A large language model operating without human input into any individual design decision produced confirmed protein binders against 14 of 15 testable targets, including several considered difficult by conventional computational methods, according to a technical report published August 18, 2026 by Anthropic.
The work, led by Amir Shanehsazzadeh at San Francisco-based Anthropic, describes a system in which Claude Opus 4.8 and a second model, Mythos Preview, each ran complete protein binder design campaigns — from target research and epitope selection through structure generation, sequence optimization, and candidate ranking — guided solely by a ~16,000-word protocol prompt. Once the expert-written protocol was fixed, no specialist scientific input was provided during individual design decisions. The authors describe the work as foundational research with no current clinical translation announced.
What the system did
The agent orchestrated a pipeline of open-source tools to generate novel single-chain miniproteins of 50–120 residues designed to bind biologically relevant surfaces on target proteins. Across 16 structurally and functionally diverse targets — including cytokines, cell-surface receptors, viral glycoproteins, and enzymes — the system generated 1,440 designs in total, of which 1,320 were analyzed. Experimental validation was performed independently and blinded by two contract research organizations: Switzerland-based Adaptyv Bio, using cell-free expression and surface plasmon resonance (SPR)/bio-layer interferometry (BLI), and South San Francisco-based Twist Bioscience (Nasdaq: TWST), expressing designs as human IgG1 Fc fusions in HEK293 cells with high-throughput SPR arrays.
Binding was confirmed for designs against 14 of 15 targets with interpretable data; one target, mature GDF-8, was excluded due to aggregation artifacts. The overall hit rate across 1,320 analyzed designs was 27% (354 confirmed binders), rising to 49% for top-ranked designs. Of the 354 confirmed binders, 194 bound below 100 nM, 90 below 10 nM, and 42 below 1 nM. Cross-species reactivity was achieved without explicit per-target engineering: 154 of 179 binders tested also bound the cynomolgus ortholog, and 130 of 233 bound the mouse ortholog.
Difficult targets and a competitive benchmark
Two results stand out scientifically. Against TNF-alpha (TNFα), a compact homotrimer whose receptor-binding grooves span subunit interfaces, the system produced 12 confirmed binders on four distinct backbones, with the tightest apparent KD of 0.70 nM. The authors note that multiple prior de novo design efforts had reported no binders against this surface.
Against RBX1, an E3 ligase subunit that had been the subject of an open community design competition in which 9 of 245 submitted designs bound, Claude's campaigns produced 28 of 90 binding designs, with a top binder at 3.9 nM — compared to 45 nM for the competition's winning entry measured on the same assay plate. The authors note that Claude's designs largely avoided the CUL1-occluded face of RBX1, with only 17% of footprint residues overlapping that region.