Chai Discovery separates molecular prediction from AI design

Chai Discovery develops AI models for molecular structure prediction and molecular design. Those are related but different tasks: predicting how an existing set of molecules may interact is not the same as generating a new molecule with desired properties. The company offers an open Chai-1 model and more controlled access to newer design capabilities. A useful evaluation starts by choosing the scientific question and the accessible model, then separating computational output from experimental evidence.

01 / ProductOne company with different model generations and access routes

Chai’s Chai-1 introduction describes a multimodal model for molecular structure prediction, covering proteins, small molecules, DNA, RNA and modifications. The current repository provides the inference software and describes an Apache 2.0 license for both code and weights. This gives researchers a concrete, inspectable route into the company’s technology rather than relying entirely on a sales demonstration.

Chai-2 shifts the emphasis toward generating binders, including antibody and miniprotein designs. Chai reports experimental results for selected target sets, but those are company research findings under particular conditions. They should not be read as a universal probability that any requested design will work, much less as evidence of a medicine’s safety or clinical value.

By 2026, the company’s Series C update identifies Chai-3 as its latest model. A Pfizer license announcement describes early Chai-3 access and a custom model using Pfizer data. Yet the public product form still names Chai-2. The correct reading is a company with several access arrangements, not a freely downloadable latest-generation design model.

02 / AudienceFor research teams with a defined experimental question

A computational biology group evaluating structure predictions has a different need from a discovery team commissioning new binders. The first can begin by examining Chai-1’s supported inputs, outputs and reproducibility on known examples. The second needs to define desired properties, experimental validation and the commercial rights surrounding generated candidates. Selecting a model before defining that distinction makes the evaluation harder to interpret.

Chai’s platform page describes target structures or sequences, antibody formats and design criteria as inputs to its design offer. These are capabilities to discuss with the vendor, not a promise that every target or format is accepted for immediate access. The page explicitly distinguishes commercial access from limited non-commercial academic access and routes both through an expression of interest.

Insilico Medicine is a useful comparison for a broader AI drug-discovery platform and development strategy. Recursion provides another perspective on combining large-scale biological data and experimental work. Chai’s specific appeal here is molecular prediction and design models that can be used within a research organization’s own discovery process. A team without scientific review and an experimental path cannot assess the value of a plausible-looking structure alone.

03 / WorkflowA proposed evaluation that keeps prediction and design separate

A proposed evaluation should begin with a bounded, non-sensitive research question and examples whose relevant properties are already understood. Decide whether the objective is structure prediction, comparison of computational hypotheses or access to a newer design capability. This article does not supply or test a molecular design; it describes how a research team could organize an evaluation responsibly.

For Chai-1, review the published software requirements and choose a reproducible environment. The repository specifies Linux, a suitable CUDA GPU and supported numerical capabilities, and recommends pinning the package version. Record the model and software version, input provenance and any optional information supplied to the run. The practical purpose is to ensure that another researcher can understand why two outputs differ, rather than treating every change as a model improvement.

Separate the predicted structure from the conclusion drawn from it. An appealing configuration may suggest a hypothesis, but it does not establish binding, function or suitability as a therapeutic candidate. Compare output with appropriate known examples and retain uncertainty. If the project moves to a design model, obtain the actual available model and access conditions instead of assuming that the public Chai-1 code reproduces Chai-2 or Chai-3.

For a commercial design evaluation, agree in advance how computational suggestions will be assessed by the team’s established experimental process. Define the acceptance criteria, the negative results that will be retained and the boundary between vendor-reported performance and results generated for this project. Keep computational, laboratory and downstream development milestones separate. The result of the pilot should be a decision about whether the model improves a specific research step, not a claim that it compresses the entire path to an approved medicine.

04 / PricingOpen licensing and negotiated access are different offers

There is no single public price that describes Chai Discovery’s portfolio. Chai-1 has a published software-and-weights license; the design-model route is an access request; major pharmaceutical arrangements are negotiated collaborations or licenses. The public access page does not disclose a numerical tariff, billing unit or standard commitment. Keep compute expenditure, software rights and project terms distinct.

An open license for one model does not grant access to a different internal model or a partner’s custom software. Likewise, a partnership announcement establishes an arrangement between named organizations, not a standard public plan. A serious proposal should identify the model version, what the customer receives, which data is used, how results may be used and who owns the relevant outputs and improvements. These are unresolved commercial questions, not assumptions this blueprint makes on the customer’s behalf.

05 / DistinctionsDesign quality extends beyond a successful binding result

Chai’s antibody research update explicitly distinguishes binding from broader developability. It discusses full-length antibody designs, biochemical properties and structural validation. That is a more useful frame than treating a single binding-rate figure as the complete measure of a design system. A candidate can succeed on one experimental dimension while remaining unsuitable on another.

The original Chai-2 announcement also describes a defined target panel and experimental characterization. The evaluation lesson is to inspect what was measured, on which targets, with what selection process and what definition of success. A percentage without that context cannot be transported into a new discovery program. This review has not independently reproduced the company’s results or compared models under a shared protocol.

Commercial adoption is a separate kind of evidence. The Bristol Myers Squibb collaboration describes using Chai’s models and platform for antibody discovery. Together with the Pfizer arrangement, it supports the company’s prominence and real enterprise interest. It does not establish that a particular generated molecule has reached a clinical milestone or that a buyer outside those partnerships receives the same capabilities.

06 / QuestionsResolve the model version and the experimental evidence

The public access materials are not completely synchronized in their model naming. The product page foregrounds Chai-2, while current company and partner announcements discuss Chai-3. A prospective user should ask which version is actually being offered, what supporting documentation is available and whether results from earlier generations are relevant to it. Avoid using the newest name in a purchasing decision without a defined delivery scope.

For the open route, local execution creates its own obligations. The team must manage compute, inputs and the reproducibility of its analysis. A downloadable model is useful for control and inspection, but it does not provide a complete research service. Establish who maintains the environment and how changes to software or optional inputs are reviewed before they affect comparisons.

For all routes, distinguish a computational hypothesis, an experimentally observed property and a development conclusion. The company’s research may justify further investigation, but it does not remove the need to evaluate the target, assay context and downstream requirements of a particular program. Retain unsuccessful cases in the pilot record. A selective gallery of successful outputs would make it difficult to judge whether the tool improves the research decision that motivated the purchase.

07 / DecisionChoose the model route before designing the evaluation