Building Scalable AI Pharma: Lessons from Chai’s Lean, High-Impact Approach

On August 7, 2025, San Francisco–based startup Chai Discovery announced a $70 million Series A funding round, led by Menlo Ventures with participation from new and existing investors including Anthology Fund, Yosemite, DST Global Partners, and OpenAI. Within just one year of its founding, Chai has completed two rounds of financing totaling $100 million.

In June 2025, Chai released its latest foundation model, Chai-2, capable of zero-shot de novo antibody design. In simple terms, the model can “imagine” new antibodies from scratch and hit the target antigen in a single round of lab experiments, achieving a 15%–20% average hit rate which is an improvement of several orders of magnitude compared to traditional methods.

In the field of AI-driven protein modeling, DeepMind’s AlphaFold is regarded as a technological milestone. Its groundbreaking ability to predict protein structures stunned the scientific community and led many to believe that AI had already “hit the ceiling” in protein research and that structural prediction seemed almost solved. However, AlphaFold’s strength lies in predicting the static 3D folds of proteins, not in designing functional, high-affinity molecules that can be directly applied to drug development.

While AlphaFold can “decode structures”, it does not efficiently solve the key challenge of designing new proteins with clear functions and stable structures. This gap in drug discovery has now become a central focus for many AI biotech startups, who aim to go beyond mere structural prediction and move toward designing novel protein molecules that are both synthesizable and experimentally verifiable.

At a time when the AI drug discovery field remains stuck in the bottleneck of sequence generation without experimental validation, Chai has broken the cycle of “model inaction” with an integrated "generation + validation" approach. Focused specifically on antibody generation, this AI biotech team has launched two generations of models within just two years.

A small but exceptional team aiming to do big and profound work

Founded in 2024 and headquartered in San Francisco, Chai Discovery is a rapidly rising AI-driven antibody discovery company. Despite having fewer than ten members, this small but highly focused team has quickly entered the ranks of leading global bio-computing innovators thanks to its streamlined technical approach and cutting-edge expertise.

The two co-founders, Joshua Meier and Jack Dent, have been close friends since university and long-term collaborators at the intersection of AI and biomedicine. Meier previously worked at OpenAI and Meta FAIR (Facebook AI Research), specializing in protein structure modeling and generative model development. Dent, meanwhile, brings engineering experience from Google and multiple biotech startups, with expertise in translating foundation models into practical, deployable systems.

The founders’ original intention in creating Chai was not merely to build a large model capable of “writing sequences”, but to establish a complete end-to-end loop from AI generation to experimental validation in driving a paradigm shift in drug discovery from experiment-driven to design-driven. In practice, this means that AI is not only generating hypotheses, but also integrated into synthesis, expression, and validation processes, thereby enabling a truly systematic antibody discovery platform with a rapid feedback mechanism.

Chai-1 Generation: Closed-loop multi-format antibody discovery within two weeks, accelerating efficient bispecific and ADC design

The discovery of antibody drugs has traditionally been a painstaking process. Conventional approaches generally fall into two categories: the first involves immunizing animals (such as mice) to produce antibodies, followed by humanization, screening, and optimization; the second relies on phage display libraries, screening through tens of millions—or even billions—of antibody sequences to identify candidate molecules. Both methods typically require months of affinity optimization, stability engineering, and expression tuning. And within these vast libraries, the probability of hitting a truly effective target is usually less than 0.1% [1].

It is against this backdrop that Chai Discovery’s model framework introduces the possibility of fundamentally re-engineering efficiency. Its core breakthroughs focus on three key areas:

First, in controllability of generative models, Chai-2 adopts a multimodal generative architecture that integrates all-atom structural prediction with sequence generation. Each candidate antibody sequence is generated alongside an evaluation of its 3D conformation and binding pocket accessibility, significantly improving the plausibility of outputs and the success rate of downstream validation.

Second, Chai has established an end-to-end “generation–experiment” closed-loop workflow. From a single design run, the model generates ≤20 candidates, which are then expressed and tested for binding in a standard 24-well plate format. The results are fed back into the model for the next iteration. This entire cycle can be completed within two weeks—dramatically faster than conventional screening processes.

Finally, in terms of antibody format and functional diversity, Chai-2 supports not only the traditional IgG format but also scFv, VHH, and miniproteins, enabling the platform to design complex modalities such as bispecific antibodies, antibody-drug conjugates (ADCs), and fusion proteins.

When its first foundation model Chai-1 was released in September 2024 [2], the company integrated multimodal biological structure data into its AI framework. Chai-1 can take as input protein sequences, small molecules, or DNA/RNA fragments, and output the 3D structure of their complexes.

Chai-1 performed strongly across multiple international benchmarks—including the PoseBusters structure prediction benchmark and the CASP15 protein folding competition—where it demonstrated more precise conformational reconstruction and stronger geometric consistency compared with AlphaFold.

Chai-2 Iteration: Moving beyond structure prediction to boost discovery hit rates

In early 2025, Chai released Chai-1r, its first version integrating structure-constrained generation. Users can provide experimentally derived binding-site information (such as contact residues or binding pocket geometry), and the model incorporates this guidance during sequence generation to improve the physical plausibility and target-binding potential of the final designs.

Co-founder Joshua Meier noted that by incorporating antigen epitope information as structural guidance, the Chai-1r model significantly outperformed both its predecessor and AlphaFold2 in standard antibody–antigen docking evaluations. Specifically, Chai-1r achieved a DockQ score of 43.7, compared with 35.6 for the previous model and 20.6 for AlphaFold2, demonstrating substantially higher predictive accuracy.

In June 2025, Chai reached a major milestone with the official release of its next-generation model, Chai-2 [1], showcasing for the first time its zero-shot de novo antibody design capability in experimental settings. “Zero-shot” refers to the model’s ability to generate candidate antibody sequences for entirely new antigens without requiring pretraining samples or curated libraries.

According to data published on Chai’s website, the team tested 52 previously unseen antigen targets, achieving a 15–20% hit rate in single-round designs, with some targets reaching up to 50% binding success. This performance is orders of magnitude higher than the 0.1–0.5% average hit rate typical of traditional antibody library screening methods.

Chai-2 is not limited to antibody design; it also supports multiple formats such as single-chain antibodies (scFv), nanobodies (VHH), and miniproteins. In miniprotein binder design, Chai-2 achieved a laboratory-validated hit rate of up to 68% (Fig. 4), with binding affinities reaching the picomolar (pM) range.

In addition, antibodies generated by Chai-2 exhibit outstanding drug-like properties, including high specificity, nanomolar-level affinity, and strong developability, laying the foundation for rapid translation into therapeutic applications.

Currently, Chai-1 is available on GitHub as an open-source package (chai_lab), allowing users to access the model for structure prediction or molecular design via code. The API availability for Chai-1r and Chai-2 has not yet been announced.

Oracle, NVIDIA, and Sino Biological are its strong backers

Chai Discovery is able to generate and validate candidate molecules within an extremely short timeframe, thanks to a synergistic ecosystem built on cloud computing, experimental facilities, and open scientific collaboration.

Specifically, Chai collaborates closely with Oracle Cloud Infrastructure (OCI), leveraging Oracle’s cloud platform to meet the advanced computational demands of its breakthrough AI model, Chai-2, including high-performance GPU clusters for model training and inference.

At the same time, Chai’s experimental validation is handled by Beijing Sino Biological and the Canadian synthetic biology platform Adaptyv, covering critical steps such as protein expression, binding screening, and functional verification.

Additionally, Chai participates in the UK government–led OpenBind initiative, an open science program aimed at generating over 500,000 protein–ligand complex structures and affinity data to support the training and evaluation of AI-driven drug discovery models.

This combination of cloud computing power + third-party experimental platforms + open scientific data enables Chai to build a highly efficient, self-driven, and scalable AI antibody design platform, offering a reference model for the infrastructure of future AI biotech enterprises.

The “generation + validation” cycle: potentially the key to revolutionising AI drug discovery

The case of Chai Discovery further suggests that strong generative capabilities alone do not constitute an industrial-grade advantage. The true breakthrough lies in deeply integrating AI models into experimental workflows, creating a rapid “design–validate” iteration cycle that significantly improves the real-world conversion efficiency of candidate molecules.

Moreover, Chai’s lean team of fewer than ten people developed three generations of models and an experimental closed-loop system in under two years, prompting other companies to rethink the efficiency of traditional “large team + long chain” R&D models. Future efforts may require more flexible, lightweight, and modular organizational structures.

Looking ahead, whether in antibody discovery, protein design, or small-molecule screening, a combination of "generation + validation + platform" ecosystem is likely to become the core capability set for next-generation AI drug discovery companies. In a context of falling computational costs and increasing data sharing, those who can first build an efficient, scalable closed-loop system are poised to stand out in the next phase of global biopharmaceutical innovation.

References:

[1] Zero-shot antibody design in a 24-well plate Chai, Discovery Team, Jacques Boitreaud, Jack Dent, Danny Geisz, Matthew McPartlon, Joshua Meier, Zhuoran Qiao, Alex Rogozhnikov, Nathan Rollins, Paul Wollenhaupt, Kevin Wu bioRxiv 2025.07.05.663018; doi: https://doi.org/10.1101/2025.07.05.663018。

[2] Chai-1: Decoding the molecular interactions of life, Chai Discovery, Jacques Boitreaud, Jack Dent, Matthew McPartlon, Joshua Meier, Vinicius Reis, Alex Rogozhnikov, Kevin Wu bioRxiv 2024.10.10.615955; doi: https://doi.org/10.1101/2024.10.10.615955

Original article by Huang Yurou of VCBeat: https://www.vbdata.cn/1519037666