Chai-2: Understand the Design Workflow and Its Limits
Chai Discovery’s Chai-2 designs molecular binders. Understand how design differs from structure prediction, what to ask before access, and what evidence your research plan needs.
Updated 5 min read
Chai Discovery’s Chai-2 is a generative molecular design series for creating binders, including antibodies. It addresses a different question from predicting the structure of supplied molecular components: you are asking for new candidates to test. Chai’s Chai-2 introduction (opens in a new tab) describes its approach and vendor-run experiments. Tools like fenbs can keep your own research questions, source links, and experiment notes so you can continue the evaluation in another AI.
What does designing a binder mean?
A binder is a molecule intended to attach to a target. In a design task, the target and desired interaction guide the search for new candidates. The result is a proposal for further research, rather than proof that a candidate will behave as intended in every experimental setting.
Chai’s published Chai-2 evaluation includes experimentally assayed antibody, nanobody, and miniprotein binders. An assay is an experimental test of a property, such as binding under defined conditions. Keep the candidate type and the property tested attached to any result you discuss; “worked” is too broad to describe a scientific finding.
For example, say a research team wants candidates for a permitted target. Its first question is what evidence would justify advancing a candidate to the next check. Writing that question early prevents a design output from quietly becoming a claim about a finished treatment or product.
How is this different from Chai-1?
| Question | Structure prediction | Binder design |
|---|---|---|
| What are you trying to learn? | A proposed structure involving supplied molecular components. | New candidate binders for a defined target and objective. |
| Which Chai series addresses it? | Chai-1. | Chai-2. |
| What follows the output? | Review the prediction and determine appropriate checks. | Evaluate candidates and perform appropriate experimental checks. |
| What should the notes preserve? | Inputs, version, settings, output references, and validation findings. | Design objective, approved inputs, candidate references, and experimental findings. |
Chai-1’s official repository (opens in a new tab) provides public code and a documented local inference path. Those conditions do not establish Chai-2’s access or license. Treat the two series as separate tools when planning who can run the work and how it will be reviewed.
If you are comparing predictions for supplied components, see Chai-1 vs AlphaFold. That comparison should remain separate from a design study. The AI models hub provides a broader starting point for choosing the kind of model your question needs.
Where do you start with access?
Start at Chai’s Platform page (opens in a new tab) and use Get access. The page invites commercial organizations to contact the team and offers an interest-based path for limited noncommercial academic access. It does not provide a verified self-service Chai-2 model-download command.
- Write a short description of the research question and intended use.
- Open the Platform page and choose Get access.
- Use the access form to explain the work and your organization.
- Confirm the available workflow, permitted inputs, access terms, and output handling with the team before planning a study around it.
Do not use Chai-1 installation instructions as a Chai-2 setup procedure. Likewise, an earlier public statement about Chai-1’s web access does not establish current Chai-2 eligibility or quotas. Treat unresolved access questions as planning dependencies, with a person responsible for confirming them.
What should your design request specify?
Prepare a research description before you discuss a run. Name the target, the kind of candidate you want to evaluate, the reason for the study, and the evidence you already have. These are planning questions for your team; they are not a list of verified fields in a Chai-2 interface.
Identify which information can leave your approved research environment. Obtain institutional approval before sending proprietary sequences or confidential research to external services. If the permitted boundary is unclear, keep the initial discussion to public sources and a general description of the research objective.
- Objective: the specific research question and what would count as progress.
- Inputs: approved target references and the versions used.
- Constraints: scientific requirements the team must confirm the workflow can accept.
- Outputs: where candidate files will live and how they will be identified.
- Review: who checks candidates and what experimental step follows.
Give each unknown a next action. “Confirm whether this constraint can be represented” is useful; leaving an empty cell is easy to misread later. Save the permitted summary in fenbs so a research assistant can help prepare the next conversation without guessing what was agreed.
How should you read the reported results?
Read the Chai-2 experimental findings as results from the vendor’s specified model, targets, and assays. They do not guarantee success on a new target, demonstrate patient efficacy, or establish regulatory approval. A designed binder still needs appropriate experimental validation.
Separate the unit being counted from the conclusion. A finding about individual designs and a finding about targets with a successful binder answer different questions. When you summarize a paper, record the experiment and its conditions rather than copying an impressive headline into your own expected-outcome column.
Use the same discipline for your own evaluation. State how candidates were selected for testing, what the assay measured, and what remained untested. The guide to reading research papers with AI explains how to check the source behind a research summary. Our AI model evaluation guide covers choosing criteria before results arrive.
What belongs in a reproducible evaluation plan?
Keep a versioned account of the question, approved inputs, model or service version you can identify, run references, and review criteria. Reproducible here means another researcher can understand what you did and what would be needed to repeat the study. It does not promise identical designs from a service.
Separate planned, attempted, completed, and reviewed work. For example, a candidate can have a saved file while still waiting for experimental testing. Record that status plainly and include the next action. Preserve unsuccessful attempts too, because they can explain why the team changed a requirement or abandoned an approach.
Keep raw experimental data in the scientific systems chosen by your organization. Your shared research notes need enough permitted detail to explain the finding, a reference to the evidence, and the remaining question. They do not need a second copy of every molecular file.
Never lose the work you did in an AI session
After a useful AI session, save it in fenbs: the key observations, findings, research and context. Then carry on from exactly where you left off, on any device and in any other AI. The next AI you connect reads it at the start of the session, so you never have to explain the same problem again.
Create a free accountFree for up to 5 people and 50 open tasks · No credit card · Works with Claude, ChatGPT, Cursor and more
How to keep your research plan with fenbs
Save the work so you can carry on from exactly where you left off, on any device and in another AI. Start with a short, readable record of your own research plan and the public sources you checked.
- Create a free account.
- Write the research question and add the source links you have read.
- Save the approved plan: what you want to test, what you still need to confirm, and who will review it.
- After each check, add what happened and the next thing to do. Keep confidential research in your approved systems.
- Use Integrations to connect a supported research assistant, then ask it to read the saved work before helping you continue.
fenbs does not design binders, run Chai-2, or validate experimental results. Chai’s reviewed pages do not document a native fenbs connection. Use fenbs for your own research notes; choose a scientific tool when the job is molecular analysis or experimental data management.