Key Moments
🔬Bio-security is an AI Arms Race - Eric Nguyen (CEO, Radical Numerics)
Latent Space PodcastKey Moments
AI is unlocking the ability to write DNA, leading to breakthroughs in disease prediction and drug design, but also raising concerns about biosecurity risks and the potential for an AI arms race.
Key Insights
Radical Numerics' Omni model can outperform specialized models on disease-variant prediction and reason across biological sequences, even beginning to 'show its work' in designing molecules.
The development of AI that can both read and write DNA has significant implications for scientific discovery and understanding human health, with the potential to revolutionize disease treatment.
The creation of the first functional genome from scratch using AI, a bacteriophage, signifies a turning point, enabling the design of entirely novel organisms and raising ethical considerations.
Omni aims to unify various biological languages and modalities, moving beyond DNA to incorporate RNA, proteins, and epigenetics, with the ultimate goal of achieving general biological intelligence.
The ability to design biological functions also necessitates developing defensive capabilities, leading Radical Numerics to focus on biosecurity, including detection, source identification, and countermeasures for biological threats.
The concept of 'chain of thought' is being applied to DNA language models, enabling them to demonstrate their reasoning process for designing molecules and potentially leading to more robust AI-driven biological discovery.
From Reading to Writing: The Leap in Genomic AI
The conversation introduces the evolution of AI in understanding biological sequences, moving from simply reading DNA to actively writing it. Eric Nguyen explains that while early genomic language models (GLMs) could interpret DNA, the development of generative models like Evo marked a significant shift, enabling the creation of new DNA sequences. This capability is crucial because much of our genome remains poorly understood; GLMs offer a way to map functions from raw DNA sequences. Early models, like Hyena DNA, utilized efficient convolution instead of attention mechanisms to process lengthy DNA sequences (up to a million characters), demonstrating the ability to predict function and long-range interactions within the genome. The introduction of Evo then pioneered the field of generative genomics, showing that AI could not only read but also generate novel DNA sequences, accelerating biological discovery in ways previously unimaginable.
Omni: A Generalist Model Outperforming Specialists
Unlike previous models like Evo, which sometimes struggled to outperform specialized DNA models, the new Omni model from Radical Numerics demonstrates a significant leap in capability. Omni is designed to be a generalist, capable of handling a wide range of tasks in genomics. The key innovation lies in its comprehensive alignment during and after training, a process that makes pre-trained models more practically useful for scientists. While pre-training focuses on tasks like predicting the next token or filling in gaps, Omni's alignment phase guides the model towards specific, real-world applications. This includes tasks like identifying disease-causing variants from wild-type and mutant sequences, which are not directly apparent from pre-training alone. Omni's performance is notable for its ability to excel across multiple domains, pushing the boundaries in predicting variant impacts and disease-causing mutations, particularly in the complex non-coding regions of the genome.
Designing Novel Biological Systems: CRISPR and Beyond
One of the earliest and most striking demonstrations of generative genomic AI was the design of new CRISPR-Cas systems. Evo was tasked with creating a novel CRISPR-Cas system, and it successfully identified one, a feat that garnered significant attention and was featured on the cover of Science. This capability extends to designing not just single types of sequences but multiple, complex molecular systems. The ability to design DNA, RNA, and proteins from a single framework is a powerful testament to the potential of AI in synthetic biology. Furthermore, AI has achieved the generation of an entire functional genome from scratch, specifically a bacteriophage (a virus), a feat previously impossible for humans. This capability raises profound questions about our ability to control the very fabric of life and its potential implications.
The Biosecurity Imperative: An AI Arms Race
The dual nature of AI's ability to design biological systems—both for therapeutic advancements and potential misuse—necessitates a focus on biosecurity. Radical Numerics views this as a critical area, believing that the teams developing design capabilities are best positioned to build defensive capabilities as well, as they are fundamentally similar models. The defensive side, currently lagging, needs to be elevated to match the offensive potential. This dual mandate involves not only advancing design capabilities but also developing tools to protect against misuse and emerging biological risks. The company is developing a three-pronged approach: detection and surveillance (identifying pathogens in the environment), source identification (determining the origin of a threat), and countermeasures (developing anti-pathogen agents). This focus is crucial as AI-generated biological sequences could be intentionally designed to evade current detection systems, posing a significant challenge.
Understanding Variant Impact: A Key Application
A significant application highlighted is the prediction of variant impacts within DNA. Changes in even a single DNA base can sometimes cause disease, while other times they have no effect. The vastness of the human genome (over 3 billion bases) means much of these variants' effects are unknown. AI models, trained on extensive genomic data, can now better predict whether a specific variant is likely to cause disease. Omni shows particular strength in analyzing non-coding regions of the genome—areas previously difficult to study but which are now understood to play a crucial role in many diseases. This capability surpasses traditional bioinformatics tools that often focused only on the 1.5-2% of the genome that codes for proteins.
Chain of Thought in Genomics: Reasoning and Design
The 'chain of thought' prompting, successful in natural language processing, is being explored for DNA language models. This involves demonstrating the model's step-by-step reasoning process, which can improve its performance and provide insights into its logic. In biology, this translates to showing how a model arrives at a designed molecule or predicts a function. Experiments using RNA aptamers have shown that by presenting the model with progressively better sequences and their associated fitness scores, it can continue this path autonomously and even improve upon them. This suggests a powerful new paradigm for AI-driven biological design, where the model can iteratively refine designs based on learned patterns and desired outcomes.
The Broadening Scope: From DNA to General Biological Intelligence
Radical Numerics aims to move beyond DNA to a more unified understanding of biological systems. The long-term vision is to develop general biological intelligence by integrating various biological modalities—including RNA, proteins, epigenetics, chromatin accessibility, and methylation patterns—alongside DNA. The company believes that DNA is the foundational language from which other biological processes stem. By unifying these different 'languages' and leveraging their shared underlying structure, they aim to create models that can understand and reason across complex biological systems. This ambition includes exploring applications in diverse life forms, from bacteria and viruses to complex organisms, with the ultimate goal of accelerating human health advancements and addressing global challenges.
Navigating the Arms Race: Design vs. Defense
The accelerating pace of AI in biology creates a dynamic akin to an arms race. While offensive capabilities in design are advancing rapidly, defensive capabilities need to keep pace. The challenge lies in building robust biosecurity tools that can not only detect known threats but also identify novel, engineered pathogens that might evade current detection methods. The company emphasizes that while perfect defense may be elusive, continuous improvement and the development of advanced AI-driven tools are essential to mitigate risks. The focus is on enabling the biological defense community with state-of-the-art AI, moving beyond simple sequence matching to understanding functional similarity and potential threats, thereby providing a crucial starting point for building a more secure future in synthetic biology.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●People Referenced
Common Questions
A Genomic Language Model (GLM) is a large language model trained on DNA sequences, similar to natural language chatbots but focused on the basic building blocks of life. It's crucial because AI's ability to read and write DNA can revolutionize scientific discovery and human health understanding.
Topics
Mentioned in this video
Company co-founded by Eric Nguyen, focused on genomic AI models for both biological design and biodefense. It emerged from the work on Evo 2.
A company that developed EVE, a system that optimizes Evo 2, demonstrating the potential for sophisticated models through optimization.
An AI research and deployment company, mentioned in the context of Jason Wei's work on 'chain-of-thought' and Greg Brockman's involvement in supporting Evo 2.
A technology company that supported the Evo 2 project, contributing to its scale and development.
A technology company that conducted work on 'paraphrasing' in the context of protein language models, demonstrating how sequence variations can maintain similar functions.
An e-commerce and cloud computing company, used as an analogy to describe the ease of ordering custom DNA sequences from synthesis companies.
CEO and co-founder of Radical Numerics, who started his PhD in Chris Ray's group and was the lead author and visionary behind the Evo generative model.
Eric Nguyen's PhD advisor at Stanford, whose group was involved in early genomic AI research.
A researcher from OpenAI credited with showcasing the initial examples of 'chain-of-thought' prompting in natural language models.
Eric Nguyen's lab colleague with whom he collaborated on the initial design of Pythia, a convolutional architecture.
Co-founder of OpenAI, who took a four-month break from OpenAI to assist the team with Evo 2, debug code, and scale the project.
One of the first generative genomics platforms, developed by Eric Nguyen and his team, which demonstrated the ability to not only read but also generate DNA. It was showcased on the cover of Science.
The institution where Eric Nguyen completed his PhD, focusing on long-context models that eventually led to his work on genomic AI.
The successor to the Evo model, which led to the founding of Radical Numerics. It was an improved genomic model.
The latest genomic language model from Radical Numerics, which significantly outperforms previous models like Evo 2, especially in tasks related to human genomics and variant effect prediction in non-coding regions.
An early large language model for DNA that used convolutions instead of attention mechanisms, allowing it to process much longer sequences (up to a million base pairs) for functional prediction.
A gene-editing system that can cut DNA. Generative models like Evo have been used to design new CRISPR-Cas systems.
A DNA model referenced in the performance benchmark, which is a supervised model predicting functional genomic tracks like chromatin accessibility or gene expression, distinct from language models.
A public archive of human variants and phenotypes, used as a reference benchmark for assessing the ability of genomic models to predict disease-causing variants.
Another reference benchmark used for assessing the ability of genomic models to predict disease-causing variants.
A large language model from OpenAI, mentioned as an analogy to explain the importance of model alignment and usability for scientists in genomics.
An early convolutional architecture designed by Eric Nguyen and Michael Poli, which was effective in processing long sequences.
A benchmark for proteins where the first checkpoint of the Evo model was tested and found to be competitive with specialized protein models.
A messaging platform used for communication and collaboration, specifically mentioned for debugging code for Evo 2 at late hours.
More from Latent Space
View all 254 summaries
122 min🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow
143 minWhy We Made Jev — Diogo Almeida, TypeSafe Co-founder & CEO
70 minA Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI
87 minThe Watchdogs of AGI — Rune Kvist of AI Underwriting Company
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free