In a recent research piece put out in Nature, it was demonstrated how artificial intelligence research can spearhead and accelerate new scientific discoveries and inventions. A devoted, interdisciplinary team was assembled in the pursuit of leveraging artificial intelligence to push fundamental research forward: getting together specialists from the domains of structural biology, physics, and machine learning to apply bleeding-edge strategies to forecast the 3D structure of a protein only on the basis of its genetic sequence. 

AlphaFold – detailed in peer-reviewed papers now put out in Nature and PROTEINS – is the result of many years of work, and develops on decades of historical research leveraging large genomic datasets to forecast protein structure. The 3D models of proteins that AlphaFold produces are far more precise than any that have come prior, indicating considerable progress on one of the fundamental challenges in biology. The AlphaFold code leveraged at CASP13 is available on Github for anybody with the interest to learn more or replicate the outcomes from the research. The research has already served as inspiration for other independent implementations, which includes the model detailed in other research papers, and as a community – built, open-source implementation. 

Proteins are large, complicated molecules critical to life as we know it. Almost every function that our body carries out – contraction of muscles, sensing of light, or conversion of food into energy is dependent on proteins and how the move and alter. What any specific protein can do is dependent on its unique 3D structure. For instance, antibodies leveraged by our immune systems are Y-shaped and for unique hooks. By holding on to viruses and bacteria, these antibody proteins are able to identify and go about tagging illnesses – causing microorganisms for elimination. Collagen proteins have the shape of cords, which communicate tension between cartilage, ligaments, skin, and bones. Other variants of proteins consist of Cas9, which leveraging CRISPR sequences as guidance, function like scissors to cut and paste portions of DNA; antifreeze proteins, whose 3D structure enables them to bind to ice crystals and avert organisms from freezing, and ribosomes, which function like a programmed factory assembly line, assisting to develop proteins themselves. 

Genes are the recipes for these proteins – and they are subsequently encoded in our DNA makeup. An error within the genetic recipe may have the outcome of a malformed protein, which could in turn have the outcome of illness or death of an organism. Several illnesses, hence, are basically linked to proteins. But just because you are aware of the genetic recipe for a protein doesn’t imply you are automatically aware of its shape. Proteins are made up of chains of amino acids (also referred to as amino acid residues). However, DNA only contains data about the sequence regarding amino acids, not how they go about folding into shape. The larger the protein, the more tough it is to go about modelling, as there are more interactions amongst amino acids to be considered. As illustrated by Levinthal’s paradox, it would take more than the age of the known universe to arbitrarily enumerate all potential configurations of a conventional protein prior to reaching the actual 3D structure – however, proteins themselves fold spontaneously, within a matter of milliseconds. Forecasting how these chains will fold into the complex 3D structure of a protein is what’s referred to as the protein folding problem – a hurdle that researchers have been hard at work on for decades. This unresolved issue has already provided inspiration to numerous advancements, from spurring IBM’s initiatives within supercomputing (BlueGene) to novel citizen science initiatives (Folding@Home and Foldit) to brand new realms of engineering, like rational protein design. 

Researchers have long been fascinated in determining the structures of proteins as a protein’s form is thought to dictate its functionality. After a protein’s shape has been comprehended, its function within the cell can be open to guessing, and researchers can produce drugs that function with the protein’s unique shape. 

Over the previous half-a-century, scientists have been able to decide shapes of proteins in laboratories leveraging experimental strategies like cryo-electron microscopy, nuclear magnetic resonance, and X-ray crystallography, but every method is dependent on a lot of trial and error, which can take years of input, and cause expenditures of tens or hundreds or thousands of dollars for each protein structure. This is why scientists are looking to artificial intelligence methodologies as an alternative to this protracted and tedious process for tough to crack proteins. The capability to forecast a protein’s shape through computing from only its genetic code – over determining it via expensive experimentation – could assist in quickening up research. 

Luckily, the domain of genomics is quite data flush owing to the swift reduction in the expenses of genetic sequencing. As an outcome, deep learning strategies to the forecasting problem that are reliant on genomic information have become more and more popular and widespread over the course of the previous few years. To spur research and quantify progress on the latest methods for enhancing the precision of forecasting, a biennial international competition referred to as CASP (Critical Assessment of Protein Structure Prediction) was setup in 1994, and has since become the gold standard foe evaluating predictive strategies. We have a massive debt to pay to decades of historical work and research by the CASP organizers, and additionally to the thousands of experimentalists whose structures facilitate this kind of evaluation. 

AlphaFold was the outcome of DeepMind’s research and inquiry into this issue, which was provided to CASP13. The team concentrated particularly on the issue of modelling target shapes from the ground up, without leveraging prior solved proteins as a base. They accomplished a high degree of precision when forecasting the physical attributes of a protein structure, and then leveraged two distinct methodologies to develop predictions of complete protein structures. 

Both of these methodologies were reliant on deep neural networks that received training to forecast attributes of the protein from its genetic sequence. The attributes their networks forecast are: a) the distances between pairings of amino acids and b) the angles amongst chemical bonds that bring together those amino acids. The first progression is an advancement on typically leveraged strategies that estimate if parings of amino acids are close to one another. 

They went about training a neural network to forecast a distribution of distances between each pairing of residues in a protein. These probabilities were then brought together into a score that went about estimating how precise a proposed protein structure is. A separate neural network also received training that leverages all distances in aggregate to go about estimating how close the proposed structure is to the correct solution. 

Leveraging these scoring functions, they were able to look around the protein landscape to identify structures that correlated to their forecasts. The first methodology developed on strategies typically leveraged in structural biology, and repeatedly substituted pieces of a protein structure with new protein fragments. They went about training a generative neural network to invent fresh fragments, which were leveraged to improve the score of the proposed protein structure on an ongoing basis. 

The second methodology was concerned with score optimization through gradient descent – a mathematical strategy typically leveraged within machine learning for making small, gradual improvements, which had the outcome of highly precise structures. This strategy was applied to entire protein chains over pieces that must have folding separately, prior to being assembled into a larger structure, in order to go about simplifying the forecasting procedure. 

The AlphaFold version leveraged at CASP13 is available on Github for anybody with interest, or anyone who wishes to replicate their protein folding outcomes. 

While it is thrilling that the protein folding model is a success, there’s still a lot of work to be conducted in the domain of protein biology, and the researchers are more than thrilled to persist with their initiatives in the domain. They’ve made a commitment to setting up ways that artificial intelligence can contribute to fundamental scientific discovery, with the idea of influencing practical-world impact. This strategy might function to eventually enhance our comprehension of our bodies and how it functions, facilitating researchers to target and develop new, efficient cures for illnesses in a more effective fashion. Researchers have just mapped structures for approximately 50% of the cumulative proteins made by human cells. Some rare illnesses consist of mutations within a singular gene, having the outcome of a malformed protein which can have considerable impacts on the health of the entire organism. A utility such as AlphaFold might assist rare illness researchers forecast the shape of a protein of interest swiftly and economically. As researchers obtain more know-how with regards to the shapes of proteins and how they function through simulating and models, this methodology may ultimately assist us in contributing to efficient and effective drug discovery, while also minimizing the expenditures connected with experimentation. The hope is that artificial intelligence will be good for illness research, and eventually enhance the quality of life for millions of patients globally. 

However, prospective advantages are not limited to health only – comprehending protein folding will help in protein design which could unlock a massive number of advantages. For instance, progression in biodegradable enzymes – which can be facilitated by protein design could assist in management of pollutants such as plastic and oil, enabling us to break down waste in a manner that is more eco-friendly. As a matter of fact, researchers have already started engineering bacteria to secrete proteins that will make waste produce biodegradable, and simpler to go about processing. 

The success of this preliminary foray into protein folding is suggestive of how machine learning frameworks can integrate with divergent sources of data to assist researchers come up with innovative solutions to complicated issues at speed. Just as we’ve observed how artificial intelligence can assist individuals master complicated games through frameworks like AlphaGo and AlphaZero, the hope is that at some point in time, artificial intelligence breakthroughs will help function as a platform to progress our comprehension of basic scientific issues, as well. 

It’s thrilling to witness these preliminary indicators of progress within protein folding, depicting the utility of artificial intelligence for scientific discovery. Although there’s a lot more research to be conducted prior to being able to influence a quantifiable impact on treating illnesses, handling waste, and more, we are well aware that the prospects are enormous. With a devoted team concentrated on diving into how machine learning can progress the world of science, many are looking ahead to observing the many ways these technologies can be a difference maker.