 Juan Carlos Lopez |
 Andy Marshall |
Although therapeutic antibodies represent a $160 billion-dollar annual market and comprise a third of all approved drugs, discovering new antibody molecules remains a labor-intensive process, requiring slow experimental approaches with low hit rates, such as animal immunizations and or the panning of phage- or yeast-displayed antibody libraries. The drug hunter’s dream would be to design an antibody to any target by simply entering information about that epitope into a computer. Now that dream is one step closer with a recent proof of principle peer-reviewed paper published in Nature on work disclosed last year from the team of 2024 Nobel Laureate David Baker. Baker and his colleagues at the University of Washington introduce the first generalizable machine-learning method for designing epitope-specific antibodies from scratch without relying on immunization, natural antibody repertoires, or knowledge of pre-existing binders.
|
|
Unlike small-molecule drug development, which has benefitted from an explosion of interest in the use of machine-learning models, in-silico design of antibody binders has lagged far behind. One reason for this is the paucity of high-resolution structures of human antibody–antigen pairs—currently only ~10,000 structures for 2,500 antibody-antigen pairs have been lodged in SAbDab (a subset of the RCSB Protein Data Bank). Most of these structures are soluble protein antigens, but there’s little data to model antibody binders to GPCRs, ion channels, multipass membrane proteins and glycan-rich targets, which are of most commercial interest. Overall, the antibody–antigen structural corpus is orders of magnitude smaller, noisier and narrower than that available for small molecules, lacking information on binding affinities and epitope competition maps via PDBBind/BindingDB/ChEMBL.
For these reasons, most companies have focused on machine learning prediction of developability properties—low aggregation, high thermostability, low non-specific binding, high solubility, low chemical liability/deamidation and low viscosity—for an antibody’s scaffold, rather than in-silico design of the six complementarity determining-regions (CDRs) on the end of an antibody’s two binding arms.
Even so, several recently founded startups have claimed to be using machine-learning models to predict/design antibody binders from scratch. These include Xaira Therapeutics, Nabla Bio, Chai Discovery and Aulos Bioscience.
Xaira debuted last year with >$1 billion in funding to advance models originating from the Baker lab. Nabla Bio also raised a $26 million series A in 2024, publishing preprints in 2024 and 2025 that describe its generative model (‘JAM’) for designing VHH antibodies with sub-nanomolar affinities against the G-protein coupled receptor (GPCR) chemokine CXC-motif receptor 7 (CXCR7), including several agonists. In August, Chai announced a $70 million series A financing based on its ‘Chai-2’ generative model disclosed in a preprint that details de novo antibodies/nanobodies against 52 protein targets, including platelet derived growth factor receptor (PDGFR), IL-7Rα, PD-L1, insulin receptor and tumor necrosis factor alpha, with “a 16% binding rate” and “at least one successful binder for 50% of targets”.
Finally, Aulos emerged with a $40 million series A in 2021 as a spinout from Biolojic Design. This program has generated computationally designed de novo CDR binders with picomolar affinities for epitopes on HER2, VEGF-A, and IL-2. The IL-2 antibody (imneskibart; AU-007)—designed to selectively bind the CD25-binding portion of IL-2, while still allowing IL-2 to bind the dimeric receptor on effector T cells and natural killer cells—reported positive phase 2 results in two types of cancer just last week. Absci, another more established company, has also been developing de novo antibodies, publishing a generative model for de novo antibody design of CDR3 loops against HER2, VEGF-A and SARS-CoV-2 S protein receptor binding domain.
|
|
Overall, though, computational efforts have largely optimized existing antibodies or proposed variants once a binder already exists. Recent generative approaches have often needed a starting binder, leaving de novo, epitope-specific antibody creation as an unmet goal. The Baker paper now provides a generalizable, open-source machine-learning approach that can find low nanomolar antibody binders to a wide range of targets.
To accomplish this task, the authors use RFdiffusion, a generative deep-learning framework for protein design, extending its capabilities by fine-tuning it specifically on antibody–antigen structures. Their goal was to enable the in-silico creation of heavy-chain variable domains (VHHs), single-chain variable fragments (scFvs), and full antibodies that target user-defined epitopes with atomic-level structural accuracy.
Their approach integrates three major components: backbone generation with a modified RFdiffusion model, CDR sequence design via the algorithm ProteinMPNN, and structural filtering using a fine-tuned RoseTTAFold2 predictor (the authors note that improved predictions can now be obtained by swapping out RoseTTAFold2 for AlphaFold3 developed last year by Google Deepmind and Isomorphic Labs). The refined RFdiffusion model can design new CDRs while preserving a fixed antibody framework and sampling diverse docking orientations around a target epitope. The resulting models generalize beyond training data, producing CDRs unlike any found in natural antibodies.
Baker and his colleagues created VHHs against several therapeutically relevant targets, including influenza H1 haemagglutinin, Clostridium difficile toxin B (TcdB), SARS-CoV-2 receptor-binding domain, and other viral or immune epitopes. High-throughput screening via yeast display or purified expression led to the identification of multiple binders, typically with initial low affinities in the tens to hundreds of nanomolar range. Cryo-EM confirmed near-perfect structural agreement between design models and experimental complexes, particularly for influenza haemagglutinin and TcdB, demonstrating atomic-level accuracy across the binding region and the designed CDR loops. To enhance affinity, the authors used OrthoRep, an in-vivo continuous evolution system, for the affinity maturation of selected VHHs. The affinity of the resulting VHHs improved by roughly two orders of magnitude while retaining the original binding orientation.
Baker and his team further challenged their method with the more difficult problem of de-novo scFv design, which requires simultaneous construction of six CDR loops across two amino acid chains. The team introduced a combinatorial assembly strategy in which heavy and light chains from structurally similar designs were mixed to overcome cases where a single imperfect CDR would compromise binding. This enabled the discovery of scFvs targeting the Frizzled epitope of TcdB and a PHOX2B peptide–MHC complex. Cryo-EM validation of two scFvs showed that all six CDR loops matched the design model with near-atomic precision.
Future work is needed to extend de novo antibody prediction via this method to tougher target classes, such as membrane proteins. Clearly, modeling across all six CDR loops and the heavy and light chains remains a hard problem; indeed, the paper’s marquee result was designing a single scFv where all six CDRs matched the designed pose at high resolution; more generally, scaling reliable heavy- and light-chain co-design beyond a few cases remains an open engineering challenge that future methods will need to solve. For the field to gather momentum, benchmarking efforts like the AIntibody challenge will be needed, together with public efforts to create datasets of negative binding data, akin to those described in a paper published earlier this year.
Overall, the Baker paper is seminal work that establishes a practical and accurate approach to designing epitope-specific antibodies from scratch. It represents a major advance in the development of therapeutic antibody discovery.
Next Phase Newsletter 