Agent-guided epitope excision enables economical de novo binder design
Abstract
De novo protein design campaigns remain bottlenecked by the inference compute time required to generate and rank candidates, limiting wider functional exploration. Here, we introduce and evaluate an epitope excision approach that truncates targets to local interface windows to accelerate both backbone diffusion and structure prediction. Across 15 targets and 3,666 designs, excised windows closely reproduced full-target binding interfaces (median CαRMSD of 0.37 Å) and made it practical to design binders against the 719-residue extracellular domain of fibroblast activation protein alpha (FAP) by cutting RFdiffusion and AlphaFold3 GPU time by ~ 88% in one of the epitopes. Pairing this workflow with an agentic optimization loop allowed the system to autonomously identify catalytic-site targeting strategies. Under a fixed compute budget, a human-in-the-loop (HITL) workflow raised the computational hit rate from 5.06% to 9.33% while further cutting GPU hours by ~ 42%. We also found that omitting target multiple sequence alignments during refolding roughly substantially increased the apparent yield. Wet-lab validation with yeast display identified nanomolar binders from the generated designs with the HITL and the fully autonomous agent (human-out-of-the-loop, HOOTL) arms recovering more enriched designs than the PI-alone baseline. Together, target excision and agent-guided search offer a practical way to scale de novo binder design to large, therapeutically relevant proteins.
Related
- Conservation guided reverse vaccinology for the design of a broadly immunogenic multi-epitope vaccine against monkeypox virus
- Half-match recombination drives bridge RNA-guided excision and off-target insertion
- CRISPR RNA-Guided Integrase Enables Targeted Genome Engineering and Prophage Functional Genomics in Plant-Pathogenic Pseudomonas
- Native hyphal alignment enables spatially anisotropic functions in Pleurotus eryngii-derived membranes for guided bone regeneration
- CRISPRi-mediated silencing of racR reveals molecular strategies for surviving lethal stress coupled to Rac prophage excision
BibTeX
@misc{kim2026agentguided,
title = {{Agent-guided epitope excision enables economical de novo binder design}},
author = {Jun W. Kim and Artiom Myaskovsky and Oscar Donosa and Lawrence Chan and Jamie Rogers and Avinatan Hassidim and Yossi Matias and Annalisa Pawlosky},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.33195},
eprint = {2610.33195},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.33195},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Agent-guided epitope excision enables economical de novo binder design
AU - Jun W. Kim
AU - Artiom Myaskovsky
AU - Oscar Donosa
AU - Lawrence Chan
AU - Jamie Rogers
AU - Avinatan Hassidim
AU - Yossi Matias
AU - Annalisa Pawlosky
PY - 2026
DA - 2026/10/23
PB - arXiv
DO - 10.48550/arxiv.2610.33195
UR - https://arxiv.org/abs/2610.33195
AB - De novo protein design campaigns remain bottlenecked by the inference compute time required to generate and rank candidates, limiting wider functional exploration. Here, we introduce and evaluate an epitope excision approach that truncates targets to local interface windows to accelerate both backbone diffusion and structure prediction. Across 15 targets and 3,666 designs, excised windows closely reproduced full-target binding interfaces (median CαRMSD of 0.37 Å) and made it practical to design binders against the 719-residue extracellular domain of fibroblast activation protein alpha (FAP) by cutting RFdiffusion and AlphaFold3 GPU time by ~ 88% in one of the epitopes. Pairing this workflow with an agentic optimization loop allowed the system to autonomously identify catalytic-site targeting strategies. Under a fixed compute budget, a human-in-the-loop (HITL) workflow raised the computational hit rate from 5.06% to 9.33% while further cutting GPU hours by ~ 42%. We also found that omitting target multiple sequence alignments during refolding roughly substantially increased the apparent yield. Wet-lab validation with yeast display identified nanomolar binders from the generated designs with the HITL and the fully autonomous agent (human-out-of-the-loop, HOOTL) arms recovering more enriched designs than the PI-alone baseline. Together, target excision and agent-guided search offer a practical way to scale de novo binder design to large, therapeutically relevant proteins.
ER -