A catalogue of genetic mutations to help pinpoint the cause of diseases - Related to imperfect, a, synthid, watermarking, mastering

A catalogue of genetic mutations to help pinpoint the cause of diseases

Research A catalogue of genetic mutations to help pinpoint the cause of diseases Share.

New AI tool classifies the effects of 71 million ‘missense’ mutations Uncovering the root causes of disease is one of the greatest challenges in human genetics. With millions of possible mutations and limited experimental data, it’s largely still a mystery which ones could give rise to disease. This knowledge is crucial to faster diagnosis and developing life-saving treatments. Today, we’re releasing a catalogue of ‘missense’ mutations where researchers can learn more about what effect they may have. Missense variants are genetic mutations that can affect the function of human proteins. In some cases, they can lead to diseases such as cystic fibrosis, sickle-cell anaemia, or cancer. The AlphaMissense catalogue was developed using AlphaMissense, our new AI model which classifies missense variants. In a paper , we show it categorised 89% of all 71 million possible missense variants as either likely pathogenic or likely benign. By contrast, only [website] have been confirmed by human experts.

AI tools that can accurately predict the effect of variants have the power to accelerate research across fields from molecular biology to clinical and statistical genetics. Experiments to uncover disease-causing mutations are expensive and laborious – every protein is unique and each experiment has to be designed separately which can take months. By using AI predictions, researchers can get a preview of results for thousands of proteins at a time, which can help to prioritise resources and accelerate more complex studies. We’ve made all of our predictions freely available for commercial and researcher use, and open sourced the model code for AlphaMissense.

AlphaMissense predicted the pathogenicity of all possible 71 million missense variants. It classified 89% – predicting 57% were likely benign and 32% were likely pathogenic.

What is a missense variant? A missense variant is a single letter substitution in DNA that results in a different amino acid within a protein. If you think of DNA as a language, switching one letter can change a word and alter the meaning of a sentence altogether. In this case, a substitution changes which amino acid is translated, which can affect the function of a protein. The average person is carrying more than 9,000 missense variants. Most are benign and have little to no effect, but others are pathogenic and can severely disrupt protein function. Missense variants can be used in the diagnosis of rare genetic diseases, where a few or even a single missense variant may directly cause disease. They are also significant for studying complex diseases, like type 2 diabetes, which can be caused by a combination of many different types of genetic changes. Classifying missense variants is an significant step in understanding which of these protein changes could give rise to disease. Of more than 4 million missense variants that have been seen already in humans, only 2% have been annotated as pathogenic or benign by experts, roughly [website] of all 71 million possible missense variants. The rest are considered ‘variants of unknown significance’ due to a lack of experimental or clinical data on their impact. With AlphaMissense we now have the clearest picture to date by classifying 89% of variants using a threshold that yielded 90% precision on a database of known disease variants. Pathogenic or benign: How AlphaMissense classifies variants AlphaMissense is based on our breakthrough model AlphaFold, which predicted structures for nearly all proteins known to science from their amino acid sequences. Our adapted model can predict the pathogenicity of missense variants altering individual amino acids of proteins. To train AlphaMissense, we fine-tuned AlphaFold on labels distinguishing variants seen in human and closely related primate populations. Variants commonly seen are treated as benign, and variants never seen are treated as pathogenic. AlphaMissense does not predict the change in protein structure upon mutation or other effects on protein stability. Instead, it leverages databases of related protein sequences and structural context of variants to produce a score between 0 and 1 approximately rating the likelihood of a variant being pathogenic. The continuous score allows people to choose a threshold for classifying variants as pathogenic or benign that matches their accuracy requirements.

An illustration of how AlphaMissense classifies human missense variants. A missense variant is input, and the AI system scores it as pathogenic or likely benign. AlphaMissense combines structural context and protein language modelling, and is fine-tuned on human and primate variant population frequency databases.

AlphaMissense achieves state-of-the-art predictions across a wide range of genetic and experimental benchmarks, all without explicitly training on such data. Our tool outperformed other computational methods when used to classify variants from ClinVar, a public archive of data on the relationship between human variants and disease. Our model was also the most accurate method for predicting results from the lab, which presents it is consistent with different ways of measuring pathogenicity.

AlphaMissense outperforms other computational methods on predicting missense variant effects.

Left: Comparing AlphaMissense and other methods’ performance on classifying variants from the Clinvar public archive. Methods shown in grey were trained directly on ClinVar and their performance on this benchmark are likely overestimated since some of their training variants are contained in this test set.

Right: Graph comparing AlphaMissense and other methods’ performance on predicting measurements from biological experiments.

Building a community resource AlphaMissense builds on AlphaFold to further the world’s understanding of proteins. One year ago, we released 200 million protein structures predicted using AlphaFold – which is helping millions of scientists around the world to accelerate research and pave the way toward new discoveries. We look forward to seeing how AlphaMissense can help solve open questions at the heart of genomics and across biological science. We’ve made AlphaMissense’s predictions freely available to both commercial and scientific communities. Together with EMBL-EBI, we are also making them more usable through the Ensembl Variant Effect Predictor. In addition to our look-up table of missense mutations, we’ve shared the expanded predictions of all possible 216 million single amino acid sequence substitutions across more than 19,000 human proteins. We’ve also included the average prediction for each gene, which is similar to measuring a gene's evolutionary constraint – this indicates how essential the gene is for the organism’s survival.

Examples of AlphaMissense predictions overlaid on AlphaFold predicted structures (red=predicted as pathogenic, blue=predicted as benign, grey=uncertain). Red dots represent known pathogenic missense variants, blue dots represent known benign variants from the ClinVar database.

Left: HBB protein. Variants in this protein can cause sickle cell anaemia.

Right: CFTR protein. Variants in this protein can cause cystic fibrosis.

Accelerating research into genetic diseases A key step in translating this research is collaborating with the scientific community. We have been working in partnership with Genomics England, to explore how these predictions could help study the genetics of rare diseases. Genomics England cross-referenced AlphaMissense’s findings with variant pathogenicity data previously aggregated with human participants. Their evaluation confirmed our predictions are accurate and consistent, providing another real-world benchmark for AlphaMissense. While our predictions are not designed to be used in the clinic directly – and should be interpreted with other information of evidence – this work has the potential to improve the diagnosis of rare genetic disorders, and help discover new disease-causing genes. Ultimately, we hope that AlphaMissense, together with other tools, will allow researchers to enhanced understand diseases and develop new life-saving treatments.

"The Thinking Part" by Daniel Warfield using MidJourney. All images by the author unless otherwise specified. Article originally made available on Int...

Introducing a context-based framework for comprehensively evaluating the social and ethical risks of AI systems.

Generative AI systems are already bei...

Quantum computers have the potential to revolutionize drug discovery, material design and fundamental physics — that is, if we can get them to work re...

Watermarking AI-generated text and video with SynthID

Technologies Watermarking AI-generated text and video with SynthID Share.

Announcing our novel watermarking method for AI-generated text and video, and how we’re bringing SynthID to key Google products Generative AI tools — and the large language model technologies behind them — have captured the public imagination. From helping with work tasks to enhancing creativity, these tools are quickly becoming part of products that are used by millions of people in their daily lives. These technologies can be hugely beneficial but as they become increasingly popular to use, the risk increases of people causing accidental or intentional harms, like spreading misinformation and phishing, if AI-generated content isn’t properly identified. That’s why last year, we launched SynthID, our novel digital toolkit for watermarking AI-generated content. Today, we’re expanding SynthID’s capabilities to watermarking AI-generated text in the Gemini app and web experience, and video in Veo, our most capable generative video model. SynthID for text is designed to complement most widely-available AI text generation models and for deploying at scale, while SynthID for video builds upon our image and audio watermarking method to include all frames in generated videos. This innovative method embeds an imperceptible watermark without impacting the quality, accuracy, creativity or speed of the text or video generation process. SynthID isn’t a silver bullet for identifying AI generated content, but is an essential building block for developing more reliable AI identification tools and can help millions of people make informed decisions about how they interact with AI-generated content. Later this summer, we’re planning to open-source SynthID for text watermarking, so developers can build with this technology and incorporate it into their models.

How text watermarking works Large language models generate sequences of text when given a prompt like, “Explain quantum mechanics to me like I’m five” or “What’s your favorite fruit?”. LLMs predict which token most likely follows another, one token at a time. Tokens are the building blocks a generative model uses for processing information. In this case, they can be a single character, word or part of a phrase. Each possible token is assigned a score, which is the percentage chance of it being the right one. Tokens with higher scores are more likely to be used. LLMs repeat these steps to build a coherent response. SynthID is designed to embed imperceptible watermarks directly into the text generation process. It does this by introducing additional information in the token distribution at the point of generation by modulating the likelihood of tokens being generated — all without compromising the quality, accuracy, creativity or speed of the text generation.

Pause video Play video SynthID adjusts the probability score of tokens generated by a large language model.

The final pattern of scores for both the model’s word choices combined with the adjusted probability scores are considered the watermark. This pattern of scores is compared with the expected pattern of scores for watermarked and unwatermarked text, helping SynthID detect if an AI tool generated the text or if it might come from other data.

A piece of text generated by Gemini with the watermark highlighted in blue.

The benefits and limitations of this technique SynthID for text watermarking works best when a language model generates longer responses, and in diverse ways — like when it’s prompted to generate an essay, a theater script or variations on an email. It performs well even under some transformations, such as cropping pieces of text, modifying a few words and mild paraphrasing. However, its confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten or translated to another language. SynthID text watermarking is less effective on responses to factual prompts because there are fewer opportunities to adjust the token distribution without affecting the factual accuracy. This includes prompts like “What is the capital of France?” or queries where little or no variation is expected like “recite a William Wordsworth poem”. Many currently available AI detection tools use algorithms for labeling and sorting data, known as classifiers. These classifiers often only perform well on particular tasks, which makes them less flexible. When the same classifier is applied across different types of platforms and content, its performance isn’t always reliable or consistent. This can lead to a text being mislabeled, which can cause problems, for example, where text might be incorrectly identified as AI-generated. SynthID works effectively on its own, but it can also be combined with other AI detection approaches to give enhanced coverage across content types and platforms. While this technique isn’t built to directly stop motivated adversaries like cyberattackers or hackers from causing harm, it can make it harder to use AI-generated content for malicious purposes.

How video watermarking works At this year’s I/O we presented Veo, our most capable generative video model. While video generation technologies aren't as widely available as image generation technologies, they’re rapidly evolving and it’ll become increasingly significant to help people know if a video is generated by an AI or not. Videos are composed of individual frames or still images. So we developed a watermarking technique inspired by our SynthID for image tool. This technique embeds a watermark directly into the pixels of every video frame, making it imperceptible to the human eye, but detectable for identification. Empowering people with knowledge of when they’re interacting with AI-generated media can play an significant role in helping prevent the spread of misinformation. Starting today, all videos generated by Veo on VideoFX will be watermarked by SynthID.

SynthID for video watermarking marks every frame of a generated video.

Bringing SynthID to the broader AI ecosystem SynthID’s text watermarking technology is designed to be compatible with most AI text generation models and for scaling across different content types and platforms. To help prevent widespread misuse of AI-generated content, we’re working on bringing this technology to the broader AI ecosystem. This summer, we’re planning to publish more about our text watermarking technology in a detailed research paper, and we’ll open-source SynthID text watermarking through our updated Responsible Generative AI Toolkit, which provides guidance and essential tools for creating safer AI applications, so developers can build with this technology and incorporate it into their models.

Responsibility & Safety How can we build human values into AI? Share.

Drawing from philosophy to identify fair principles for ethic...

Research Google DeepMind at ICLR 2024 Share.

Developing next-gen AI agents, exploring new modalities, and pioneering foundational l...

Impact MuZero, AlphaZero, and AlphaDev: Optimizing computer systems Share.

As part of our aim to build increasingly capable and gen...

Mastering Stratego, the classic game of imperfect information

Research Mastering Stratego, the classic game of imperfect information Share.

DeepNash learns to play Stratego from scratch by combining game theory and model-free deep RL Game-playing artificial intelligence (AI) systems have advanced to a new frontier. Stratego, the classic board game that’s more complex than chess and Go, and craftier than poker, has now been mastered. , we present DeepNash, an AI agent that learned the game from scratch to a human expert level by playing against itself. DeepNash uses a novel approach, based on game theory and model-free deep reinforcement learning. Its play style converges to a Nash equilibrium, which means its play is very hard for an opponent to exploit. So hard, in fact, that DeepNash has reached an all-time top-three ranking among human experts on the world’s biggest online Stratego platform, Gravon. Board games have historically been a measure of progress in the field of AI, allowing us to study how humans and machines develop and execute strategies in a controlled environment. Unlike chess and Go, Stratego is a game of imperfect information: players cannot directly observe the identities of their opponent's pieces. This complexity has meant that other AI-based Stratego systems have struggled to get beyond amateur level. It also means that a very successful AI technique called “game tree search”, previously used to master many games of perfect information, is not sufficiently scalable for Stratego. For this reason, DeepNash goes far beyond game tree search altogether. The value of mastering Stratego goes beyond gaming. In pursuit of our mission of solving intelligence to advance science and benefit humanity, we need to build advanced AI systems that can operate in complex, real-world situations with limited information of other agents and people. Our paper exhibits how DeepNash can be applied in situations of uncertainty and successfully balance outcomes to help solve complex problems. Getting to know Stratego Stratego is a turn-based, capture-the-flag game. It’s a game of bluff and tactics, of information gathering and subtle manoeuvring. And it’s a zero-sum game, so any gain by one player represents a loss of the same magnitude for their opponent. Stratego is challenging for AI, in part, because it’s a game of imperfect information. Both players start by arranging their 40 playing pieces in whatever starting formation they like, initially hidden from one another as the game begins. Since both players don't have access to the same knowledge, they need to balance all possible outcomes when making a decision – providing a challenging benchmark for studying strategic interactions. The types of pieces and their rankings are shown below.

Left: The piece rankings. In battles, higher-ranking pieces win, except the 10 (Marshal) loses when attacked by a Spy, and Bombs always win except when captured by a Miner.

Middle: A possible starting formation. Notice how the Flag is tucked away safely at the back, flanked by protective Bombs. The two pale blue areas are “lakes” and are never entered.

Right: A game in play, showing Blue’s Spy capturing Red’s 10.

Information is hard won in Stratego. The identity of an opponent's piece is typically revealed only when it meets the other player on the battlefield. This is in stark contrast to games of perfect information such as chess or Go, in which the location and identity of every piece is known to both players. The machine learning approaches that work so well on perfect information games, such as DeepMind’s AlphaZero, are not easily transferred to Stratego. The need to make decisions with imperfect information, and the potential to bluff, makes Stratego more akin to Texas hold’em poker and requires a human-like capacity once noted by the American writer Jack London: “Life is not always a matter of holding good cards, but sometimes, playing a poor hand well.” The AI techniques that work so well in games like Texas hold’em don’t transfer to Stratego, however, because of the sheer length of the game – often hundreds of moves before a player wins. Reasoning in Stratego must be done over a large number of sequential actions with no obvious insight into how each action contributes to the final outcome. Finally, the number of possible game states (expressed as “game tree complexity”) is off the chart compared with chess, Go and poker, making it incredibly difficult to solve. This is what excited us about Stratego, and why it has represented a decades-long challenge to the AI community.

The scale of the differences between chess, poker, Go, and Stratego.

Seeking an equilibrium DeepNash employs a novel approach based on a combination of game theory and model-free deep reinforcement learning. “Model-free” means DeepNash is not attempting to explicitly model its opponent’s private game-state during the game. In the early stages of the game in particular, when DeepNash knows little about its opponent’s pieces, such modelling would be ineffective, if not impossible. And because the game tree complexity of Stratego is so vast, DeepNash cannot employ a stalwart approach of AI-based gaming – Monte Carlo tree search. Tree search has been a key ingredient of many landmark achievements in AI for less complex board games, and poker. Instead, DeepNash is powered by a new game-theoretic algorithmic idea that we're calling Regularised Nash Dynamics (R-NaD). Working at an unparalleled scale, R-NaD steers DeepNash’s learning behaviour towards what’s known as a Nash equilibrium (dive into the technical details in our paper). Game-playing behaviour that results in a Nash equilibrium is unexploitable over time. If a person or machine played perfectly unexploitable Stratego, the worst win rate they could achieve would be 50%, and only if facing a similarly perfect opponent. In matches against the best Stratego bots – including several winners of the Computer Stratego World Championship – DeepNash’s win rate topped 97%, and was frequently 100%. Against the top expert human players on the Gravon games platform, DeepNash achieved a win rate of 84%, earning it an all-time top-three ranking.

Expect the unexpected To achieve these results, DeepNash demonstrated some remarkable behaviours both during its initial piece-deployment phase and in the gameplay phase. To become hard to exploit, DeepNash developed an unpredictable strategy. This means creating initial deployments varied enough to prevent its opponent spotting patterns over a series of games. And during the game phase, DeepNash randomises between seemingly equivalent actions to prevent exploitable tendencies. Stratego players strive to be unpredictable, so there’s value in keeping information hidden. DeepNash demonstrates how it values information in quite striking ways. In the example below, against a human player, DeepNash (blue) sacrificed, among other pieces, a 7 (Major) and an 8 (Colonel) early in the game and as a result was able to locate the opponent’s 10 (Marshal), 9 (General), an 8 and two 7’s.

In this early game situation, DeepNash (blue) has already located many of its opponent’s most powerful pieces, while keeping its own key pieces secret.

These efforts left DeepNash at a significant material disadvantage; it lost a 7 and an 8 while its human opponent preserved all their pieces ranked 7 and above. Nevertheless, having solid intel on its opponent’s top brass, DeepNash evaluated its winning chances at 70% – and it won. The art of the bluff As in poker, a good Stratego player must sometimes represent strength, even when weak. DeepNash learned a variety of such bluffing tactics. In the example below, DeepNash uses a 2 (a weak Scout, unknown to its opponent) as if it were a high-ranking piece, pursuing its opponent’s known 8. The human opponent decides the pursuer is most likely a 10, and so attempts to lure it into an ambush by their Spy. This tactic by DeepNash, risking only a minor piece, succeeds in flushing out and eliminating its opponent’s Spy, a critical piece.

The human player (red) is convinced the unknown piece chasing their 8 must be DeepNash’s 10 (note: DeepNash had already lost its only 9).

See more by watching these four videos of full-length games played by DeepNash against (anonymised) human experts: Game 1, Game 2, Game 3, Game 4.

“ The level of play of DeepNash surprised me. I had never heard of an artificial Stratego player that came close to the level needed to win a match against an experienced human player. But after playing against DeepNash myself, I wasn’t surprised by the top-3 ranking it later achieved on the Gravon platform. I expect it would do very well if allowed to participate in the human World Championships. Vincent de Boer, paper co-author and former Stratego World Champion.

Future directions While we developed DeepNash for the highly defined world of Stratego, our novel R-NaD method can be directly applied to other two-player zero-sum games of both perfect or imperfect information. R-NaD has the potential to generalise far beyond two-player gaming settings to address large-scale real-world problems, which are often characterised by imperfect information and astronomical state spaces. We also hope R-NaD can help unlock new applications of AI in domains that feature a large number of human or AI participants with different goals that might not have information about the intention of others or what’s occurring in their environment, such as in the large-scale optimisation of traffic management to reduce driver journey times and the associated vehicle emissions. In creating a generalisable AI system that’s robust in the face of uncertainty, we hope to bring the problem-solving capabilities of AI further into our inherently unpredictable world. Learn more about DeepNash by reading our paper in Science. For researchers interested in giving R-NaD a try or working with our newly proposed method, we’ve open-sourced our code.

Responsibility & Safety Introducing the Frontier Safety Framework Share.

Our approach to analyzing and mitigating future risks pose...

Note: This blog was first . Following the paper’s publication in Science on 8 Dec 2022, we’ve made minor updates to the text to...

Applying our AI research to help enrich the lives of billions of people around the world.

Building useful products with new technologies has always be...