NovaDDE is an AI workbench for protein binder design

From target to ranked binder candidates, NovaDDE brings a conversational agent, an expert-grade design studio, and a purpose-built folding model together in a single workflow.

Drug discovery is, at its core, a search problem. Navigating an astronomically large space of sequences and structures while satisfying constraints from physics, biology, and developability can be tedious, inefficient and challenging. Historically, this search has been fragmented across tools and disciplines, with frequent handoffs between different stages of the process. Take protein binder design as an example: a scientist identifies target and epitopes, a designer proposes candidates; a structural biologist evaluates whether they are likely to adopt the intended structure; a wet-lab scientist sends selected candidates for synthesis and validation. Each step introduces another handoff and another round of waiting, so the feedback loop can stretch from weeks to months, or, for some programs, years.

We built NovaDDE to shorten that loop in an automated fashion. NovaDDE is an AI workbench for protein binder design that puts a conversational agent, an expert-grade design studio, and a purpose-built structure prediction model behind a single interface. Using NovaDDE, a scientist can go from "I want a binder against this target" to a set of binder candidates without leaving the conversation.

NovaDDE has three components

The agent is the primary surface. It's an LLM-powered assistant that understands the vocabulary of protein binder design well enough to turn a plain-language research goal into a concrete design job, deciding which generative model to invoke, what constraints to apply, and how many candidates to sample, and explaining its choices in real time.

The design studio is where that same work gets expert control, for decisions — which pose to keep, which liability to accept — that benefit from direct interaction rather than a sentence.

NovaAtom, the folding model, enhances the design process by predicting whether a candidate sequence is likely to fold as intended and how well that structure fits the design objective.

What sessions look like

A researcher designing a binder against a novel target can perform the whole task inside the agent: a short description of the target, the epitope of interest, and any developability constraints that matter for the program — liabilities to avoid, a framework to stay within should the binder be an antibody, a size range of the binder. The agent translates that into a generative design job, runs it, folds the resulting candidates, and ranks them in one session. The agent reports back a ranked shortlist of designs with the structural predictions attached, with its read on which candidates are worth a closer look and why.

The design studio gives scientists hands-on tools to shape designs directly: selecting hotspots on the structure, defining the peptide design space through parameters such as length and cyclization, and, for antibody designs, specifying the framework and tuning CDR lengths for different interface requirements. Once a design is specified, scientists can submit jobs directly, without relying on an agent to initiate them.

Let's take a look at another example. Suppose we want to design an antibody binder against human CCR6. The agent proposes candidate epitopes from structural evidence, here a CCL20-facing patch on ECL2, and reasons about potential anchor residues.

In parallel, the agent produced an interactive session that lets you inspect the target structure, chains, and proposed hotspots before submitting a design job.

The agent then asks the user to specify the desired antibody format, framework template, and regions to design. Here, the user selects a VHH format with the h-NbBCII10 framework, and chooses to generate CDR1, CDR2, and CDR3 de novo, using the native CDR lengths. Only the framework regions from the template are retained—the template CDR sequences and structures are not used. The agent then prepares a submission package that the user can review and submit through Studio.

Our first BoltzGen campaign with a 200-candidate design budget identified candidates with promising features but the relatively low interface-confidence scores. These in silico scores suggest uncertainty in the predicted binding poses, motivating further computational optimization before advancing candidates to experimental validation.

Additionally, we ran OASis analysis on the shortlisted sequences, providing a repertoire-based measure of sequence humanness to inform further sequence optimization. Taken together, these analyses provide computational criteria for prioritizing candidates and identifying areas for further refinement.

Where this is going

We're actively working on scaling the folding and generative pipelines to better design and evaluate candidate binders. The underlying model, the "Lite" tier of NovaAtom, our internal folding model, is the first step of a larger effort, and we expect the prediction quality and range of modalities it covers to keep expanding. NovaAtom is now also available as Geodesic API Platform for teams who want the model outside the NovaDDE workbench.

As these capabilities mature, we're also exploring iterative refinement as a way to improve design quality: promising candidates from an initial generation can seed additional rounds of sequence generation, folding, evaluation, and filtering against predefined structural criteria. This lets us progressively explore sequence space rather than treating each design run as a single pass.

That same iterative workflow can then extend all the way to the lab, helping scientists move faster from design to experimental validation. Today, a design session ends with a ranked shortlist; the next piece we're building will let scientists order synthesis for selected candidates directly from the same session, with the option to send them to a CRO or directly into NovaLab, our in-house wet lab. This closes the loop between computational design and experimental validation: instead of exporting a sequence list and picking the work back up in a separate ordering system, scientists will be able to design, fold, evaluate, refine, and order protein binder candidates within the NovaDDE system.