Chai Discovery launches CHI-2, an antibody foundation model achieving 15-20% success rates and potentially eliminating high-throughput lab screening

Summary

Chai Discovery launched CHI-2, a new antibody foundation model that achieves hit rates of 15–20% on antibody generation tasks, up from the 0.1% typical of conventional high-throughput screening methods. The improvement is large enough that the company believes it can bypass the standard lab screening step entirely, taking a target, generating candidate protein sequences computationally, and moving directly to focused lab validation.

The scale of progress caught even the company off guard. Chai Discovery's stated goal for all of 2025 was to reach a 1% success rate on this task. By the midpoint of the year, the model is already exceeding 15%, a signal that the underlying technology is compounding faster than internal forecasts anticipated.

The immediate clinical relevance centers on monoclonal antibodies and the emerging class of bispecific antibodies, which bind two targets simultaneously or deliver enhanced potency against a single target. Roughly half of all new drug approvals are now biologics, making antibody design one of the highest-value problems in pharmaceutical R&D. CHI-2 can generate candidate sequences in 24 hours, compared to the months-long timelines traditional discovery requires.

Chai builds its own foundation models in-house and does not use AlphaFold, citing access restrictions on the latest version. Approximately six months after founding, the company open-sourced its protein structure prediction model, which it describes as now in use across most major pharmaceutical companies. Proprietary development has since shifted toward more complex inference pipelines, which Joshua from Chai compares structurally to how OpenAI's o3 chains multiple reasoning steps rather than running single-pass inference.

On the question of business model, Chai is deliberately keeping its options open. The analogy to LLMs is explicit: agility matters more than committing early to either a pure API-provider path or a fully integrated drug developer model. The longer-term technical ambition is what the company calls zero-shot drug candidates, sequences produced directly by the model that carry sufficient drug-like properties to enter clinical development without significant downstream engineering. Hyperpersonalized medicine, where a patient sample informs real-time drug generation, is described as a plausible but multi-leap extension of that vision.