📄 bioengineering

Agentic campaign control for high-throughput de novo binder design

This paper introduces T-REX, an agentic framework that orchestrates multiple protein generative models and evaluators using LLMs to dynamically decide between rescuing, exploring, or exploiting design strategies, thereby optimizing high-throughput de novo binder creation within finite compute budgets.

Original authors: Jeon, M., Song, J., Kim, J., Zhong, E. D.

Published 2026-09-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jeon, M., Song, J., Kim, J., Zhong, E. D.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The quest to design new proteins from scratch is one of the most promising frontiers in modern biology. Proteins are the molecular machines that carry out nearly every task inside living cells, from building tissues to fighting infections. For decades, scientists have tried to engineer custom proteins to act as medicines or industrial catalysts, but the process has been slow and expensive. It requires generating millions of potential shapes, testing them on computers to see if they fold correctly, and then discarding the vast majority that fail. In recent years, artificial intelligence has revolutionized this field by creating tools that can predict protein structures and generate new sequences with remarkable speed. However, these tools are often specialized; one might be excellent at creating a specific shape, while another is better at refining the chemical sequence that holds it together. The challenge for researchers has become not a lack of tools, but a lack of strategy: how does one effectively manage a finite amount of computing power when faced with a suite of different, powerful, and sometimes contradictory AI programs?

A team of researchers at Princeton University and the Massachusetts Institute of Technology has introduced a new approach to solve this problem, treating the design process not as a single task but as a dynamic campaign. They developed a system called T-REX, which acts as an intelligent manager for a fleet of different AI protein designers. Instead of running one program until it finishes or blindly trying every possible combination, T-REX uses a large language model to constantly monitor the results of ongoing experiments. It watches for signs of success, failure, or stagnation, and then decides in real time whether to rescue a promising but flawed design, explore a completely new method, or double down on a path that is already producing good results. This system orchestrates six different generation tools, a sequence refinement tool, and a structure evaluator, coordinating them across multiple graphics processing units to maximize the number of unique, high-quality protein binders created within a fixed budget of computing time.

The researchers tested this system against seven different biological targets, which are specific molecules that a new protein is designed to attach to. They compared T-REX against several other strategies, including using a single AI tool in isolation and using simpler, non-intelligent algorithms to manage the workflow. The results were clear: T-REX consistently produced more structurally distinct hits than any of the other methods. Across all seven targets, the system generated a significantly higher number of unique protein designs that met strict quality standards. In some cases, it produced more than four times as many successful designs as the next best alternative. The system achieved this by adapting its strategy as the campaign progressed. When a particular method stopped yielding new results, T-REX would shift its resources to a different tool. When a design came close to success but failed on a specific metric, the system would direct a specialized tool to fix that specific weakness rather than starting over.

One of the most significant findings was how the system handled failure. In traditional design campaigns, a stalled process often leads to wasted time as researchers wait for a single tool to finish or manually intervene. T-REX, however, treats different types of failure as distinct signals. If a design fails because its overall shape is unstable, the system might try a different generation method. If it fails because the chemical sequence is slightly off, it might switch to a refinement tool to tweak the sequence while keeping the shape. This ability to diagnose the specific reason for a setback and choose a targeted response allowed the system to recover from stalled periods much faster than the other methods. In simulations where the system was forced to stop using its "rescue" strategy, it took significantly longer to find new successful designs, proving that this adaptive recovery mechanism is a key driver of its success.

The designs produced by T-REX were not only more numerous but also more diverse. The system generated proteins with a wide variety of shapes, including those rich in helical structures and those dominated by sheet-like formations. It also found binders that attached to the target molecules in different locations, suggesting a broad exploration of the possible ways a protein could interact with its target. This diversity is crucial for drug discovery, as different shapes and binding positions can lead to different therapeutic effects or avoid resistance mechanisms. The researchers verified that these designs were not just statistical anomalies but held up under rigorous testing, passing multiple layers of computational evaluation that check for stability and binding strength.

While the study was conducted entirely in a computer simulation, the implications for real-world science are substantial. The work demonstrates that the future of protein design may not lie in building a single, perfect AI model, but in creating intelligent systems that can manage and combine many different models. By treating the design process as a continuous loop of observation, reasoning, and action, T-REX showed that a relatively brief period of reasoning can guide much more expensive and time-consuming computations. The system is now available as open-source software, allowing other scientists to use this orchestration framework to accelerate their own work. This approach suggests that the next leap in biological engineering will come from better management of our computational tools, turning a collection of specialized programs into a cohesive, adaptive team capable of solving complex design challenges.

Technical Summary: Agentic Campaign Control for High-Throughput De Novo Binder Design

Problem Statement
The field of de novo protein design has rapidly expanded to include a diverse ecosystem of specialized computational tools, including generative models for backbone structures, inverse folding for sequence design, and structure evaluators. However, a critical bottleneck remains: how to effectively orchestrate these complementary tools within a finite compute budget. Traditional approaches often rely on static workflows or single-generator baselines that fail to adapt to the specific failure modes of a design campaign. As campaigns progress, candidate designs may miss in silico metrics, duplicate existing structures, or fail due to specific weaknesses (e.g., poor interface confidence vs. poor fold confidence). Without adaptive control, computational resources are often wasted on unproductive routes or fail to capitalize on promising but imperfect candidates. The challenge is to dynamically allocate compute across generation, redesign, and evaluation tasks based on accumulating evidence to maximize the throughput of structurally distinct, high-quality binder hits.

Methodology: T-REX Framework
The authors introduce Target-adaptive Rescue–Explore–eXploit (T-REX), an agentic campaign controller designed to orchestrate multiple state-of-the-art protein generative models and evaluators. T-REX operates as a closed-loop system comprising three main components:

Structured Evidence Representation:
Instead of relying on scalar hit counts, T-REX constructs an EvidenceSummary that aggregates campaign outcomes into interpretable categories. It identifies four operational failure modes:
- No Output: Generation tools produce no usable structure.
- Awaiting Eval: Results are pending evaluation.
- Near Miss: Candidates narrowly miss qualification criteria (e.g., pLDDT or iPAE thresholds).
- Duplicate: Qualified designs cluster with previously found structures.
  Based on these modes, the system assigns one of seven campaign states (e.g., "low evidence," "stall," "deep stall," "productive-with-duplication") to provide high-level context for decision-making.
Agentic Reasoning (LLM Agents):
T-REX employs two Large Language Model (LLM) agents to reason over the structured evidence:
- Planner: Analyzes the EvidenceSummary and campaign history to formulate HypothesisCards. These cards propose evidence-linked hypotheses and specific follow-up computational tests (e.g., "Test broader Feynman-Kac search because recent designs failed only on iPAE").
- Supervisor: Takes the valid candidate jobs proposed by the Planner and assigns them to one of three strategic priorities:
  - Rescue: Directs computation to address a specific weakness in a promising design or route (e.g., sequence redesign for a near-miss).
  - Explore: Tests underused methods or alternative configurations to diversify the search.
  - eXploit: Allocates additional compute to routes that are currently yielding new structural hits.
    The Supervisor also proposes a dynamic allocation ratio (R/E/X) for subsequent jobs.
Deterministic Execution Layer:
To ensure reliability and resource efficiency, a deterministic controller validates the LLM's proposed actions against the fixed action space and resource constraints. It manages asynchronous execution across multiple GPUs, allowing new jobs to start as soon as resources become available without waiting for all concurrent jobs to finish. This layer enforces the "Check + Build" validation, ensuring that LLM proposals are feasible before dispatch.

Key Contributions

Agentic Orchestration: T-REX is the first framework to integrate LLM-based contextual reasoning with a structured failure-mode analysis for de novo binder design campaigns, moving beyond static pipelines or simple scalar-reward bandits.
Adaptive Strategy: The system dynamically shifts between Rescue, Explore, and eXploit strategies based on real-time campaign states, allowing it to recover from stalls and capitalize on productive routes more effectively than non-agentic baselines.
Open-Source Release: The authors release T-REX as open-source software, including the agent prompts and campaign-control framework, to support compute-efficient binder design workflows.

Results
The authors evaluated T-REX across seven distinct protein targets (CD45, BetV1, CbAgo, HER2-AAV, SC2RBD, PDL1, and IL7RA) under a fixed compute budget of 144 H100 worker GPU-hours.

Throughput Superiority: T-REX achieved the highest throughput of structurally distinct hits (SUs) on all seven targets. It outperformed a non-agentic tree-search controller (PUCT) by a geometric mean of 2.43-fold and the strongest generator-only baseline for each target by 2.48-fold.
Robustness: The performance gains were consistent across different structural diversity definitions (TM-score thresholds of 0.5, 0.6, and 0.8) and sequence clustering (70% identity). T-REX also retained the highest number of hits under stricter post-hoc quality thresholds (e.g., binding free energy predictions).
Stall Recovery: Ablation studies demonstrated that the Rescue priority is critical for recovery. When the Rescue priority was removed, the system took significantly longer to recover from "stall" or "deep stall" states (median 5.04 vs. 1.77 GPU-hours for "stall") and spent a much larger fraction of time in stalled states (27.7% vs. 3.8%).
Diversity: The designs produced by T-REX spanned diverse secondary structures (helix-rich, mixed, and $\beta$ -strand-rich) and, in the case of the CbAgo target, occupied a broader range of predicted binding positions compared to PUCT.

Significance and Claims
The paper claims that adaptive orchestration is a complementary and necessary direction to the continued improvement of individual protein design models. By leveraging LLM agents to interpret heterogeneous outcomes and a deterministic controller to manage resources, T-REX demonstrates that "evidence-guided" campaign control can significantly improve the efficiency of de novo design.

The authors remain modest regarding the scope of their claims:

The evaluation was conducted entirely in a computational setting using established in silico metrics (pLDDT, iPAE, scRMSD). While the underlying generators (Proteína-Complexa, BindCraft, BoltzGen) have produced experimentally validated binders in prior studies, T-REX itself was not experimentally validated in this work.
The framework is designed to be extensible to future stages of the design pipeline (e.g., binding affinity optimization, specificity, developability) by adding corresponding tools and adapting the evidence summary, but such extensions were not implemented in this study.
The results highlight that the throughput gains are not specific to a single target or initialization, suggesting that adaptive control is a generalizable strategy for allocating compute across complementary methods in scientific campaigns.

Technical Summary: Agentic Campaign Control for High-Throughput De Novo Binder Design

More like this