Origin-1: Experimentally Validated Generative AI Platform for De Novo Antibody Design Against “Zero-Prior” Epitopes

Simon Levine-Gottreich, Paulina Kulytė, Jacob Stern, David Grayson, Jonathan Edward King, Umberto Lupo

Keywords: Generative AI, Antibody Design, Zero-Prior Epitopes, All-Atom Modeling, Co-Design, Experimental Validation, Cryo-Electron Microscopy, Affinity

TL;DR: Origin-1-1 integrates all-atom generative modeling, confidence-guided scoring, and experimental validation to design functional antibodies against zero-prior epitopes.

Abstract: Generative artificial intelligence has advanced antibody discovery, yet \textit{de novo} design against epitopes lacking structural precedent remains a fundamental challenge. Here we present Origin-1, a generative AI platform that integrates epitope-conditioned all-atom structure generation (AbsciGen), paired CDR sequence design (IgDesign2), and a co-folding--based scoring protocol (AbsciBind) to select high-confidence binders. We test Origin-1's ability to design antibodies against ``zero-prior'' epitopes, or target sites with no available antibody--antigen or protein--protein complex structures. Across ten human protein targets, Origin-1 identified antibodies for four targets in fewer than 100 design attempts each. Cryogenic electron microscopy confirmed atomic accuracy (3.0-3.1Å resolution; DockQ 0.73-0.83) and AI-guided affinity maturation produced an IL36RA antagonist with 104 nM potency. These results establish a framework for antibody design without structural precedent.

Submission Number: 48