Wuya Discovery · 无涯探索Try Agent

GoldenThink™ - Antibody Optimization Platform

Don't wait until manufacturing.

Design developability from day one.

All positive and negative data return to L1. The more cycles the loop completes, the more accurate the next round of design usually becomes.

  1. L1 Us · Design and rank

    Starting with your existing antibody, we use your sequence, structure and experimental data to score candidates on four metrics together: affinity, developability, homology, novelty. We have predicted and screened out obvious risks. You receive two deliverables: a candidate sequence panel and a ranking.

  2. L2 You or a partner CRO · Basic wet-lab testing

    The primary criterion is binding: affinity, kon, koff. First, check whether the molecule can be expressed, folds correctly and binds. For those that pass, screen for thermal stability, stability under acidic and alkaline conditions, monomer fraction, aggregation and solubility.

  3. L3 Partners · Function, specificity and mechanism

    Does it affect normal functions that should remain unaffected? Does it cause unintended competition or blocking? Are there obvious off-target effects? Does it trigger inappropriate cellular responses? The key question: is the biological activity as intended?

  4. L4 Partners · Final-format assembly and retesting

    After monoclonal antibodies are assembled into a bispecific or trispecific antibody, can it still be expressed and purified properly? Is it stable? Can both parts work at the same time?

  5. L5 Partners · Final cellular function

    Does the final assembled format actually work at the cellular level?

Both positive and negative data from L2 through L5 return to L1, not just successful results, to guide the next round of design.

The same method applies to monospecific, bispecific and trispecific antibodies.

FAQ

What can you do now, and what do you deliver?

We start with your existing antibody (hit). We assess developability as part of design, without waiting until functional testing is complete and it is time to scale up production. You receive two deliverables:

  1. A candidate sequence panel covering different affinity ranges. We have predicted and screened out obvious risks to developability.
  2. A ranking based on combined scores for affinity, developability, homology, novelty. It identifies which sequences to test first.

Only higher-ranked molecules proceed to wet-lab testing, which takes time and money. Both deliverables fit directly into your existing R&D workflow. You use the ranking to plan experiments.

Can you assess developability before candidate selection? What about stability and aggregation in bispecific and trispecific antibodies?

Before candidate selection, we can already assess which molecules carry lower risk and where changes are needed. But a single prediction score or experiment cannot establish whether a molecule can or cannot become a drug. We score four metrics together: affinity, developability, homology, novelty. The early output is a risk assessment and a set of priorities. If you already have experimental developability data for the molecule, such as thermal stability or aggregation data, we can use them directly to fine-tune the model, which can help improve prediction accuracy and the developability of newly designed molecules. Aggregation risk is related to hydrophobic regions in the sequence, so the sequence alone allows us to screen for and predict this risk before extensive experiments. As molecules become more complex, meeting all four objectives at once becomes harder. Bispecific and trispecific antibodies therefore need to be modeled as complete molecules during design to screen out obvious risks.

Beyond monoclonal antibodies, how do you handle bispecific antibodies?

The same method applies to monoclonal, bispecific and trispecific antibodies. We consider a bispecific antibody as a complete molecule. Most of the data the current model has learned from come from monoclonal antibodies, so work on bispecific and trispecific antibodies needs to be tailored to your molecule. If you already have wet-lab data, design can begin directly. If you do not, we recommend designing and iterating in parallel.

Can your AI help choose how to arrange the two arms of a bispecific antibody (C-terminus or N-terminus)?

There is no universal answer yet. AI can provide informed recommendations. We can predict the structures of different arrangements of the two binding arms, such as attachment at the N-terminus or C-terminus, one by one. We check for steric hindrance and structural conflicts (steric clash) between the arms and confirm that the binding region of each arm can still reach its respective target epitope. Applying this to an actual molecule requires customization based on the target molecule and its wet-lab data, followed by several rounds of learning and iteration. Predictions of the molecular configuration usually become more accurate with support from real data.

How do you manage our data after we share it with you?

We currently offer three options, depending on your needs:

  1. Project isolation: your data are used only for this project and are not reused. The contract specifies how the data and trained models are handled after the project ends.
  2. Your data stay on your side, in your private cloud. We run computations on masked or encrypted data.
  3. Everything runs on your own hardware. Hardware requirements and costs are assessed separately.

During the project, you can retrieve your data and the results we produce at any time. If none of these options fit, we can discuss a custom arrangement based on your specific needs.

How does your AI handle different affinity requirements across projects?

The aim is to match your target affinity, not simply to maximize it.

  1. Specify the affinity range you want for the project as an input parameter for AI.
  2. AI uses this target to generate a set of candidate variants with different affinities.
  3. You then have two options: test affinity yourself and return the data to us for further optimization, or have our partner CRO handle all of it. The two options differ in time and price.
Sharing our molecule with you reveals the target. How do you ensure confidentiality?

Molecules and experimental data can indeed reveal the target. We recommend signing a confidentiality agreement before exchanging information and specifying three things in the contract:

  1. Your data are used only for your project.
  2. They are not used for other clients' projects without your written consent.
  3. How the data, models and backups are handled after the project ends.

If you have particularly strict confidentiality requirements, local deployment can provide physical isolation.