acceptodds
Under review as a conference paper at ICLR 2027

Agentic campaign control for high-throughput de novo binder design

Abstract

Progress in artificial intelligence has produced a rapidly growing ecosystem of powerful methods for de novo protein design. With access to many capable and complementary protein design models, how do we choose and use them effectively, especially under a finite computational budget? Here, we introduce Target-adaptive Rescue–Explore–eXploit (T-REX), an agentic protein design framework that adaptively orchestrates multiple state-of-the-art protein generative models and structure evaluators over the course of a design campaign. Within \trex, large language model (LLM) agents interpret accumulating campaign outcomes to decide whether to rescue promising candidates, explore alternative settings, or exploit productive ones, while a deterministic execution layer validates and schedules proposed actions. We find that LLM agents can integrate structured feedback, including different categories of failure modes, with semantic reasoning to propose new design strategies. Across seven de novo protein-binder design targets, T-REX achieved the highest throughput of structurally distinct hits under matched GPU budgets, while adapting its allocation across models and strategies in a target-dependent manner. T-REX outperformed a non-agentic tree-search controller by 2.43-fold and the strongest individual generator for each target by 2.48-fold on average. These results suggest that as protein-design models become increasingly capable, adaptive orchestration of these models can provide a complementary route to improving design campaigns.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, de novo binder design, scientific agents, agentic campaign control, adaptive compute allocation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_UepCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

this paper introduces T-REX, a training-free binder design framework that orchestrates diverse generative models under an efficient compute budget. It interprets campaign evidence via three strategies: rescue, explore and exploit. the In-silico experiments show it tops deduplicated hits on all 7 targets under matched budgets, outperforming PUCT and the strongest generator-only baseline per target.

Critical strengths

T-REX has a carefully tailored harness for novel protein binder design, including four failure mode, three allocation priorities, and gpu allocation for effectiveness execution. which well track the evidence such as evidence summary, campaign states.

the experimental gains are significant compare with generator-only baselines, rule-based controllers, and tree-search controllers.

the paper is well written, appendix includes sufficient details, and the figure presentations are look good.

Critical weaknesses

it is well established that tool + LLM planning alone boosts performance. baselines are narrow without peer agentic comparisons such as autoBinder, the claude Science autonomous binder, or recent OpenAI bio agents/skills, nor a generic coding agent as a resource allocator. without these, the marginal value of T-REX's LLM orchestration over tool-using planning alone remains unclear. notably, given the strong capability of current coding agents, it remains unclear whether they could directly plan and allocate jobs to solve the same campaign without T-REX-style harness.

Questions for authors

no token cost for orchestration planning nor module-level effectiveness, leaving the efficiency claim unconvincing.

ablations do not cover other llm model (only qwen3.6-27b)

7 comments

We thank the reviewer for their positive assessment of T-ReX, particularly recognizing T-REX's carefully tailored design, the significant improvements over existing baselines, and the clarity of the paper and appendix. We appreciate the constructive suggestions and have conducted additional experiments to address the reviewer's concerns.

[W1] Comparison with additional agentic baselines

Thanks for the suggestion. We agree that comparisons with general-purpose agentic systems are valuable for understanding the contribution of the T-REX's structured framework. As the reviewer suggested, we implemented an additional baseline based on the Claude Code autonomous binding-design campaign protocol [1]. We note that the Claude binder-design campaign used has a substantially larger compute budget (~$10k per target, roughly 2,500 H100-hours). For a fair comparison, we adapted the protocol to match T-ReX's design objective, target, available design methods, and compute budget, removing platform-specific services such as Slack, cloud-account integration, and external databases. We evaluated both approaches on CD45 using a budget of 50 worker GPU-hours across two random seeds (Section 3.1; SI Section E.5.2, Table S26).

We evaluated the Claude campaign protocol with two LLMs: Qwen3.6-27B, the same model used by T-ReX, to compare the campaign-control frameworks, and Claude Sonnet 5.5, to evaluate the performance of a frontier LLM agent.

FrameworkLLMSeed 0Seed 1MeanLLM token cost ($ per run, mean)Cost per SU ($)
T-REXQwen3.6-27B1391411402.341.69
Claude campaignClaude Sonnet 5.514370106.522.992.42
Claude campaignQwen3.6-27B721068932.743.00

Cost per SU = (50 worker GPU-h x $4.69 + LLM token cost) / SU, pooled over the two seeds (CD45; LLM via API).

For the token price, Qwen 3.6-27B is $0.25 / $2.00 per M input/output tokens while Claude Sonnet 5.5 is $2 / $10 per M input/output tokens.

With the same underlying LLM, T-ReX produced 1.57× more structurally unique designs on average than the adapted Claude campaign protocol. We also observed substantially lower GPU idle time (3%/6% versus 23%/17% across the two seeds), demonstrating the benefit of separating LLM-guided planning from deterministic job scheduling.

Notably, T-ReX achieves these results using the open-weight Qwen3.6-27B model. Even when using the more expensive frontier model Claude Sonnet 5.5, the Claude campaign produced fewer SUs on average (106.5 versus 140), although it slightly outperformed T-ReX in one seed. These results suggest that a specialized campaign-control framework can achieve competitive or better performance without relying on a frontier LLM.

We also appreciate the reference to AutoBinder Agent [2]. AutoBinder automates a predefined design pipeline (MaSIF → Rosetta grafting → ProteinMPNN → AlphaFold3), whereas T-ReX adaptively allocates compute across multiple design methods. Since AutoBinder does not report a comparable compute-constrained binder-design benchmark, a quantitative comparison is not currently possible. We will discuss this work in our revised related-work section.

References

[1] Claude Science and A. Shanehsazzadeh, "Autonomous de novo protein binder design with Claude," Anthropic technical report, August 2026.

[2] Ge et al., "AutoBinder Agent: An MCP-Based Agent for End-to-End Protein Binder Design," arXiv:2602.00019 (2026).

[Q1] Token costs and module-level effectiveness

We have expanded our cost analysis to include both token-based LLM inference costs and the full cost of dedicated GPU serving (see our response to Reviewer VK4s, [W3]; SI Section E.5.1, Tables S12 and S25). Even when charging the full LLM-server reservation, T-ReX retains its throughput advantage over all baselines, with geometric-mean improvements of 1.82× over PUCT and 1.90× over the strongest generator-only baseline.

Regarding module-level effectiveness, our Rescue-priority ablation examines the contribution of explicitly identifying and recovering from stalled design campaigns. Removing Rescue increased the fraction of compute spent in stalled states from 3.8% to 27.7%, demonstrating the importance of using structured failure information to guide subsequent design decisions.

[Q2] Generalization to other LLMs

Following the reviewer’s suggestion, we additionally evaluated T-ReX with Claude Sonnet 5.5, replacing Qwen3.6-27B while retaining the same framework and 50-worker-GPU-hour budget on CD45 (SI Section E.5.3, Table S26).

FrameworkLLMSeed 0Seed 1MeanLLM token cost ($ per run, mean)Cost per SU ($)
T-REXQwen3.6-27B1391411402.341.69
T-REXClaude Sonnet 5.512810411651.812.47
Claude campaignClaude Sonnet 5.514370106.522.992.42

Cost per SU = (50 GPU-h x $4.69 + LLM token cost) / SU, pooled over the two seeds; token prices as in [W1].

These results demonstrate that T-ReX can successfully operate with different underlying LLMs, although performance varies with the model. Our primary contribution is the agentic campaign-control framework rather than a comparison of LLM capabilities. A systematic evaluation across LLM architectures and scales is beyond the scope of this work, but represents an interesting direction for future research, particularly as LLM reasoning capabilities continue to improve.

thanks author for the carefully responses, which address some of my concerns; while the author also acknowledge and include more baseline and experiments, this is good helping readers to fairly understand the gains. and it looks T-Rex is more efficient. while my concerns are still on the significance and innovation between T-Rex and existing agents solutions. thus I maintain my original rating.

Reviewer_VK4sOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This work proposes T-REX, an agentic campaign-control framework for de novo protein binder design. However, as an application-oriented study, the paper is written in a highly engineering-focused manner. Apart from introducing the protein binder design task, much of the abstract is devoted to describing the system design and implementation details, for example: “T-REX leverages large language model (LLM) agents to reason over accumulating outcomes and decide whether to rescue promising candidates, explore alternative settings, or exploit productive routes, while a deterministic controller validates and schedules proposed actions.” This makes it difficult to clearly understand the actual scientific objective of the work with respect to the protein binder design task.

Critical strengths
  1. The experimental results are relatively comprehensive, and the experimental procedures and analyses are described in considerable detail.
  2. The motivation is clear. Rather than proposing a new protein generation model, the work focuses on improving the use and coordination of existing models.
Critical weaknesses
  1. The presentation is highly engineering-oriented and shows a strong AI-style writing pattern, including the way the supplementary materials are written and organized.
  2. The architecture diagram is also highly engineering-oriented, making it difficult to see a clear connection between the framework design and the protein binder design task itself.
  3. The computational cost of LLM serving is not explicitly included in the reported budget. It is therefore unclear whether the claimed performance and computational-efficiency advantages would remain if this cost were taken into account.
Questions for authors

How much of the performance improvement is actually attributable to LLM reasoning? If the H100 cost for LLM serving, as well as token/inference costs, are included in the total computational budget, does T-REX still maintain its advantage over PUCT, ϵ\epsilon-greedy, and the generator-only baselines? The paper is written in a highly engineering-oriented style. Could the authors clarify how AI tools were used in conducting this work and preparing the manuscript, and to what extent they contributed to the overall work?

5 comments

We thank the reviewer for recognizing the clear motivation and comprehensive experimental evaluation. We appreciate the feedback and address each point below.

[W1] Presentation and writing style

We respectfully disagree with the characterization of the manuscript as exhibiting an “AI-style writing pattern.” The manuscript was carefully written and revised over many drafts to maximize its clarity, especially given the complexity of agentic systems. We created all figures from scratch (in Keynote). As disclosed in our AI Use Statement and submission form, coding assistants helped implement methods, and AI tools helped find related work and check grammar. If the reviewer has any specific concerns or feedback on our presentation, please let us know so that we can address them!

[W2] Connection between the framework and binder design

Thank you for this feedback. We apologize that the connection between Figure 1 and the protein binder design task was not sufficiently clear. We intended panels (a–c) to motivate the design of T-ReX and panels (d–e) to illustrate the T-ReX controller and its execution. The central objective of T-ReX is to answer a concrete binder-design question: Under a fixed GPU budget, which design tools and settings should be run next to maximize the number of structurally distinct, qualifying binder designs? Each panel connects directly to this objective:

  • Fig. 1a defines the protein design task, available generators (Complexa, BindCraft, BoltzGen), and success criteria (pLDDT, iPAE, scRMSD).
  • Fig. 1b identifies distinct failure modes that inform subsequent design decisions.
  • Fig. 1c introduces the Rescue–Explore–eXploit allocation priorities for responding to these outcomes.
  • Fig. 1d illustrates how the T-ReX controller integrates evidence through the course of a design campaign to propose, validate, and prioritize subsequent design actions.
  • Fig. 1e shows how these actions are scheduled and executed across GPUs under a fixed compute budget. We have clarified these connections in the revised figure caption.

[W3] LLM-serving costs

We thank the reviewer for raising this important point. We focused on worker GPU-hours as the primary compute budget, since the LLM is invoked only intermittently (approximately once every 6.5 minutes, with an average inference time of 27 seconds per call). Across seven campaigns, cumulative LLM-call latency was 23.5 hours, approximately 7% of the 336 (7 targets x 48 GPU-h for LLM server) hours for which the dedicated single-GPU LLM server was reserved. (SI Section E.1.3).

Importantly, even when charging the full LLM-server reservation, T-ReX maintains its throughput advantage across all seven targets (Table S12):

ComparisonRange across 7 targetsGeometric mean
vs PUCT (best among non-LLM controllers)1.03–3.27×1.82×
vs strongest generator-only baseline1.07–4.52×1.90×

Moreover, the computational cost of T-ReX is dominated by protein generation and evaluation, rather than LLM inference. Across seven campaigns, LLM inference cost just $37.39 in total ($5.34 per campaign), representing approximately 0.78% of the total estimated cost when using API-based inference. This highlights that LLM-guided orchestration introduces relatively little additional cost compared with the computationally intensive protein design models it coordinates.

We added a comparison of total monetary cost per structurally unique design (SU), accounting for both worker GPU costs and LLM inference costs:

Monetary cost per SU.

MethodCost per SUBaseline / T-REX (API)
T-REX (LLM via API: worker GPUs + tokens)$3.26-
T-REX (paper setup: worker GPUs + full vLLM server)$4.31–
PUCT$8.192.5×
ε-greedy$23.067.1×
SMAC3 (100 GPU-h budget)$9.983.1×
Complexa-only$8.462.6×
BindCraft-only$7.992.5×
BoltzGen-only$12.713.9×

With a dedicated server, T-REX is 1.9–5.3× cheaper than every baseline.

[Q1] How much of the improvement comes from LLM reasoning?

Thank you for this question. We compare against three no-LLM controller baselines that use scalar SU/GPU-h reward: tree search (PUCT), ε-greedy selection, or Bayesian optimization (SMAC3). All these baselines share T-REX’s action space and evaluation pipeline. T-REX produced 2.5-7.1x more SUs at the same budget. As shown in [W3], the advantage still remains after LLM costs are included.

Below, we show the comparison between the no-LLM controller and T-REX total SUs (Table S11). The numbers are summed over seven targets.

No-LLM ControllerWorker GPU-h/targetT-REX SUsComparator SUsRatio
PUCT1441,4615772.5×
ε-greedy1441,4612057.1×
SMAC31001,1083293.4×

This comparison keeps the action space and evaluation fixed and changes only the allocation policy: LLM reasoning over structured evidence versus scalar-reward search (SI C.5).

[Q2] Use of AI tools

We appreciate the opportunity to clarify. The scientific ideas, methodological design, experiments, analyses, and conclusions were developed by the authors. We wrote and revised the manuscript over several iterations with assistance from AI tools, but the scientific content and writing decisions were our own. While some plots were generated with the help of AI-assisted coding tools, all figures were created from scratch in Keynote.

As disclosed in our AI Use Statement and submission form, coding assistants helped implement methods, and AI tools assisted with finding related work and checking grammar. If the reviewer has any specific concerns, we would be happy to address them.

Dear Reviewer VK4s,

We want to note that the reviewer-author discussion period will be closed tomorrow. As mentioned in our previous comments, we tried to resolve all your concerns and suggestions in our response and revision.

We would like to ask the reviewer to be willing to update their score in a positive direction. If the reviewer thinks that there still exists room for improvement, please let use know. We will do our best to improve our submission.

Reviewer_1Z1oOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces T-REX, an agentic framework that coordinates protein generators and structure evaluators under a compute budget. This framework is feedback/outcome-conditioned, and reusing previous discoveries efficiently to produce new designs. The LLM agent plays the role of deciding which generator method to use: four Complexa search variants, BindCraft, BoltzGen, ProteinMPNN redesign, or simple conduct a AF2 structure evaluation.

Critical strengths

The paper is a good example of how to design well-informed system that leverages LLM's strength but also relied on years of genetics' models (AlphaFold 2, BindCraft, ProteinMPNN). This design space could have been huge and it's not directly trivial. What should the role of LLM be in protein design? [1] believes LLM should do literature/database search, produce captions, and then embed the text to then jointly train another sequence denoising model. However, [2] thinks LLM is capable of outputting an entire protein sequence on its own. Clearly this paper took a very reasonable stance that LLM should serve as an orchestrator to use existing protein design models. This puts LLM as a scientist, leveraging its tool use capabilities and knowledge. I find this design to be non-trivial, clever, and well-thought-through. Is this the final design or even the best design? I'm not quite sure. However, the entire LLM for Bio field is trying to figure out what is the best way to set up this task -- and this is a very valuable contribution to the effort.

Critical weaknesses
  1. The evaluation size is very small (7 targets). This is not a super super big problem. Every research group / lab has different budget level. The novelty of idea and carefulness of thought matter the most. However, [1] evaluated on 20 targets, and [2] evaluated on more. Again, this is not asking the researchers to simply evaluate more -- the size does matter but it's not the most important part.

  2. In figure 1, panel e, having 4 GPUs, 1 for LLM serving, 3 for evaluation and generation (re-design), is certainly interesting -- almost mimicking the idea of "async RL". However, this problem is not nearly as complex as async / pipeline RL. It is unclear to me what is the point of asynchronous evaluation here? Do you compare against a synchronous baseline and notice performance parity and massive speed gain? It feels like the authors just put the figure there to mimic pipeline RL without truly doing anything substantial.

[1] Lee, Joseph, Yanjing Li, Bo Qiang, Gregory Lau, Allyson Ettinger, Noah A. Smith, Rory Donovan-Maiye, David Baker, Hanna Hajishirzi, and Pang Wei Koh. "Designing Protein Structures from Natural Language." arXiv preprint arXiv:2610.13917 (2026).

[2] Kshirsagar, Meghana, et al. "RosettaSearch: Multi-Objective Inference-Time Search for Protein Sequence Design." arXiv preprint arXiv:2604.17175 (2026).

Questions for authors
  1. For "scRMSD", correct me if I'm wrong, I'm under the impression that this compares a generated sequence with a reference sequence and is only a valid metric if the two share the same length. How do you handle this metric (feeding it back to T-REX for the next experiment) if it's N/A?

  2. Are you guys able to send some of the high quality designed proteins to the wet lab and validate your search results? Also not a deal breaker, just curious!

2 comments

We thank the reviewer for their positive feedback, particularly their recognition of T-ReX as “non-trivial, clever, and well-thought-through” and “a very valuable contribution,” as well as their assessment of the paper's clarity and soundness. We address their comments below.

[W1] Number of evaluation targets

Thank you for the suggestion. Given our limited compute budget, we focused on an in-depth evaluation of seven targets to demonstrate the effectiveness of our approach. Across these targets, our study comprises 74 campaigns, including ablation studies and technical replicates, totaling approximately 10,000 H100 worker GPU-hours. Following the suggestion, we added two new targets, and T-REX again outperformed the strongest adaptive baseline (PUCT) on both (1.38x and 1.26x; Section 3.1, SI Section E.5.4, Table S27).

New targets.

TargetT-REX SU (SU/GPU-h)PUCT SU (SU/GPU-h)T-REX / PUCT
TrkA283 (5.66)205 (4.10)1.38×
InsR fragment241 (4.82)192 (3.84)1.26×

SU = Foldseek TM0.6 structure-unique hits under the same qualification criteria as the paper. Because of the compute resource limitation, these campaigns use a 50 worker-GPU-h budget instead of 144 (used in the original submission). We also note that [1] and [2] (that reviewer mentioned) address a complementary design level of the problem:

  • [1], posted after our submission, is a text-conditioned structure generator built on Proteina-Complexa.
  • [2] uses an LLM as an inference-time optimizer that refines sequences for a fixed backbone (backbone-conditioned sequence design).
    Both are design methods that could be added to T-REX's action space, and we can make note of this in our revised manuscript.

[W2] Purpose of asynchronous execution

We appreciate the opportunity to clarify this point. We do not claim a novel asynchronous algorithm or a measured speedup over a synchronous baseline. Asynchronous execution is a practical design choice motivated by the widely varying runtimes of protein generation, redesign, and evaluation jobs. For example, in the CbAgo campaign, a BindCraft job used 7.78 GPU-h on one GPU, while two ProteinMPNN redesign jobs on another GPU took one minute each and moved on to evaluation. A synchronous execution scheme would require the second GPU to wait for the longer-running job before proceeding. In T-ReX, freed workers can immediately execute the next validated job, and completed results can inform subsequent planning decisions without waiting for all running jobs. Our primary contribution is the evidence-guided allocation of compute across design strategies, rather than asynchronous scheduling itself. We will clarify this distinction in the Figure 1 description.

[Q1] scRMSD when lengths differ

scRMSD measures structural self-consistency, rather than agreement with an external reference sequence. Specifically, we compare the designed binder structure against the AF2-predicted structure of that same designed sequence. The corresponding binder structures therefore have identical sequence lengths by construction, so the length-mismatch issue does not arise (SI Section C.1).

[Q2] Wet-lab validation

We agree that having wet-lab validation is an important next step, and are in the process of adding wet lab validation. We note that the underlying generators, including Proteina Complexa, BindCraft, and BoltzGen, have demonstrated experimentally validated binders in their respective studies.

Dear Reviewer 1Z1o,

As we approach the end of the author-reviewer discussion period tomorrow, we kindly request the reviewer to consider updating their score. If the reviewer thinks that there still exists room for improvement, please let us know. We will do our best to improve our submission.

Comments on the paper

1 comment

Hi reviewers,

Thank you for your reviewing service. The authors have provided rebuttals with some new results. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.