Key Moments
🔬Biology Is Turning Into Software — Matt McPartlon and Neil Patil, Chai Discovery
Latent Space PodcastWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AI is transforming biology into an engineering discipline, enabling rapid drug design. However, challenges remain in validation speed and talent acquisition, hindering faster progress.
Key Insights
Chai Discovery secured partnerships with major pharmaceutical companies like Eli Lilly, Pfizer, Novartis, and argenx, indicating a significant shift in how pharma adopts AI design tools.
Chai's protein design models, like Chai-2, have demonstrated success in designing antibodies against 50 targets, achieving binders for approximately half with a 20% hit rate on average.
The development of advanced structure prediction models, such as AlphaFold 2 and Chai's own Chai-1, was crucial for unlocking the potential of protein design.
Chai's product suite is moving beyond a chatbot interface towards a visual design suite, akin to tools like Autodesk or Figma, to facilitate molecule design.
The cost of developing a successful drug can be up to $2.6 billion, making AI-driven acceleration in early-stage discovery, saving millions, highly attractive.
A significant bottleneck in AI for biology is the slow validation loop, where experimental results can take months, hindering rapid iteration compared to software development.
Biology is becoming an engineering discipline with AI
The conversation highlights a fundamental shift in biology, moving from a purely scientific experiment to an engineering discipline, driven by advancements in AI. Chai Discovery's co-founders, Matt McPartlon and Neil Patil, explain that AI models are enabling the precise engineering of molecules, much like software or mechanical engineering. This allows for declarative definition of desired outcomes, with AI filling in the gaps to create drug candidates. This paradigm shift is supported by powerful models like Chai-2, which have shown significant success in designing antibodies, achieving a substantial hit rate in binder discovery and even enabling complex functionalities like GPCR agonist activity.
Pharma's rapid adoption of AI design tools
A key indicator of this shift is the sudden willingness of major pharmaceutical companies to purchase AI design tools rather than building their own internal pipelines. Chai Discovery has secured significant partnerships with industry giants like Eli Lilly, Pfizer, Novartis, and argenx. This adoption is driven by the compelling value proposition of AI models that can accelerate the lengthy and expensive drug discovery process. Companies are now relying on neutral software factories like Chai to design medicines, marking a departure from traditional, more experimental approaches.
The evolution of protein design models
The progress in protein design is strongly linked to advancements in structure prediction models. The advent of models like AlphaFold 1 and 2 was a major breakthrough, enabling accurate prediction of protein structures, which is foundational for design. Chai's own models, starting with Chai-1 (an open-sourced structure prediction model), laid the groundwork. Chai-2 then represented a significant leap into design capabilities, moving from predicting existing protein structures to generating novel molecules that can bind to specific targets. This evolution from structure prediction to generative design is what has convinced pharma of AI's utility.
Chai's product vision: from chatbot to design suite
Chai's product strategy is evolving from a simple chatbot interface to a sophisticated visual design suite. Inspired by tools like Autodesk, SolidWorks, and Figma, the platform aims to provide users with a visual environment to design molecules. Features include 'paint tools' for epitope design and 'content-aware fill' for generating binders. This visual approach is crucial for making complex AI tools accessible and usable for medicinal chemists, who are historically skeptical of AI and require intuitive, understandable interfaces to trust and adopt new technologies.
Addressing selectivity and cross-reactivity
A critical challenge in drug design is ensuring molecules are selective for their intended target and avoid binding to other unintended targets (cross-reactivity). Chai's platform incorporates both modeling and product-side strategies to address this. Users can intentionally design for selectivity by identifying conserved regions or specific binding sites. The product suite allows for the explicit definition of what to bind and what to avoid, helping to mitigate the risks of toxicity and off-target effects that often lead to drug failures. This precision engineering is key to developing safer and more effective therapeutics.
Overcoming validation and talent bottlenecks
Despite significant AI advancements, a major bottleneck remains the slow validation loop in biological research. Unlike software development where feedback is rapid, experimental validation of drug candidates can take months. This significantly slows down iteration and improvement. Another challenge is talent acquisition; there's a perceived scarcity of smart individuals focusing on biology compared to software engineering or LLMs. Chai emphasizes that a strong biology background isn't always necessary, aiming to attract computational talent by making the field more accessible and visual. The company also focuses on engineering primitives and simplicity to manage complexity, recognizing that well-engineered infrastructure is crucial for scaling AI in science.
The economic driver: saving millions in a multi-billion dollar process
The high cost of drug development, potentially reaching $2.6 billion per successful drug, makes AI-driven acceleration in early discovery highly attractive. While saving a few million dollars might seem small in comparison, it represents a significant saving in the overall campaign. Furthermore, AI enables the pursuit of more ambitious and complex drug modalities, such as bispecific antibodies or ADCs (Antibody-Drug Conjugates), which are difficult or impossible to discover with traditional methods. By improving the probability of finding binders and optimizing them, AI can unlock new therapeutic avenues and portfolios of drugs, rather than just making existing processes faster.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Drugs & Medications
●People Referenced
Common Questions
Chai Discovery is a protein design startup that leverages AI models to accelerate drug discovery. It aims to transform the lengthy, expensive, and often 'brute force' process of finding therapeutic candidates into a more precise and efficient engineering discipline, focusing on designing novel proteins and antibodies.
Topics
Mentioned in this video
One of Chai Discovery's major pharma partners.
One of Chai Discovery's major pharma partners.
Mentioned as an example of visual design software, similar to how Chai's product aims to function for molecule design.
Used as an analogy for Chai's visual design suite for molecules.
The protein design startup being discussed, which is about two and a half years old and has made significant advancements in the field.
A SAS company where Neil Patil was one of the first employees.
One of Chai Discovery's major pharma partners, working closely on developing the design suite.
One of Chai Discovery's major pharma partners.
Where Chai's CEO, Josh, worked on the original ESM papers.
Co-led Chai Discovery's seed round and allowed Chai to use their vacant office space in early days.
A company providing a framework for durable execution in software engineering, heavily utilized by Chai for orchestrating long-running jobs and handling infrastructure flakiness.
Mentioned in the context of the intense competition for compute resources, with hyperscalers and big AI labs dominating the market.
Implied through discussion of GPU units like B300, highlighting their role in the compute market for AI models.
Used as an analogy for engineering simplicity, referencing the evolution of their Raptor engines from complex to streamlined designs.
Referenced in the context of a research paper where they distilled a very large dataset, demonstrating efficiency in data usage.
A company started by Neil Patil's brother, focused on design for LLMs and helping enterprises with their language data for specialized tasks.
Cited as one of the biggest and earliest venture outcomes in Silicon Valley, illustrating the venture capital model in bio-pharma.
Referenced through its CEO, Satya Nadella, and his vision for the future of work with AI.
Used as an analogy to highlight how software enables precision engineering in electrical engineering, similar to what Chai aims for in biology.
Referenced as visual design software, drawing a parallel to Chai's approach to protein design.
Contrasted with Chai's visual design suite to highlight the difference in interface approach for complex biological design tasks.
Models developed at Meta, whose early signs of life suggested scaling laws in protein design.
Protein structure prediction software that Matt McPartland saw early signs of life in during his PhD, particularly AlphaFold 1 and 2.
A version of AlphaFold capable of predicting the shape of two proteins at once, crucial for unlocking design.
Chai's structure prediction model, open-sourced to contribute to the community and build necessary infrastructure for future models.
A version of Chai's model released after Chai 2, which included a study on the developability of molecules.
The latest series of Chai models, which focuses on producing molecules with high binding affinity and developability to reach therapeutic grade.
Mentioned as an example of a complex system with many submodules, making it difficult to optimize and study.
Co-founder of Chai Discovery, with a background in AI biology and theoretical computer science, specializing in protein structure prediction.
Leads platform and product at Chai Discovery, with a background in app development, robotics, self-driving cars, and SaaS, focusing on infrastructure and productization.
One of Chai's first hires as a hardcore lab scientist, who initially skeptical, helped battle-test Chai's models.
CEO of Microsoft, whose vision of 'making everyone a manager of infinite minds' is discussed in the context of AI's impact on engineering roles.
More from Latent Space
View all 242 summaries
103 minThe Inference Frontier: 10x Faster Models to Self-Optimizing AI — Philip Kiely & Ali Taha, Baseten
71 minOpenAI’s Plan to Make ChatGPT the Everything App — Akshay Nathan, OpenAI
117 minPoolside’s Model Factory, Laguna S, Open Models, and the Race to AGI — Eiso Kant, Poolside AI
90 minCausal Models Need Causal Data - Xaira’s X-Cell model (Bo Wang & Ci Chu)
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