Practical De Novo Nanobody Discovery with Tens of Experimental Candidates

Conventional antibody discovery relies on animal immunization or large-scale library screening,
requiring experimental interrogation of millions to billions of candidates with limited programmability
over epitope specificity and molecular properties. Recent advances in generative
modeling have demonstrated the possibility of de novo antibody design, yet achieving experimentally
practical biologics discovery at low experimental throughput across therapeutically
relevant targets remains challenging.

Here we present MMDesign, a de novo VHH discovery platform combining generative
sequence–structure optimization with multi-tier computational filtering. Starting from only
a target protein and specified epitope residues, MMDesign generates tens of thousands of
nanobody candidates through optimization over the confidence landscape of MMFold, a proprietary
antibody–antigen structure prediction model, together with a protein language model.
Successive filtering stages incorporating structural reliability, sequence naturalness, and physicsbased
interface evaluation compress the candidate pool to only tens of experimental testable
designs per target. We evaluated MMDesign across 11 therapeutically relevant targets spanning
cytokines, immune checkpoints, receptors, a viral protein, and a surface antigen. Experimental
testing of only 14–50 candidates per target yielded confirmed binders for 10 of 11 targets (90.9%),
with best measured affinities ranging from picomolar to nanomolar levels, demonstrating broad
target-level success at experimentally tractable scale. Notably, MMDesign achieved a 50% hit rate
on TNFα, a shallow trimeric cytokine target that has proven difficult for prior low-throughput de
novo VHH discovery campaigns.

Together, these results suggest that biologics discovery may increasingly transition from largescale
stochastic screening toward programmable molecular engineering, in which computational
generation and prioritization substantially reduce experimental burden while enabling rapid
epitope-targeted therapeutic binder discovery.