Related Experiment Video
Updated: Jun 27, 2026
Validation of Nanobody and Antibody Based In Vivo Tumor Xenograft NIRF-imaging Experiments in Mice Using Ex Vivo Flow Cytometry and Microscopy
Published on: April 6, 2015
Physics-Informed Artificial Intelligence Design of Picomolar Nanobodies Enables Deep Tumor Penetration and
Ning Shi1,2,3, Caiping Ren1,3,4, Liang Zhang2
1Department of Neurosurgery, Xiangya Hospital, Xiangya School of Basic Medical Science, Central South University, Changsha, Hunan 410008, China.
Abstract:
The clinical utility of nanobodies in solid tumor therapy is constrained by a fundamental biophysical trade-off: rapid renal clearance necessitates half-life extension, which in turn demands ultrahigh affinity to prevent dissociation from the target under systemic washout conditions. While generative artificial intelligence has substantially advanced structure prediction, it often fails to resolve the subtle energetic frustrations at protein-protein interfaces required for affinity maturation. Here, we present a physics-informed artificial intelligence framework that integrates AlphaFold 3 structural priors with molecular dynamics simulations to rationally design a picomolar anti-carcinoembryonic antigen nanobody. By employing variable dielectric molecular mechanics/generalized Born surface area decomposition, we identified interfacial residues that were structurally permissible but thermodynamically suboptimal. We subsequently constructed a focused library to resolve these bottlenecks through electrostatic optimization, desolvation penalty minimization, and van der Waals packing refinement. This strategy achieved a 99% binding positivity rate and yielded variants with picomolar affinity (KD ≈ 44 pM)-an ~306-fold improvement over the parental clone-without compromising thermal stability (T m > 63 °C). To translate these biophysical gains into therapeutic efficacy, we engineered bispecific nanobodies fusing the affinity-matured domains with an anti-human serum albumin binder. In vivo longitudinal imaging of colorectal cancer xenografts revealed a "lock-and-hold" phenotype, characterized by deep intratumoral penetration and sustained retention (>168 h). This work demonstrates that coupling geometric deep learning with rigorous physical principles overcomes the inefficiencies of stochastic screening, providing a valuable framework that may be adapted for the rational development of high-potency biologics across various therapeutic targets.
