Towards an Agentic AI Framework for Generating, Optimizing and Filtering Protein Binders

Mohammad Amaan Sayeed, Boulbaba Ben AmorGenBio @ ICML 2026 · OpenReview

TL;DR

This paper introduces an "Agentic AI" framework that shifts the focus of protein binder design from mere generation to rigorous wet-lab risk minimization. By combining a sequence-based generator (PBIND42) with a multi-agent filtering system (DYRA), the researchers successfully reduced hundreds of candidates down to just 24 high-probability designs, recovering nanomolar-affinity binders for challenging oncology targets.

What Problem Does This Paper Address?

In the world of de novo protein design, we are currently "rich in models but poor in hits." While tools like RFdiffusion can generate thousands of plausible-looking protein structures, many of these candidates are "hallucinations"—they look great on a computer screen but fail miserably in the real world. They might not fold correctly, they might be impossible to manufacture (poor expression), or they might stick to everything in a test tube except the intended target.

The authors identify a "hallucination pathology" where models optimize for structural scores at the expense of biological reality. The gap they tackle is the wet-lab false-positive problem: how do we triage thousands of in silico designs to find the handful that will actually work in a laboratory setting without wasting massive amounts of time and money on failed synthesis?

Methods

The authors propose a five-stage pipeline that treats binder design as a resource-allocation problem rather than a simple ranking task.

  1. Hotspot-Conditioned Generation (PBIND42): Unlike traditional models that require a 3D structure as input, PBIND42 is a purely sequence-based Protein Language Model (pLM). It uses "instruction fine-tuning" to generate binders based on the target sequence and specific "hotspot" residues (the specific "epitope" or landing site on the target).
  2. The Boltz-2 Epitope Gate: An early-stage filter using the Boltz-2 model to ensure the generated binder actually contacts the intended hotspot before any further computation is wasted.
  3. Tri-model Gradient-Surgery Optimization: This is the technical "secret sauce." The system optimizes the sequence using three competing signals: structural/interface quality (Boltz-2), inverse-folding compatibility (ProteinMPNN), and a "manifold prior" (PBIND42). They use gradient surgery to ensure that improvements in binding geometry don't push the sequence into "unnatural" territory that would make it impossible to express in a lab.
  4. DYRA (Dynamic Risk Assessment): A four-agent deterministic triage system. Each "agent" specializes in a different risk category:
  5. Structural Quality Agent: Evaluates global plausibility.
  6. BLI Surface Properties Agent: Predicts if the protein will behave well in common laboratory assays (Bio-Layer Interferometry).
  7. Interface Quality Agent: Checks for "brittle" or unrealistic contacts.
  8. Expression Quality Agent: Uses ESM2 to ensure the sequence looks like a "natural" protein that a cell can actually produce.

Datasets

  • Training: The PBIND42 model was fine-tuned on DIPS-Plus, a curated dataset of 42,112 protein-protein complex structures derived from the PDB.
  • Evaluation/Benchmarks: The pipeline was validated on two oncology targets from the AlphaProteo benchmark: BHRF1 (an EBV protein, considered "easy") and PD-L1 (a major cancer immunotherapy target, considered "difficult" due to its flat, polar surface).

Key Insights and Ideas

  • Sequence-First is Viable: The paper demonstrates that you don't necessarily need a high-resolution 3D structure to start the design process. Language models can capture enough "protein grammar" to generate viable seeds from sequences alone.
  • Assay Compatibility is the Real Bottleneck: One of the most striking findings is that "assay developability"—not predicted interface energy—is the dominant failure mode in the wet lab. The DYRA framework gives the highest weight (55%) to the BLI compatibility agent for this reason.
  • Manifold Preservation: By using "gradient surgery," the authors prevent the optimizer from "cheating." In many AI design tasks, the model finds "adversarial" solutions that score high on metrics but are physically impossible. Keeping the design on the "natural protein manifold" is essential for experimental success.
  • Agentic Triage: Rather than using one "master" score, the framework uses specialized agents to veto candidates. If a design is a "High Risk" for expression, it's out, regardless of how well it's predicted to bind.

Key Results

The framework successfully compressed a pool of 500 candidates per target down to a single 24-design synthesis plate.

  • BHRF1: Recovered a lead binder with a mean $K_D$ (dissociation constant) of 176 nM.
  • PD-L1: Recovered two binders, with the best reaching a mean $K_D$ of 217 nM.
  • Efficiency: The workflow achieved an 8.3% hit rate from a very small synthesis batch.
  • Comparison: In head-to-head replicated lab tests, their BHRF1 binder was ~7x tighter and their PD-L1 binder was ~18x tighter than replicated reference sequences from other state-of-the-art methods (like AlphaProteo) under identical assay conditions.

Limitations and Future Work

  • Target Diversity: The experimental validation currently only covers two oncology targets. The authors are extending the work to others like VEGF-A and TNF-$\alpha$ to prove broader generalization.
  • Hand-tuned Weights: The weights for the DYRA agents are currently set manually based on the authors' experience. Future versions might use machine learning to optimize these weights.
  • Oracle Bias: Since Boltz-2 is used for gating, optimization, and ranking, there is a risk of "residual oracle bias"—the system might be over-fitting to the quirks of that specific model. The authors partially mitigated this by using AlphaFold 3 as an independent post-hoc check.

Summary updated

Abstract

De novo protein binder design is increasingly limited by experimental false positives: many candidates appear plausible in silico but fail during expression, assay, folding or binding affinity characterization. We present an Agentic AI framework for protein binder generation, optimization, and filtering that explicitly treats candidate selection as a wet-lab risk-minimization problem. The framework combines (i) epitope-conditioned generation with PBind42, (ii) tri-model optimization using Boltz-2, ProteinMPNN, and a PBind42 prior, and (iii) deterministic multi-agent filtering system, DyRA, that evaluates structural confidence, assay compatibility, interface quality, and expression plausibility before synthesis. Across two oncology benchmark targets, BHRF1 and PD-L1, the workflow reduces 500 generated candidates per target to compact 24-design plates and recovers two confirmed binders for each target. The best BHRF1 binder reaches a mean KD of 176 nM, while the best PD-L1 binder reaches a mean KD of 217 nM, with a second PD-L1 binder at 704 nM. These results show that protein-language-model generation can be made experimentally actionable when coupled to manifold-preserving optimization and agentic pre-synthesis filtering, substantially reducing the number of designs that must be synthesized and assayed.
BibTeX
@inproceedings{sayeed2026towards,
  author = {Mohammad Amaan Sayeed and Boulbaba Ben Amor},
  title = {{Towards an Agentic AI Framework for Generating, Optimizing and Filtering Protein Binders}},
  booktitle = {GenBio @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=11gDwrQyrO},
}
Figure