Therapeutic antibody development must reconcile the immune system’s natural ability to produce potent binders with additional requirements for biochemical and physiological suitability for medicine. The authors present a machine learning-based strategy to generate humanized anti-PD-1 antibodies that aims to preserve favorable therapeutic features while exploring sequence and structural variants that could change potency. The approach is framed as an interpolation within a chemically informed latent space, enabling systematic generation of novel designs informed by existing clinical antibodies.
This work focuses on antibodies that target Programmed Death‑1 (PD‑1), a clinically validated immune checkpoint target. The authors emphasize that redesigning antibodies by traditional methods can be slow and limited in scope; their objective is to use AI-guided generative modeling to expand the design space and accelerate identification of viable therapeutic candidates.
The generative model used is a conditional kernel-elastic autoencoder (CKEA). The CKEA was employed to interpolate between two clinically relevant anti-PD-1 antibodies, nivolumab and pembrolizumab, which both engage the FG-loop hotspot on PD-1 but do so in substantially different orientations. The authors note that the binding orientations of these two antibodies differ by 174 degrees, providing diverse structural templates for the latent-space interpolation.
By conditioning the latent space on biochemical information and kernel-elastic constraints, the framework is intended to produce designs that retain favorable therapeutic attributes from the templates while sampling novel variants that may alter binding potency or other functional properties. The study frames the CKEA approach as a generative tool for producing humanized antibody sequences for downstream structural and experimental evaluation.
To assess structural and functional viability of generated designs, the authors performed molecular dynamics (MD) simulations on the modeled antibody–PD-1 complexes. MD analysis provided detailed information about free-energy landscapes and helped identify stable binding conformations among the generated variants.
These computational evaluations were used to prioritize candidates for experimental testing. The MD-derived landscapes and conformational stability metrics were described as providing a strong basis for subsequent laboratory validation, enabling the team to select designs with plausible structural integrity and binding modes for expression and biochemical characterization.
Designs generated by the CKEA were taken forward to laboratory testing. The authors reported expression and purification of six designed antibodies followed by assays to determine binding to PD-1. Of these six designs, three exhibited some level of measurable binding to PD-1 after expression and purification.
The authors note that the binding properties observed in the expressed designs could likely be improved through additional computational approaches such as saturation mutagenesis or via laboratory evolution methods. The experimental results therefore serve as an initial validation of the generative workflow, demonstrating that a subset of AI-designed molecules can show target engagement after synthesis.
Taken together, the computational and experimental findings demonstrate that AI-guided interpolation methods like the CKEA can generate novel antibody designs that retain measurable target binding. The identification of three binders out of six designs indicates the method can produce viable candidates, though further optimization is needed to improve potency and therapeutic attributes.
The authors suggest that following this proof-of-concept, designs could be further refined using computational saturation mutagenesis or empirical laboratory evolution to enhance affinity, specificity, and other desirable properties for therapeutic development. The report frames the CKEA-based workflow as a promising strategy for next-generation antibody discovery that integrates generative modeling, structural simulation, and experimental triage.
This work is presented as a preprint on bioRxiv and has not been certified by peer review. The manuscript lists funding from the National Institutes of Health and includes grant identifiers declared by the authors. The authors declared no competing interests.
Additional metadata provided by the source include authorship and institutional affiliation (Yale University) and the posted date of August 04, 2026. The DOI and preprint accession are reported in the source record.