henhaobio
Sign in

bioRxivBiochemistryPreprint

Reliable and Efficient Agentic Protein Design with Toolset Harnessing

Jasper ButcheriD, Jie CheniD, Andrew Hunt, Cyrus M. HaasiD, Wenzhao DaiiD, Shajesh SharmaiD, Raktim MitraiD, Saman SalikeiD, Jeremy G. KoobiD, Aiko MuraishiiD, Zhe LiiD, Harley PylesiD, Michael Dominik TykaiD, Magnus S. BaueriD, Rafi I. Brent, Anant MittaliD, Neil P. KingiD, David BakeriD

PDFDOICode

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (20 Oct 2026) · CC BY 4.0

Chaperone, a DataFrame-based framework and agent harness for protein design tools, cut tokens 86% and GPU-hours 64% while raising pass rate from 6% to 12%.

Protein design pipelines chain independently built tools that lack standardized data formats or APIs, and end-to-end systems such as BindCraft are mostly predefined, so swapping in new models or metrics means editing their code. Chaperone offers a lightweight shared DataFrame interface that doubles as an agent harness.

Key results

  1. ShownAcross nine autonomous design tasks, GPT-5.6 Sol agents using the Chaperone harness used far fewer tokens and GPU-hours than baseline agents. They also reached a higher mean computational pass rate on held-out scorers.

    Nine tasks, 81 runs (three replicates at each of three reasoning efforts): mean pass rate 6.1% vs 11.7%; tokens per run down 86%, GPU-hours per run down 64%. Harness higher in six tasks, similar in two, lower for PD-L1. · Abstract; 3 In silico evaluation; Figure legends

  2. ShownRaising agent reasoning effort inflated baseline token use much more than harnessed use. The authors read this as extra reasoning being spent on workflow plumbing rather than on design exploration.

    For baseline agents, each step up in reasoning effort added a median of 25 million tokens. Harnessed Chaperone agents stayed under 10 million tokens in total, even at the highest effort setting (27 runs per effort). · 3 In silico evaluation; Figure legends

  3. ShownA two-week agent-assisted cortisol small-molecule binder campaign gave experimentally confirmed low-nanomolar binders with no affinity maturation. The best binder beat the earlier 68 nM cortisol binder from the prior work cited.

    Of 27 candidates tested by fluorescence polarization, 15 gave supported fits without QC flags, nine gave flagged fits, three showed no clear response. Best KD was 1.32 nM, vs 68 nM for Lee et al. after maturation. 104 small-scale runs preceded scale-up. · 4.1 Small molecule binder design; Figure legends

  4. SuggestedChaperone reproduced a multi-tool tetrahedral nanoparticle design workflow that includes RPXdock, but only a small fraction of tested designs assembled.

    15 designs were selected for experimental validation; one assembled into a nanoparticle (6.7% success), confirmed by negative-stain electron microscopy. · 4.2 Nanoparticle design; Figure legends

  5. SuggestedFor C2/D3 symmetric-complex linker design, a pilot sweep fixed linker length and MPNN variant, then the same pipeline was scaled to roughly 10x the pilot. Crystals of design H11 were observed by brightfield microscopy.

    Scale-up covered about 4,500 placements with seven lengths each; 4,355 placements (97.9%) used the pilot window centre. H11 has a 17-residue β-hairpin linker bridging an 8.9 Å gap; crystals were seen only by brightfield microscopy. · 4.3 Protein fusion; 5 Conclusions; Figure legends

Methods, data and reproducibility

Each tool is wrapped as a block that maps a Polars DataFrame of designs (one row each, with structure history and metrics) to an extended DataFrame; an engine chunks work across jobs via Submitit. The library, docs, examples and skills are given to LLM agents and tested against held-out scorers on nine tasks.

  • Agent evaluation on nine design tasks (protease, symmetric trimer, peptide, RNA, PD-L1, DNA, biotin ligand, antibody CDR, glycosidase); 96 designs submitted per run · n = 81 runs across 9 tasks · PDB 1EC6, 4ZQK, 1VFB (task inputs)
  • Cortisol binder campaign: 80 RFD3 runs in the sweep, then candidates tested by fluorescence polarization · n = 27 candidates tested
  • Tetrahedral nanoparticle designs containing a trimeric coiled-coil motif, checked by negative-stain EM · n = 15 designs
  • C2/D3 symmetric-complex linker designs from prepositioned motifs; pilot and scaled rounds · n = about 4,500 placements
Methods & tools
Chaperone, RFdiffusion3 (RFD3), ProteinMPNN / LigandMPNN, AlphaFold3, RPXdock, CaGE, Rosetta, ESMFold2, RoseTTAFold3, Submitit, Polars, GPT-5.6 Sol via Codex (Harbor Controller)
Software & environment
Harnessed agents ran with internet disabled and containers pre-installed, using a git-state-pruned copy of the Chaperone repository; intermediate outputs stored as Parquet; Submitit-based scheduler.
Assay conditions & compute
Each agent run had 12 hours and a maximum of five B4000 GPUs via a SLURM shim, on a shared cluster with priority queuing.
Reproducibility gaps
No URLs or accessions for the evaluation tasks, held-out scorers, or FP and sequencing data; prepositioned crystal motifs come from an unpublished pipeline (manuscript in preparation).
Limitations (4)
  • Tool-use evaluation is compute intensive and stochastic, with task-by-task variability. Only GPT-5.6 Sol was evaluated in depth; other model families and harnesses are left for future work. · stated by the authors
  • The harness-versus-baseline comparison does not separate the effects of workflow orchestration from structural reasoning and experimental activity. · stated by the authors
  • Wet-lab support is thin and has no non-harness comparison: one of 15 nanoparticle designs assembled, and the H11 crystals were shown only by brightfield microscopy. · noted in this summary
  • Arms differ in setup (baseline installed tools from the web; harness had pre-installed containers), and absolute pass rates stay low (6% to 12%), with the harness lower on PD-L1. · noted in this summary

Read this if you build or evaluate LLM agents for protein design and want token, GPU-cost and pass-rate numbers for a harnessed versus baseline setup.

Citations
0

Abstract

Computational protein design workflows require the orchestration of multiple tools, data formats, and analysis steps that are often fragmented. Here, we introduce Chaperone, a lightweight and customizable framework that enables seamless integration of RFdiffusion3 (RFD3) with downstream sequence design, structure prediction, and evaluation tools, including ProteinMPNN, AlphaFold3 (AF3), and more than 50 other protein design tools. By combining a programmable agent-friendly interface with parallelized execution on high-performance computing systems, Chaperone as a harness enables rapid iteration and scalable exploration of complex protein design workflows. Across an evaluation of nine biomolecular design tasks, we find using Chaperone reduces total token usage relative to baseline agents by 86% and GPU costs by 64%, while increasing the overall computational design pass rate. We demonstrate Chaperones utility for large-scale design-space exploration through a two-week end-to-end small-molecule binder design campaign, yielding low nanomolar binders to cortisol zero shot. We further demonstrate its versatility across distinct design challenges, including the design of a symmetric 12-subunit self-assembly nanoparticle and a linker that mediates interactions between C2- and D3-symmetric protein complexes to form crystalline assemblies. Together, these results demonstrate how agent harnesses built around standardized tool interfaces can substantially improve the efficiency and success of complex, long-horizon agentic design workflows in biology. We make Chaperone freely available to the community on GitHub: https://github.com/RosettaCommons/chaperone.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISXu '26Olivieri '26Chacón-Sánc '25Kim '26Jeon '26Silke '26
71% semantic matchBioinformatics
Xiaopeng Xu, Chenjie Feng, Chao Zha+4 more·Protein Science·Mar 25, 2026· 6 citations