Category Archives: Deep Mind
Researchers at Google DeepMind claim to have built an AI model that can pinpoint which genetic mutations are likely to cause disease, according to a new study in the journal Science.
The new model, dubbed AlphaMissense, is an adaptation of AlphaFold, the DeepMind breakthrough that, back in 2020, finally cracked the protein folding problem, which had baffled the scientific community for years. According to the new study, AlphaMissense is "fine-tuned" on "human and primate" genetic variance and specifically trained to identify "missense" mutations, or genetic mutations that take place in a single letter of DNA code.
Though some missense mutations are completely benign any given human has 9,000 or so missense alleles in their DNA others can cause serious disease; sickle cell anemia, cystic fibrosis, and cancer, as DeepMind noted in a Tuesday blog post, all stem from missense genes specifically. And yet, despite the fact that missense mutations and other DNA abnormalities are a primary driver of illness, humans have only been able to independently classify a minuscule 0.1 percent of missense genes as good or bad.
Until now, that is. According to DeepMind's new study, this new AI model has been able to identify a staggering 71 million missense mutations, and from there has been able to predictively classify 89 percent of these variations as "either likely benign or likely pathogenic." Tens of millions of these predictions have since been spun into a vast online database for physicians, genetic researchers, and other diagnostic experts, who according to Google will hopefully be able to use this new resource to find and diagnose various illnesses including exceedingly rare disorders and ultimately kickstart the development of what it called "life-saving treatments."
"Today, we're releasing a catalog of 'missense' mutations where researchers can learn more about what effect they may have," DeepMind penned in its Tuesday blog, adding later that "by using AI predictions, researchers can get a preview of results for thousands of proteins at a time, which can help to prioritize resources and accelerate more complex studies."
But while that allsounds great, the news has been met with mixed reactions from the scientific community.
Some folks, like Ewan Birney, deputy director general of the European Molecular Biology Laboratory, told the BBC that AlphaMissense is a "big step forward," arguing that the model "will help clinical researchers prioritize where to look to find areas that could cause disease." But others, like Ben Lehner, a senior group leader in human genetics at the UK's Wellcome Sanger Institute, were more hesitant, telling The Guardian that the black-box aspect of the tech concerns him.
"One concern about the DeepMind model is that it is extremely complicated," Lehner told The Guardian. "A model like this may turn out to be more complicated than the biology it is trying to predict," he added, noting that because doctors don'treallyunderstand how models like AlphaMissense actually work, using their predictions to make diagnostic choices might prove problematic.
"It's humbling to realize that we may never be able to understand how these models actually work. Is this a problem?" Lehner told the Guardian. "It may not be for some applications, but will doctors be comfortable making decisions about patients that they don't understand and can't explain?"
That said, though, Lehner did note that the DeepMind model "does a good job of predicting what is broken," and that "knowing what is broken is a good first step." Still, he says, you "also need to know how something is broken if you want to fix it."
AlphaMissense, of course, doesn't quite go that far just yet. After all, genetics is endlessly complicated. As Heidi Rehm, who directs the clinical laboratory at the Broad Institute of MIT and Harvard, told The MIT Technology Review, computer predictions are only "one piece of evidence" from which physicians can make diagnostic calls.
"The models are improving, but none are perfect, and they still don't get you to pathogenic or not," Rehm continued, reportedly noting that she was "disappointed" to see Google exaggerate the medical efficacy of its new product.
So, mixed reviews. But even if DeepMind's purported step forward isn't quite as big as the venture has cracked it up to be, it may well be a step forward nonetheless. Only time will tell but in the meantime, if you're in the business of diagnosing genetic disorders, maybe just take AlphaMissense's predictions with a grain of salt.
More on healthcare innovations: Biotech Company Says It's Implanted Dopamine-making Cells in Patients' Brains
Researchers at Google DeepMind, the tech giant's artificial intelligence arm, on Tuesday introduced a tool that predicts whether genetic mutations are likely to cause harm, a breakthrough that could help research into rare diseases. The findings are "another step in recognising the impact that AI is having in the natural sciences," said Pushmeet Kohli, vice president for research at Google DeepMind.
The tool focuses on so-called "missense" mutations, where a single letter of the genetic code is affected. A typical human has 9,000 such mutations throughout their genome; they can be harmless or cause diseases such as cystic fibrosis or cancer, or damage brain development. To date, four million of these mutations have been observed in humans, but only two percent of them have been classified, either as disease-causing or benign.
In all, there are 71 million such possible mutations. The Google DeepMind tool, called AlphaMissense, reviewed these mutations and was able to predict 89 percent of them, with 90 percent accuracy. A score was assigned to each mutation, indicating the risk of it causing disease (otherwise referred to as pathogenic).
The result: 57 percent were classified as probably benign, and 32 percent as probably pathogenic -- the remainder being uncertain. The database was made public and available to scientists, and an accompanying study was published on Tuesday in the journal Science.
AlphaMissense demonstrates "superior performance" than previously available tools, wrote experts Joseph Marsh and Sarah Teichmann in an article also published in Science. "We should emphasize that the predictions were never really trained or never really intended to be used for clinical diagnosis alone," said Jun Cheng of Google DeepMind.
"However, we do think that our predictions can potentially be helpful to increase the diagnosed rate of rare disease, and also potentially to help us find new disease-causing genes," Cheng added. Indirectly, this could lead to the development of new treatments, the researchers said.
The tool was trained on the DNA of humans and closely related primates, enabling it to recognize which genetic mutations are widespread. Cheng said the training allowed the tool to input "millions of protein sequences and learn what a regular protein sequence looks like."
It then could identify a mutation and its potential for harm. Cheng compared the process to learning a language. "If we substitute a word from an English sentence, a person that is familiar with English can immediately see whether this word substitution will change the meaning of the sentence or not."
AlphaMissense promises to predict how dangerous a genetic mutation is.
Researchers at DeepMind, one of Googles artificial intelligence (AI) companies, presented a new tool that predicts whether or not genetic mutations are potentially pathogenic, a breakthrough that could help research into rare diseases.
The tool focuses on so-called missense mutations, where there is a single nucleotide change, meaning one letter of the DNA code is affected.
If you compare DNA to the alphabet, a missense mutation is like a typo, and its one that can change the resulting amino acid.
An individual has around 9,000 of these mutations, most of which are benign. However, some of these mutations are responsible for diseases such as cystic fibrosis, sickle-cell anaemia, or cancer, DeepMind said.
Currently, four million of these mutations have been observed in humans, but only two per cent of them have been classified as either pathogenic or benign.
In total, there are 71 million possible mutations of this type. They were examined by the AlphaMissense tool, which was able to categorise 89 per cent of them, the company said.
Each mutation was given a score between 0 and 1, indicating the risk of it being pathogenic, i.e. causing disease. AlphaMissense predicted that 57 per cent were likely benign and 32 per cent were likely pathogenic, with the rest remaining uncertain.
The database has been made public and available to all scientists through GitHub, a platform to store and share computer code. A study on the findings was published on Tuesday in the journal Science.
The authors demonstrate superior performance by AlphaMissense," wrote experts Joseph Marsh and Sarah Teichmann in an article also published in Science.
The tool has been trained on a database of human and primate DNA, enabling it to recognise which genetic mutations are prevalent.
Jun Cheng, a scientist at Google DeepMind, explained that the tool can tell whether a protein sequence is worrying or not.
He added that the predictions could increase the rate of diagnosis of rare diseases and help to find new genes involved in the diseases.
Indirectly, this could lead to the development of new treatments, the researchers claim, warning, however, that AlphaMissense should not be used on its own to make a diagnosis.
AlphaMissense was based on AlphaFold, another machine learning program presented by Google DeepMind in 2018, which had published the largest database of proteins with over 200 million structures available.
AIs existential threat is a completely bonkers distraction because there are like 101 more practical issues to talk about, top founder in the field…
DeepMind is headquartered in London Getty Images
Asking Google's PaLM 2 artificial intelligence to "take a deep breath and work on this problem step by step" was the most effective prompt tested by Google DeepMind researchersto improve AI's accuracy,according to a study published on September 7.
Their study aimed to see how simple prompts could improve the performance of large language models, like ChatGPT's GPT-4 or Google's PaLM 2. It is not clear how many prompts were used during the study.
When prompted with the phrase, researchers found that Google's PaLM 2 model was 80% accurate while responding to a set of grade school math problems.
The creators of the over 8,500 math problems used in the study wrote that a"bright middle school student" should be able to solve all of them.
Without the prompt, the model was merely 34% accurate in answering the math problems, per the study. Meanwhile, prompting the AI with "let's think step by step" saw an increase in accuracy to 71%.
The researchers automated the process of testing a large number of different phrases with a variety of AI models to understand which prompts would work best.
To add context to these findings, a 2022 joint study byresearchersat Google andthe University of Tokyofound that getting large language models to"think step by step" improved their accuracy.
ChatGPT's launch in November has sparked curiosity over how we should be speaking to AI toget the outcomes we desire. Some companies are even hiring "prompt engineers"who specialize in crafting questions and phrases for AI to improve its responses.
Anna Bernstein, a prompt engineer working for AI company Copy.ai, told Insider in August that to get the most out of theprompts, one shoulduse a thesaurus and pay closer attention toverbs.
Some groups have even put together "prompt libraries" to share phrases to get the most out of AI like OpenAI's Discord community, where developers share sample phrases to get ChatGPT to help with job interviews.
Google and the study's authors did not immediately respond to requests for comment from Insider, sent outside regular business hours.
Originally posted here:
Google researchers: Tell AI to 'take a deep breath' to improve accuracy - Business Insider
The latest iteration of a legacy
Founded at the Massachusetts Institute of Technology in 1899, MIT Technology Review is a world-renowned, independent media company whose insight, analysis, reviews, interviews and live events explain the newest technologies and their commercial, social and political impact.
This protein, whose structure was predicted by AlphaFold, is part of the nuclear pore complex, which is a gateway for molecules entering a cells nucleus and is a drug target.Credit: DeepMind
After Google Deepminds AlphaFold proved that it could predict the 3D shapes of proteins with high accuracy in 2020, chemists became excited about the promise of using the open-source artificial-intelligence (AI) programme to discover drugs more quickly and cheaply. Most drugs work by binding to various sites on proteins, and AlphaFold could predict the structures for proteins that scientists previously knew little about.
Last month, the biotechnology firm Recursion, based in Salt Lake City, Utah, announced that it had calculated how 36 billion potential drug compounds could bind to more than 15,000 human proteins whose structures were predicted by AlphaFold. To pull off the massive computation, Recursion used its own AI tool, MatchMaker, that matched binding pockets on the predicted structures with suitably shaped small molecules, or ligands, from a database called Enamine Real Space.
AI tools are designing entirely new proteins that could transform medicine
Lots of people have predicted how molecules would bind with proteins, says Chris Gibson, Recursions co-founder and chief executive, but this many predictions is pretty unprecedented.
But not everyone is as bullish about AlphaFold revolutionizing drug discovery at least, not yet. In a paper published in eLife the day before Recursions announcement1, a team of scientists at Stanford University in California showed that AlphaFolds prowess at predicting protein structures doesnt yet translate into solid leads for ligand binding.
Models like AlphaFold are really good with [protein] structures, but we need to put some thought into how were going to use them for drug discovery, says Masha Karelina, a biophysicist at Stanford and co-author of the paper.
Others who spoke to Nature agree that this type of effort offers impressive amounts of data, but they arent yet sure about its quality. Biotech announcements such as the one from Recursion arent typically accompanied by validation data confirmation from laboratory experiments that a model has accurately predicted binding. The calculated interactions are also based on predicted, rather than experimentally determined, protein structures which might not contain the atomic-level resolution that drug developers need to pinpoint where the strongest binding might occur. Whats more, the sheer number of predicted interactions (Recursion predicted 2.8 quadrillion) means that even a small percentage of false-positive hits can lead to costly delays while scientists waste valuable time trying to validate them, says Brian Shoichet, a pharmaceutical chemist at the University of California, San Francisco.
The result, Shoichet says, is a lot of excitement, but also a lot of questions.
The idea of using computational tools in drug discovery is to make it easier, faster and cheaper to play with all the parameters that make a good drug, says Vsevolod Katritch, a computational biologist at the University of Southern California in Los Angeles. By using AI models to find leads, a drug company might need to test only a few hundred compounds in the lab, instead of thousands. This can shave millions off the cost and bring a compound to market in years instead of decades.
What's next for AlphaFold and the AI protein-folding revolution
AlphaFold and similar programs, such as RoseTTAFold, which was developed by an international team led by researchers at the University of Washingtons Institute of Protein Design, promise to shake up the pharmaceutical industry further because the structures of many human proteins had been lacking, making it difficult to find treatments for some diseases. The programmes have become so good at predicting 3D protein shapes that of the 200 million protein structures deposited into a database last year, the European Molecular Biology Laboratorys European Bioinformatics Institute deemed 35% to be highly accurate as good as experimentally determined structures and another 45% accurate enough for some applications.
On the surface, making the leap from AlphaFolds and RoseTTAFolds protein structures to the prediction of ligand binding doesnt seem like such a big one, Karelina says. She initially thought that modelling how a small molecule docks to a predicted protein structure (which usually involves estimating the energy released during ligand binding) would be easy . But when she set out to test it, she found that docking to AlphaFold models is much less accurate than to protein structures that are experimentally determined1. Karelinas still not 100% sure why, but she thinks that small variations in the orientation of amino-acid side chains in the models versus the experimental structures could be behind the gap. When drugs bind, they can also slightly alter protein shapes, something that AlphaFold structures dont reflect.
Laksh Aithani, chief executive and co-founder of London-based Charm Therapeutics, agrees with Karelinas findings that RoseTTAFold and AlphaFold dont perform well when determining small-molecule docking.
Charm is trying a different way of evaluating proteindrug binding. The technique uses an AI tool, called DragonFold, that is built on a RoseTTAFold backbone. It models the 3D shape of the protein and ligand bound together, which Aithani says allows Charm to account for changes in protein shape that occur with ligand binding and to modify the would-be drug to create tighter, more selective binding. The effort isnt far enough along for Aithani to reveal many details, but he says the project has attracted the interest of pharmaceutical firm Bristol Myers Squibb, based in Lawrenceville, New Jersey.
In the end, the challenge for these groups, says Shoichet, isnt to design a model that will identify how well molecules bind, but to create a system that can identify compounds that bind strongly to proteins about which little is known. To make progress, validation in the lab is necessary, he says.
A Pandoras box: map of protein-structure families delights scientists
Industry should be able to do the validation, says Bonnie Berger, a mathematician at the Massachusetts Institute of Technology in Cambridge. At the moment, however, if industry is doing it, it isnt sharing that data.
Theres a lack of transparency from companies like Recursion, who make predictions without fully sharing their methods or results. Its a problem for me and for the field, she says.
Recursion responds that it has shared validation data on MatchMaker in two studies: one in Scientific Reports in 20212, and one in the Journal of Chemical Information and Modeling earlier this year3.
Sharing these exciting technical milestones in real time as they occur is our way to share how we are thinking about drug discovery with the community and the broader general public, says Recursion spokesperson Ryan Kelly.
Berger says that competitions such as the one that put AlphaFold on the map could not only help drive drug discovery forward, but also shed more light on industrys methods. AlphaFold made headlines when it won the biennial Critical Assessment of protein Structure Prediction (CASP) contest in 2020, in which researchers had to test their prediction models against a set of proteins for which structures were experimentally determined, but not yet publicly released. In the same way, an AI tools results for drugprotein interactions could be compared with lab results for binding.
Theres a huge amount of effort going on to harness models such as AlphaFold for drug discovery, Shoichet says. But things are still just ramping up.
Read the original here:
AlphaFold touted as next big thing for drug discovery but is it? - Nature.com
Every day, we handpick the 5 Top Stories stock market investors need to know. In 5 minutes, youll learn the stocks, CEOs, and money managers moving markets.
Online grocer Instacarts [CART] shares started trading at $42 on Tuesday, having sold for $30, raising $660m and pushing the firms valuation to more than $11bn. This may be a lot less than its pandemic-era valuation of $39bn, but the IPO has been hailed as sign of a return to health for the tech start-up space, reported Bloomberg. Nevertheless, Instacart shares fell by nearly 5% in premarket trading on Wednesday, in a possible sign that its debut rally is already waning.
Researchers at DeepMind, Googles [GOOGL] artificial intelligence (AI) research lab, have used an AI tool to predict whether genetic mutations will prove harmful. An article published in the journal Science details how the tool, called AlphaMissense, evaluated all of the 71 million missense mutations, wherein a single letter of the genetic code varies. On Tuesday, Mark Zuckerberg announced that the Chan Zuckerberg Initiative is to build one of the largest computing systems dedicated to non-profit life sciences.
The fact that Huaweis latest smartphone has a 5G chip produced by Chinese semiconductor maker SMIC [0981.HK] raises questions about the effectiveness of US sanctions, and may also constitute a headwind for Apple [AAPL] in China, CNBC reported. Meanwhile, Beijing has accused Washington of infiltrating Huawei servers as far as back as 2009, while the German interior ministry is moving to ban components made by Huawei and ZTE [0763.HK] from its 5G network, according to Bloomberg.
This week, the EU will begin finalising the text of a new law aimed at protecting gig workers among them, those who drive for Uber [UBER]. Anabel Daz, Head of EMEA Mobility at Uber, has duly warned that being obliged to treat such workers as employees could cause the firm to shut down in hundreds of cities across the bloc, and to drive up its prices for consumers by some 40%.
The Wall Street Journal reported that federal prosecutors are looking into perks that Tesla [TSLA] may have provided CEO Elon Musk going back as far as 2017. This forms part of their investigation into the plan for a house, known as Project 42, which the company was allegedly going to build for Musk, who has refuted the allegation. Elsewhere, in a conversation with Israeli Prime Minister Benjamin Netanyahu, Musk mentioned he is weighing charging users to access X, formerly Twitter.
Disclaimer Past performance is not a reliable indicator of future results.
CMC Markets is an execution-only service provider. The material (whether or not it states any opinions) is for general information purposes only, and does not take into account your personal circumstances or objectives. Nothing in this material is (or should be considered to be) financial, investment or other advice on which reliance should be placed. No opinion given in the material constitutes a recommendation by CMC Markets or the author that any particular investment, security, transaction or investment strategy is suitable for any specific person.
The material has not been prepared in accordance with legal requirements designed to promote the independence of investment research. Although we are not specifically prevented from dealing before providing this material, we do not seek to take advantage of the material prior to its dissemination.
CMC Markets does not endorse or offer opinion on the trading strategies used by the author. Their trading strategies do not guarantee any return and CMC Markets shall not be held responsible for any loss that you may incur, either directly or indirectly, arising from any investment based on any information contained herein.
*Tax treatment depends on individual circumstances and can change or may differ in a jurisdiction other than the UK.
Continue reading for FREE
Success! You have successfully signed up.
Tech giant Alphabet (GOOG -0.08%) (GOOGL -0.15%) dominates the search market with its Google search engine and is a powerhouse in streaming video with YouTube. In addition, its cloud division has carved out a solid third-place position in the cloud computing infrastructure market, recently becoming profitable. And the company is also investing in several "moonshot" projects, such as Waymo self-driving cars, which it groups into a segment called "other bets."
In short, Alphabet is a great business. Yet while the release of OpenAI's ChatGPT last year brought up competitive concerns, I would actually lean more toward Alphabet being in a stronger position a few years rather than a weaker one.
That's because of three innovations we already know are in its pipeline, not to mention others we might not even know about yet.
At the beginning of this year, some thought ChatGPT could put several of Alphabet's businesses in jeopardy. Notably, some wondered if a ChatGPT-enhanced Bing search engine would lure significant traffic away from Google, or if the answers generated by chatbots would be able as monetizable as traditional search queries.
While those questions are still out there, Google's share of the search business hasn't been affected much thus far. Moreover, Google's engineers were the ones who actually cracked the code on transformer technology, the architecture which allows AI models to train themselves without the need for human data labeling, and which has led to recent breakthroughs.
So Google should be able to compete in AI effectively. Earlier this year, the company combined its DeepMind and Google Brain AI laboratories into one unit. Now, the company is on the cusp of releasing its own large language model, Gemini, to take on OpenAI and others.
Gemini will attempt to take the best technologies under both DeepMind and Google Brain and combine them in a multimodal model. A multimodal model integrates not only text, but also images and other types of data. The version of ChatGPT you are probably familiar with is not multimodal; however, there are reports that OpenAI is working on a multimodal version of ChatGPT.
According to recent interviews with DeepMind CEO Demis Hassabis, Gemini will have innovative new reasoning, problem-solving, and reinforcement-learning capabilities, and will also be available in different sizes and capabilities for customers. Google is in the process of testing the model with a handful of outside developers, so Gemini should go into beta testing soon.
While we can't know how well Gemini will compete, Hassabis said early results were "promising." And Google not only has deep AI chops, but also unique computing infrastructure expertise from its many years of running Google Search. So, one can easily imagine a scenario in which Google leads the AI field with differentiated technology, just as it has long led search.
Three years from now, that could open up new revenue streams, including better search, a monetized chatbot, and/or more services on Google Cloud. Last quarter, Alphabet CEO Sundar Pichai noted that 70% of AI unicorns used Google Cloud. Meanwhile, the cloud unit could become a major profit center soon, as it just flipped to profitability for the first time in Q1 2023.
Image source: Getty Images.
Most people don't immediately think of Pixel phones and tablets as a core part of Alphabet's business, but the company has been improving its offerings and supporting the Pixel brand with more resources of late.
Results are starting to show. According to research firm Canalys, Google Pixel grew 20% in the first quarter, and reached 2.46% market share of Android phones in North America. That may not sound like much, but that made Pixel the third-most-used Android smartphone in North America. Meanwhile, second-place Motorola, which is now produced by Chinese company Lenovo, fell 40% in the first quarter, to a 4.82% market share.
That seems to indicate Google is gaining momentum in the phone market, as Motorola weakens and other Chinese smartphone brands exit the market. Furthermore, Korea's LG decided to stop making smartphones in 2021, leaving a bigger opening for Google to take market share.
And Pixel is doing even better in some overseas markets. For instance, Counterpoint Research reports that Pixel is now actually the leading Android brand in Japan, with a market share of about 9%.
Investors should also expect Pixel devices to get even better in the coming years. This is because the Pixel team is working on its own proprietary mobile system on a chip (SoC). Like many other mobile handset companies, Pixel started off using third-party hardware. However, Google has the financial resources to make Pixel's inner parts more proprietary, and it's now at work on a fully customized SoC.
The most recent Pixel smartphone actually uses some proprietary chip technology, but currently, Google is combining this with IP from Samsung's Exynos chipsets. However, Google is targeting the 2025 Pixel as the smartphone that will have a fully customized SoC. Given Google's success with in-house chipmaking for its tensor processing units (TPUs), which it uses in its cloud servers, the company could begin to make waves in the smartphone market a few years from now.
Finally, there was an interesting announcement made on Alphabet's last earnings call. CFO Ruth Porat, who has been in that role since 2015, announced she would be adding another title -- chief investment officer. More specifically, that new role will have her overseeing the "other bets" division, while also working with overseas policymakers to "unlock economic growth via technology and investment."
It's a bit unclear what that means, but Porat has been credited with helping Alphabet become more financially disciplined in recent years. And the "other bets" division could certainly use more discipline, as it lost $5.3 billion in 2021 and $6.1 billion in 2022.
With Porat now overseeing that division, it may see significant bottom-line improvements a few years from now, either through better revenue growth or perhaps cost savings from culling unprofitable projects.
Overall, Alphabet looks likely to be a steady earnings compounder for years to come. Meanwhile, it still trades at a reasonable valuation of around 20 times next year's earnings estimates -- and the stock is actually cheaper than that when factoring in Alphabet's $118 billion in cash and the depressive effect of its "other bets" losses.
Investors should look forward to more profitable growth from Alphabet over the next three years.
Read this article:
Where Will Alphabet Be in 3 Years? - The Motley Fool
A patient receives treatment for cystic fibrosis, a disease linked in some cases to missense mutations.Credit: Burger/Phanie/Science Photo Library
Google DeepMind has wielded its revolutionary protein-structure-prediction AI in the hunt for genetic mutations that cause disease.
A new tool based on the AlphaFold network can accurately predict which mutations in proteins are likely to cause health conditions a challenge that limits the use of genomics in healthcare.
The AI network called AlphaMissense is a step forward, say researchers who are developing similar tools, but not necessarily a sea change. It is one of many techniques in development that aim to help researchers, and ultimately physicians, to interpret peoples genomes to find the cause of a disease. But tools such as AlphaMissense which is described in a 19 September paper in Science1 will need to undergo thorough testing before they are used in the clinic.
A Pandoras box: map of protein-structure families delights scientists
Many of the genetic mutations that directly cause a condition, such as those responsible for cystic fibrosis and sickle-cell disease, tend to change the amino acid sequence of the protein they encode. But researchers have observed only a few million of these single-letter missense mutations. Of the more than 70 million possible in the human genome, only a sliver have been conclusively linked to disease, and most seem to have no ill effect on health.
So when researchers and doctors find a missense mutation theyve never seen before, it can be difficult to know what to make of it. To help interpret such variants of unknown significance, researchers have developed dozens of different computational tools that can predict whether a variant is likely to cause disease. AlphaMissense incorporates existing approaches to the problem, which are increasingly being addressed with machine learning.
The network is based on AlphaFold, which predicts a protein structure from an amino-acid sequence. But instead of determining the structural effects of a mutation an open challenge in biology AlphaMissense uses AlphaFolds intuition about structure to identify where disease-causing mutations are likely to occur within a protein, Pushmeet Kohli, DeepMinds vice-president of Research and a study author, said at a press briefing.
AlphaMissense also incorporates a type of neural network inspired by large language models like ChatGPT that has been trained on millions of protein sequences instead of words, called a protein language model. These have proven adept at predicting protein structures and designing new proteins. They are useful for variant prediction because they have learned which sequences are plausible and which are not, iga Avsec, the DeepMind research scientist who co-led the study, told journalists.
Foldseek gives AlphaFold protein database a rapid search tool
DeepMinds network seems to outperform other computational tools at discerning variants known to cause disease from those that dont. It also does well at spotting problem variants identified in laboratory experiments that measure the effects of thousands of mutations at once. The researchers also used AlphaMissense to create a catalogue of every possible missense mutation in the human genome, determining that 57% are likely to be benign and that 32% may cause disease.
AlphaMissense is an advance over existing tools for predicting the effects of mutations, but not a gigantic leap forward, says Arne Eloffson, a computational biologist at the University of Stockholm.
Its impact wont be as significant as AlphaFold, which ushered in a new era in computational biology, agrees Joseph Marsh, a computational biologist at the MRC Human Genetics Unit in Edinburgh, UK. Its exciting. Its probably the best predictor we have right now. But will it be the best predictor in two or three years? Theres a good chance it wont be.
Computational predictions currently have a minimal role in diagnosing genetic diseases, says Marsh, and recommendations from physicians groups say that these tools should provide only supporting evidence in linking a mutation to a disease. AlphaMissense confidently classified a much larger proportion of missense mutations than have previous methods, says Avsec. As these models get better than I think people will be more inclined to trust them.
Yana Bromberg, a bioinformatician at Emory University in Atlanta, Georgia, emphasizes that tools such as AlphaMissense must be rigorously evaluated using good performance metrics before ever being applied in the real-world.
For example, an exercise called the Critical Assessment of Genome Interpretation (CAGI) has benchmarked the performance of such prediction methods for years against experimental data that has not yet been released. Its my worst nightmare to think of a doctor taking a prediction and running with it, as if its a real thing, without evaluation by entities such as CAGI, Bromberg adds.