Today, we’re thrilled to share a major advance in our ability to design antibodies fully computationally. We developed a new AI protein design system, JAM, capable of designing high quality lead antibodies de novo given just a target protein sequence and/or structure. We thoroughly experimentally validated this new approach on a range of therapeutic applications and carefully studied the therapeutic properties of these de novo antibody designs.

Where we are today

Recent advances in deep learning have revolutionized our ability to predict the structures of proteins and design proteins de novo. While these tools hold enormous promise for medicine, there has been relatively little progress in using them to develop viable drugs. Antibodies, as the most established drug format, are a great first real-world test for de novo drug design. However, so far no AI systems have shown they are able to produce high quality lead antibodies from scratch in a way that is robust and generalizes across protein targets and antibody formats.

Our advance

We introduce a new AI system called Joint Atomic Modeling (JAM) and use it to design de novo antibodies with good affinity, epitope-specificity, and function. We show that our antibodies have strong developability and other drug-like properties with no optimization, making it possible to replace traditional discovery with de novo design for the first time.

We show generalizability across a diverse range of soluble targets with orders of magnitude higher hit rates compared to traditional discovery, including a protein target with no known structure in the PDB. Next, we apply JAM to a set of multipass membrane proteins, which as a class are considered difficult to drug, and show for the first time we can design antibodies against them from scratch, including the first computational designed binders, of any kind, against Claudin-4 (CLDN4) and the GPCR (CXCR7). When JAM is given time to “think” and introspect on its results, we see significantly improved success rates and affinities - echoing recent trends for large language models (LLMS) and teasing the exciting possibility of “test-time scaling laws” for protein design.

Our technical report is the first demonstration of de novo antibody design relevant to therapeutic discovery and the first evidence of computationally designed antibodies that can target hard-to-drug membrane proteins including GPCRs.

The bigger picture

What implications does this have for the development of new drugs?

Unlocking new target space

• Multipass membrane proteins, like GPCRs, are notoriously challenging. Our approach enables the design of antibodies that can not only drug these targets directly but also serve as precise tools for target validation and the exploration of new biology.
• Intractable targets—such as tight junction proteins, disordered proteins, and transiently stable complexes—cannot be produced reliably in recombinant form, making traditional antibody development difficult. By designing antibodies entirely in silico and with high hit-rates, we bypass the need for recombinant proteins and could even directly test for antibody function from the onset of the discovery process.

Faster antibody development

• Our results show significant efficiency gains in experimental time result from the small library sizes enabled by JAM’s high hit-rates. Importantly, this accelerates the antibody development process whilst also cutting costs.
• Excitingly, epitope-specificity enabled by JAM addresses a key attrition point in traditional discovery campaigns, where antibodies often fail to target the desired epitope. This boost in specificity increases the likelihood of discovering functional leads and cuts down on wasted effort.

Precisely controllable antibody behavior

• Species cross-species reactivity requirements can be used to guide epitope selection, streamlining the preclinical development process.
• Engineered selectivity can similarly be selected for at the epitope level, generating antibodies that either (i) target multiple related proteins with desired binding profiles, or (ii) minimizing binding to closely-related family members and/or dangerous off-targets.

Results

JAM designed antibodies have good affinity, specificity, developability and function within therapeutic range and exhibit strong early stage developability without any experimental optimization.

• Good affinity - We designed de novo antibodies against 7 disclosed (SARS-CoV-2 RDB, HER2, IL-13, PD-1, TrkA, CLDN4, CXCR7) and 1 undisclosed protein target and saw that top designs achieved double-digit to single-digit nanomolar affinities that put them within therapeutic-range for a high quality lead antibody.
• Specificity - Importantly, designs made from one target rarely bound another target, suggesting good specificity. Excitingly, three VHH designs we made to Claudin-4 were over 100x more selective for Claudin-4 than three other closely related off target Claudins (3, 6, & 9). Strong selectivity like the result we saw here is critical to ensuring therapeutic antibodies are safe and can be administered at high doses to achieve the maximum therapeutic impact.
• Early stage developability - We evaluated early stage developability properties of our de novo antibodies and found expression levels, monomericity, and polyreactivity at or better than leading monoclonals (mAbs). These are promising indications that these designs will behave well in manufacturing and formulation, two important hurdles to clear for a commercially viable drug.
• Function - We tested the functional activity of our de novo antibodies using target-specific functional assays. For SARS-CoV-2, we measured neutralization activity using a pseudovirus assay and saw IC50 values in the sub-nanomolar range (0.65 nM). We also use epitope mapping to confirm our designed antibodies bound to the desired functional sites, providing robust evidence of disease-relevant function.

We use JAM to develop the first computationally designed antibodies against multipass membrane proteins, achieving strong affinity, specificity and desired function. This has long been considered a ‘holy grail’ of antibody engineering but has not been previously reported.

• Membrane Protein Targeting - We demonstrate JAM's ability to target challenging multipass membrane proteins by designing antibodies against two therapeutic targets: claudin-4 (CLDN4) and the GPCR CXCR7. To enable efficient screening of large design libraries against these membrane-bound targets, we leveraged JAM's general protein design capabilities to create soluble protein proxies (solMPMPs) that preserve the native extracellular epitopes while enabling solution-phase characterization.
• For both targets, this soluble proxy strategy proved successful. Selected antibodies showed high-affinity binding to their native targets expressed on engineered cell lines, with maintained specificity. Notably, CLDN4-targeting antibodies also demonstrated binding to OVCAR3 cells, an ovarian cancer line expressing endogenous CLDN4, validating their potential therapeutic relevance. These results represent the first demonstration that computational design can generate antibodies against this historically challenging but therapeutically crucial target class.

JAM shows strong signs of generalization, including to novel targets not in the training set, and demonstrates early evidence of test-time scaling laws. We show how using JAM to design experimental libraries further improves performance and compresses design and affinity optimization into a single experimental step.

• Strong Generalization - JAM demonstrated broad generalization across multiple dimensions of antibody design. We validated its capabilities on diverse targets by generating both single-domain antibodies (VHHs) and full-length monoclonal antibodies (mAbs), achieving comparable affinity and function across formats. Notably, JAM successfully designed high-affinity, specific antibodies to a novel target absent from the PDB, indicating genuine generalization beyond training data.
• Scaling test-time compute - A particularly exciting finding emerged when allowed JAM to introspect on its outputs and perform multiple rounds of iterative design at inference time. By increasing these computational iterations, we observed substantial improvements: 3-fold higher success rates, 10-fold better affinities, and unexpectedly, greater structural diversity - all while maintaining developability and humanness. This previously unobserved relationship between test-time compute and protein design success warrants investigation into the best way to allocate compute between training and inference to maximize experimental success.
• One-Shot Affinity Optimization - We further enhanced JAM's capabilities by combining computational design with focused experimental sampling. Using error-prone PCR to explore sequence neighborhoods around JAM-designed binders, we achieved up to 100-fold improvements in affinity, reaching picomolar binding strengths. This hybrid approach effectively compressed de novo design and affinity optimization into a single experimental step, significantly streamlining the discovery process.

Limitations

There are a few things we're working on to improve JAM and to better deliver on the promise of de novo antibody design:

• Humanness: JAM’s antibodies are not fully human; they more closely resemble chimeric or humanized antibodies. While this is suitable for many applications and can be addressed during later stages of development, it does mean additional steps may be required to achieve fully human therapeutics and avoid immunogenicity risks.
• KDs are nanomolar: The affinities we achieve at this stage are representative of lead molecules rather than fully optimized drugs. These molecules serve as a robust starting point for further optimization to reach therapeutic-grade performance, but in many cases won't be there yet right away.
• Further developability assessments: There are a battery of developability characteristics we have not yet investigated for these antibodies but believe they are off to a good start with low polyspecificity, good expression and high monomericity. In future studies, these molecules will be further interrogated for properties including self-association, and stability at high temperatures to determine whether they can meet the rigorous demands of therapeutic development.