Computational Evolution of Anti-PD‑1 Antibodies Induces Structural Refolding for High-Affinity Interactions

Checkpoint
inhibitors targeting the PD-1/PD-L1 axis are key immunotherapies,
but the dynamic and flexible nature of PD-1 complicates rational antibody
engineering. Here, we use computational saturation mutagenesis, AlphaFold
prediction, and molecular dynamics (MD) simulations to evolve pembrolizumab
variants with suitable binding. Seven engineered antibodies form additional
salt bridges and hydrophobic contacts via refolding of both the antibody
and the PD-1 interface. One variant, m7p.5, displays improved biphasic
kinetics and high-affinity binding (KD,apparent = 62 pM). Structural changes
include an α-helix to loop transition in the antibody heavy
chain and a 4.6-Å Cα shift of a PD-1 loop. These results
show that computational evolution can access binding modes inaccessible
to traditional rigid structural design, enabling high-affinity antibodies
for flexible targets. It is demonstrated that our integrated computational
approaches including MD simulations can generate new picomolar high-affinity
antibodies targeting specific epitopes of proteins that may be intrinsically
flexible and are difficult to target with reasonable computational
cost, which would be far less than an experimental cost for finding
new antibodies with equivalent binding affinities. This study provides
a new tool that can be combined with other artificial-intelligence-based
antibody generation against PD-1 from the existing anti-PD-1 antibody
library with broad applications in protein–protein interactions.