Antibody design with steerable discrete diffusion

Antibodies are central to modern therapeutics, yet their discovery and optimization remain constrained by limited integration between computational design and empirical validation. Here, we present ADAPT, a generative antibody language model designed to operate in a seamless closed-loop cycle between \textit{in silico} generation and high-throughput experimentation. ADAPT conditions antibody sequence generation on antigen context using an order-agnostic autoregressive discrete diffusion framework, enabling both localized complementarity-determining region (CDR) design and \textit{de novo} full-length variable chain generation. We show that ADAPT learns biologically meaningful representations that capture global organization of antibody–antigen systems alongside local sequence organization within antibodies and antigens. Across computational benchmarks, ADAPT generalizes to antigens withheld during training. Experimentally, we validate ADAPT-generated nanobodies using yeast display. We further show that feedback can be used to steer generation toward specific functional properties, including improved humanness. Together, ADAPT provides a flexible and experimentally compatible framework for antibody design, enabling coordinated user defined optimization and rapid exploration of functional antibody sequence space.