Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem
Chai Discovery is revolutionizing drug discovery by treating biology as an engineering problem, leveraging AI—particularly diffusion models and the "bitter lesson" of scaling—to design molecules rather than merely discover them. Their approach has boosted antibody design hit rates from 0.1% to 16%, aiming for a "Molecular CAD" suite that collapses discovery timelines from months to days. They partner with pharma, building infrastructure and creating a data flywheel to develop higher-quality, more targeted medicines for previously undruggable diseases.
Chai Discovery aims to transform drug discovery from a trial-and-error process into an engineering discipline, akin to software development. The co-founders, Josh Meier and Matt McPartlon, envision biology as a problem solvable through abstraction layers, enabling rapid iteration similar to modern engineering. They argue that the lab remains crucial for verification, but the paradigm should shift from “drug discovery” (finding a needle in a haystack) to “drug design” (materializing a desired molecule, like designing a key for a lock). Counterintuitively, this shift might lead to more lab testing, as increased productivity boosts the return on investment for experiments.
Evolution of AI in Protein Science
The field has seen rapid advancements, particularly with deep learning. Key milestones include:
- 2018-2020: Significant performance breakthroughs in the biannual CASP protein folding competition, especially with AlphaFold 2 in 2020. Early deep learning models focused on predicting inter-amino acid distances, producing somewhat noisy protein structures.
- 2021-2022: Progression from protein folding to designing sequences that fold into a specific structure, bringing the field closer to drug design.
- Advent of Diffusion Models: This marked a pivotal moment, enabling simultaneous generation of protein structures and sequences. Diffusion models, by iteratively refining slightly broken or noisy inputs, offer the model “more time to think” and learn to build realistic proteins, allowing for more complex “prompts” that incorporate real-world constraints like specific binding sites.
Chai’s Core Philosophy: The Bitter Lesson and Simplicity
Chai’s founding in 2024 was motivated by the realization that antibody folding, once thought intractable due to data scarcity, was becoming solvable with new architectures like diffusion models. Their core principle is the “bitter lesson” from other AI fields: success comes from scaling data, models, and compute. They believe biology, despite its perceived bespoke nature, obeys these universal machine learning principles.
A central tenet for Chai is simplicity in model architecture. Josh and Matt note that complex models with many sub-modules (e.g., Chai 1 with 23 distinct sub-modules) hinder iteration and understanding. Simplifying the architecture helps identify true scaling directions and streamline the research process. Chai builds its models from scratch, rather than fine-tuning existing large language models, emphasizing rigor and a principled approach to tracking progress, especially given the inherent variability and large error bars in wet lab experiments.
Model Performance and Verifiability
Chai has focused on de novo molecular generation rather than merely optimizing existing ones. Their Chai 2 model achieved a significant improvement in antibody design success rates, increasing binding rates from a state-of-the-art of 0.1% (1 in 1,000) to 16% (150 in 1,000). This dramatically higher hit rate provides enough data for meaningful statistical analysis of molecular properties, enabling faster iteration and evaluation.
Interestingly, the founders argue that biology is a more verifiable domain than even code generation, due to objective readouts for properties like binding and manufacturability. While validation might take longer (days to weeks vs. seconds for unit tests), the results are empirically measurable. Chai’s approach aims to unlock “hard targets”—whether they are undruggable GPCRs for scientists, un-foldable proteins for ML researchers, or representation challenges for product designers—by consistently improving their foundational models rather than adding incremental, specialized modules.
The Molecular CAD Vision and Future of Drug Discovery
Chai’s ultimate goal is to develop a computer-aided design (CAD) suite for molecules, moving beyond designing just one or two molecules to fundamentally changing how medicines are discovered. This suite aims to collapse the drug design loop from approximately nine months to nine days, enabling a vastly greater number of hypotheses to be tested.
The long-term vision (2035-2100) for the industry includes:
- Higher Quality Medicines: Drugs that are more intentional, disease-specific, and free of negative interactions.
- Addressing Undruggable Diseases: Tackling conditions like Alzheimer’s, which are currently intractable.
- Economical Solutions: Making personalized medicines and treatments for rare diseases financially viable due to reduced discovery costs and accelerated timelines. This future shifts the industry goal from being “first in class” to being “last in class,” providing the definitive solution for a given disease.
Platform Business Model and Partnering with Pharma
Chai has chosen a platform business model, enabling existing pharmaceutical companies rather than competing with them as a full-stack drug developer. This decision stems from their belief that their AI will be the infrastructure for most future drug discovery. By partnering with major pharma companies (e.g., Eli Lilly, Novartis, Organon, Pfizer), Chai benefits from a scalable business model where model improvements directly lead to revenue and further investment in R&D, rather than diverting resources to clinical trials.
This partnership approach also ensures rigor: pharma partners are highly demanding and validate claims rigorously, forcing Chai’s models to work reliably and generalize broadly across various targets, preventing the company from self-deception that can occur when developing drugs in-house for a limited set of targets. The pharma industry’s competitive and innovation-driven nature drives rapid adoption of effective AI solutions.
Data Strategy and the Flywheel Effect
Chai’s models are trained on critical data sources:
- Protein Data Bank (PDB): The “gold standard” for structural data (crystal structures).
- Massive Sequence Databases: Trillions of protein sequence tokens, where language models like ESM (developed by Josh) can implicitly learn 3D structural information to predict missing amino acids.
The synergy between these approaches creates a data flywheel: as models improve, they can predict more accurate structures from sequences, generating more training data for subsequent model generations. Chai also employs “dogfooding,” where their internal science team uses the models to generate new experimental data, which in turn further improves the models. This mirrors the iterative improvement seen in LLMs, where once a base model is sufficiently capable, further refinements (like reinforcement learning from human feedback) become highly effective.
Staying Ahead and Company Culture
Chai maintains its competitive edge through rigorous evaluation, constantly pushing model capabilities (e.g., integrating manufacturability properties), and quickly productizing breakthroughs. They see their primary competitor as “nature” itself and the long-established, complex wet lab protocols, aiming to surpass the capabilities of even highly experienced scientists like Andy Young, who has decades of experience in antibody engineering.
Chai’s culture is characterized by its mission-driven nature and the deep dedication of its team, driven by seeing tangible results. While acknowledging the challenges of scaling compute infrastructure (managing large GPU clusters, maintaining long training runs, optimizing resources), these are viewed as “pains of growth.” The company prioritizes building robust, production-level code and systems, even if it means moving “a little bit slower in the short term” to ensure long-term, compounding progress. The name “Chai” stands for Chemistry and AI.
Future Outlook
Over the next 6-12 months, Chai is most excited about the real-world deployments and results emerging from their pharma partnerships, anticipating that molecules designed by Chai will begin to reach patients within a few years. Technically, they aim to achieve even greater control over molecular properties, moving towards highly precise design (e.g., targeting exact binding affinities) for specific biological challenges.