Agent-Based RBX1 Binder Design with an LLM-Guided Multi-Stage Pipeline Hypothesis Designing binders for RBX1 is a complex challenge that requires precise engagement on a functional partner-facing surface. We hypothesize that an LLM-guided, agent-based workflow can successfully translate qualitative literature evidence into explicit, actionable generative design constraints. By anchoring our design around highly supported primary residues while including exploratory secondary positions , we hypothesize that we can balance biological specificity with the structural diversity needed to prevent premature convergence on a narrow binding hypothesis.

Methodological Approach We developed an automated multi-stage pipeline utilizing Claude Code / OpenClaw, where an LLM agent serves as both an executor and a decision layer. The workflow progresses through the following stages:

Evidence-Guided Target Definition: The agent reviewed existing literature to establish a practical target model focused on primary anchors and exploratory positions.

Backbone Generation and Sequence Design: We provided these target residues to RFdiffusion to construct binder backbones specifically aimed at this spatial interface. RFdiffusion generated an initial ensemble of 10,000 diverse backbones. To broadly explore sequence space, ProteinMPNN subsequently generated 48 sequence proposals per backbone, resulting in 480,000 raw designs.

Agent-Driven Filtering: To manage the massive generated library, the agent rigorously filtered designs based on complete outputs, optimal sequence recovery, and favorable ProteinMPNN scores. This reduced the pool to 515 unique, high-confidence candidates while meticulously preserving structured metadata.

In Silico Evaluation: The agent automated Boltz-2 evaluation, computationally identifying interfaces by calculating the minimum heavy-atom distance (< 5.0 Å) to our target residues.

Asymmetric Scoring & Prioritization: To validate our biological hypothesis, the agent applied an asymmetric scoring rule: primary anchors received the highest statistical weight, while exploratory positions provided secondary value. Final candidates were ranked by combining provenance, generation context, Boltz-2 metrics, and the agent's internal evidence-validation scores.

Additional Context & Results Our agent-based framework inherently makes evidence weighting, parameter tradeoffs, and candidate prioritization fully explicit and auditable.

In-depth characterization of the top 100 matched designs revealed a selective distribution of target engagement.

Across all evaluated complexes, the pipeline demonstrated strong structural confidence with a mean Boltz-2 score of 0.864 and a mean ipTM of 0.844.

The top-performing candidates successfully engaged the critical high-confidence anchors while partially engaging exploratory positions, achieving ipTM values up to approximately 0.923.

Ultimately, this workflow successfully validated our foundational biological hypothesis by seamlessly integrating generative design, automated filtering, and rigorous structure-based rescoring.