The article outlines how artificial intelligence is reshaping therapeutic antibody discovery, spotlighting Chai Discovery’s Chai-2, a generative diffusion model introduced with support from OpenAI. Chai Discovery, which announced a Series A round in August 2025, positions Chai-2 as a breakthrough for de novo antibody design. According to the company, the model can generate protein sequences with stable 3D structures without relying on traditional databases or high-throughput screening, with the aim of accelerating development timelines and improving confidence in moving candidates toward the clinic.
Chai-2 is presented as delivering a 100-fold improvement in success rate over conventional methods and reducing design time by more than 50 percent, from months to under two weeks. The system integrates antigen-specific guidance within a full-atom diffusion framework to tailor antibodies to specific epitopes, supporting end-to-end generation of molecules with functional binding sites. More broadly, the article notes that diffusion models have become a dominant force in generative biology for their ability to transform noise into viable protein structures, enabling high-quality and diverse candidates, scalable design processes, and deeper exploration of sequence space. It cites a growing ecosystem that includes platforms such as AbDiffuser, AntiBARTy and RFdiffusion.
Despite this momentum, the piece flags persistent hurdles in artificial intelligence driven antibody development. These include limited access to complete and unbiased public datasets for training and validation, a lack of negative training data that can skew model performance, and the tendency to overlook key biochemical properties such as solubility and immunogenicity. The result, it cautions, can be molecules that perform in silico but fail to translate in laboratory settings.
To address data gaps, the article highlights Patsnap’s Lao Tzu Antibody-Antigen Dataset, described as a curated resource that combines artificial intelligence based analysis with expert manual annotation. The dataset spans more than 120,000 antibody antigen pairings, 3,300 target antigens, 20,000 affinity measurements and 24,000 IC50 or EC50 datapoints drawn from patent and non-patent sources. It is positioned to help researchers train and validate generative models more accurately and efficiently, supporting the identification and development of therapeutic antibodies.
Looking ahead, the article suggests that as diffusion-based approaches evolve and incorporate novel algorithmic frameworks and large language models, their capabilities will expand further. Patsnap plans to continue growing the Lao Tzu dataset to provide the high-quality data needed for the next wave of breakthroughs in antibody and protein design.
