Autonomous De Novo Protein Binder Design with Claude

Claude Science¹, Amir Shanehsazzadeh¹

Abstract

Deep-learning methods for protein structure generation, sequence design, and structure prediction now enable de novo binder design against many targets using only tens of designs. A design campaign nonetheless demands expertise that spans target biology, structural modeling, and a fast-moving set of computational tools, as well as days of orchestrating software and compute. We asked how much of this expertise and labor an AI agent could supply. We wrote the working knowledge of a binder design campaign into a single protocol prompt that specifies no epitope, scaffold, or sequence for any target. Working from it, and without human input into any design decision, Claude Opus 4.8 and Mythos Preview ran 24-to 48-hour campaigns against 16 targets. They researched each target, chose epitopes, installed and ran open-source protein design and structure-prediction models, optimized their candidates in silico, and delivered 30 ranked designs per target. Two independent contract research organizations synthesized every design exactly as delivered and measured its binding, with 15 of the 16 targets giving interpretable measurements. Claude designed binders against 14 of them, and 354 of 1,320 designs bound, a hit rate of 27%; among the designs ranked first for each target in each campaign, 49% bound. Designing against all targets at once in a single 48-hour session, Mythos Preview and Opus 4.8 achieved hit rates of 26.7% and 22.6%; designing against one target at a time in 24-hour sessions, Mythos Preview’s hit rate was 35.1%. On the E3 ligase subunit RBX1, recently the subject of an open design competition in which 9 of 245 de novo designs bound, 28 of Claude’s 90 designs bound. The tightest bound with a K_D of 3.9 nM, against 45 nM for the competition’s winning entry re-synthesized and measured on the same plate. Although cross-species reactivity was only a secondary objective of the prompt, 130 of the 233 binders tested against the mouse ortholog of their target also bound it. Every model Claude used is open-source, which places campaigns of this kind within reach of any laboratory. We release the prompts, computational models of all 1,440 designs, and the binding data for the 1,320 designs with reliable measurements as a reproducible protocol for autonomous binder design and a benchmark dataset for the field.¹