Two examples of protein targets in the free modeling category. Alphafold predicts very accurate compositions measured against the experimental result. Credit: Deepmind
Deepmind develops an AI solution to a 50-year-old protein challenge, creating the potential to accelerate biological research.
In major scientific advances, the latest version of Deepmind’s AI system Alphafold has been recognized as a solution to the 50-year-old grand challenge of predicting protein synthesis, often referred to as the ‘protein folding problem’. Can accelerate, unlocking new possibilities in the understanding of the disease and in search of drugs in other fields.
Today, CASP 14 results show that DeepMind’s latest alphafold system achieves unparalleled levels. Accuracy The system in structure forecasting is able to determine high-precision compositions in a matter of days. CASP is a biennial community-driven assessment of protein structure forecasting, a 1994-initiated community-led assessment, and a gold standard for evaluating predictive techniques. Participants must blindly predict the composition of proteins that have only recently – or in some cases not yet – been experimentally determined, and their predictions must be compared with the experimental data.
CASP uses the “Global Distance Test (GDT)” metric to evaluate accuracy, ranging from 0-100. The new Alphafold system achieves an average score of 92.4 GDT in all targets. The average error of the system is about 1.6 angstroms – about the width of one Atom. According to Professor John Moult, co-founder and chairman of CASP, a score of about 90 GDT is considered informal to be competitive with the results obtained from experimental methods.
Professor John Moult, co-founder and president of CASP, University of Maryland, said:
“We’re stuck on this one problem – how the protein shuts down – for almost 50 years. To see if Deepmind produces a solution to this, after working on this problem individually for so long and after so many stops and One wonders if we’ll be there, it’s a very special moment. “
Why the prediction of protein structure is important
Proteins are essential for life and their shape is closely related to their functions. The ability to predict protein structures enables a better understanding of what they do and how they function. The main database currently contains over 200 million proteins and only a fraction of their 3D structures have been mapped out.
A big challenge is the astronomical number of ways in which a protein can theoretically fold before settling into the final 3D structure. Many of the major challenges facing society, such as developing disease treatments or finding enzymes that break down industrial waste, are fundamentally linked to proteins and the role they play. Determining protein shapes and functions is a major area of scientific research, primarily using experimental techniques that can employ workers and laborers for years per structure, and requires the use of specialized multi-million dollar tools.
Deepmind’s approach to the protein folding problem
This progress builds on Deepmind’s first entry in CASP13 in 2018, where the early version of Alphafold achieved the highest level of accuracy among all participants. Now, Deepmind has developed new deep learning architectures for CASP14, drawing inspiration from the fields of biology, physics and machine learning, as well as the work of many scientists in the field of protein folding over the past half century.
The folded protein can be thought of as a “spatial graph”, where the residues are nodes and the edges connect the residues closely. This graph is important for understanding physiological interactions with proteins as well as their evolutionary history. For the latest version of Alphafold used in CASP14, Deepmind has created a focused neural network system, trained in the final stages, trying to interpret the structure of this graph while reasoning on the constructed graph. It uses evolutionarily related sequences, multiple sequence alignment (MSA) and amino representation Acid Remaining pairs to correct this graph.
By repeating this process, the system develops strong predictions of the underlying anatomy of the protein. In addition, Alphafold can predict which parts of each predicted protein structure are reliable using internal confidence measurements.
The system was trained on publicly available data containing the protein data bank’s ના 170,000 protein structures, including modern machine learning standards – about 128 TPUV. Relatively simple calculations run by 3-cores (equivalent to about -2 100-200) have been used. A few weeks.
Potential for real world impact
Deepmind is excited to collaborate with others to learn more about the potential of alphafold, and the Alphafold team is exploring how predictions of protein structure can contribute to the understanding of certain diseases with some expert groups.
There are also indications that predicting protein structure as one of the many tools developed by the scientific community could be useful in future epidemic response efforts. Earlier this year, Deepmind predicted several protein structures SARS-CoV-2 Effectively fast work by virus and experimental experts has now confirmed that Alphafold has achieved a high degree of accuracy on its predictions.
Alphafold is the most significant advance of the Deepmind to date. But as with all scientific research, much remains to be done, including a figure of how multiple protein complexes form, how they interact. DNA, RNA, Or how to determine the exact location of small molecules and all amino acid side chains.
Like the previous CASSP13 alphafold system, Deepmind is considering submitting a detailed course of operation of this system to a peer-reviewed journal in a timely manner, and is simultaneously exploring how best to provide extended access to the system in a scalable manner.
Alphafold has broken new ground to demonstrate the amazing potential of the AIA as a tool to aid in basic scientific discovery. Deepmind hopes to collaborate with others to unlock potential.
Statements of independent scientists:
Professor Venki Ramakrishnan, Nobel Laureate and President of the Royal Society
“This computational work represents a remarkable advance on the protein-folding problem, a grand challenge in biology that is 50 years old. That has happened many decades before what many in the field would have predicted. It will be exciting to see the many ways in which it will fundamentally change biological research. “
Professor Dame Janet Thornton, Director Emirates and Senior Scientist, EMBL-EBI
“What the Deepmind team has been able to achieve is amazing and will transform the future of structural biology and protein research. After decades of studying proteins, molecules that provide the structure and functions of many living creatures, I wake up this morning to find that progress has been made. ”
Arthur d. Levinson, PhD, founder and CEO Calico, former chairman and CEO, Genetech
“Alphafold is a step ahead of time, predicting protein structures with incredible speed and accuracy. This leap forward demonstrates how many promising calculation methods are in place to transform biological research and accelerate the process of drug discovery. “
Professor Andre Lupus, Director, Max Planck Institute for Developmental Biology
“AlphaFold’s surprisingly accurate models Dello has allowed us to tackle protein structure that has been stalled for almost a decade, resuming our efforts to understand how signals are transmitted to the cell membrane.”
Professor Evan Birney, Deputy Director General EMBL, Director EMBL-EBI
“I almost fell down from my chair when I saw these results. I know how rigid CASP is – it basically ensures that computational modeling must perform on the challenging task of AB-Ario protein folding. She was humbled to see that these models could do it so precisely. There will be many aspects to understand but this is a huge advancement for science. “
Deepmind / Alphabet Statements:
Demis Hasabis, PhD, Founder and CEO, Deepmind
“The ultimate vision behind Deepmind has always been to build AI and then advance knowledge about the world around us by accelerating the pace of scientific discovery. For us Alphafold presents the first proof point for that thesis. This advance is our first major breakthrough in the long-running grand challenge of science, which we hope will have a major real-world impact on disease understanding and drug discovery. “
Pushmeet Kohli, PhD, AI for Science, Deepmind
“These incredible results are a testament to DeepMind’s unique research philosophy – bringing together mission-focused, multidisciplinary teams to target ambitious scientific goals. Critical evaluations such as CASP are important for advancing research, and we look forward to furthering this work, deepening our understanding of proteins and biological mechanisms and opening up new avenues of research. “
John Jumper, PhD, Alphafold Lead, Deepmind
“Protein biology is surprisingly complex and ignores simple traits. The work of our team demonstrates that machine learning techniques are finally able to enable the complexity of describing these amazing protein machines, and we are really excited to see what new advances human health and basic biology will bring. “
Catherine Tunyasunakul, PhD, Science Engineer, Deepmind
“A.I. With the ability to predict high-precision protein structures, with potential applications of drug design and biomediation, we can change how we approach biology. Especially for experimentally challenging proteins, well-predicted techniques can make a big difference. “
Pretty Pichai, CEO, Google and Alphabet
“There are incredible advances in AI-powered in this protein folding, which will help to better understand one of the most basic building blocks of life. These huge jumps of deepmind have immediate practical effects, enabling researchers to tackle new and difficult problems, in response to future epidemics towards environmental sustainability. “
For more on this topic read the Deepmond AI solution to the 50 year old science challenge “could revolutionize medical research”.
About Deepmind
Deepmind is a multidisciplinary team of scientists, engineers, machine learning experts and more, working together to research and build secure AI systems that solve problems and learn how to advance scientific discovery for all.
Gony is known for developing Alfago, the first program to beat the world champion in the complex game, has published more than a dozen research papers containing more than a dozen of nature and science by Deepmind, and has achieved progress results in many challenging AI domains. Protein folding to Starcraft II.
Deepmind was founded in London in 2010, and joined forces with Google in 2014 to accelerate its work. Since then, his community has expanded to include Mountain View teams in Alberta, Montreal, Paris and California.
