards and Technology. These institutions have the impressive task of retaining what you might consider as the huge data available to AlphaFold and other efforts. More than 144,000 protein structures have been gathered and can be downloaded and downloaded almost half a million times a year, according to the PDB. PDB also runs the CASP challenge.

The DeepMind structure predictions are published in a format called the PDB of the consortium. This means that even the language in which DeepMind can express its scientific findings is possible by the consortium.

The fact that dedicated teams have spent decades painstakingly assembling knowledge stores from which researchers can freely extract is a striking achievement in the history of science and, in fact, humanity.

DeepMinds publication of the protein files was praised by other scientists, such as the Francis Crick Institute. In their blog post about their work COVID-19, DeepMind scientists recognize a lot of work on the virus by other institutions. We are indebted to the work of many other laboratories, they write, this work would not be possible without the efforts of researchers around the world who have responded to the COVID-19 outbreak with incredible agility.

It is a responsible and worthy recognition. It can be added that it is not only the current laboratories that have made the AlphaFold files possible, but also that generations of work carried out by public and private suits have made it possible for the collective understanding of which AlphaFold is only the latest interesting wrinkle.

More:

Google's DeepMind effort for COVID-19 coronavirus is based on the shoulders of giants - Mashviral News - Mash Viral

Written by admin |

March 8th, 2020 at 10:47 am

Posted in Alphazero