We present an end-to-end Agentic AI framework for de novo protein binder design comprising three stages: Prot42-guided generation, multi-objective optimization, and DyRA multi-agent risk assess- ment. Each stage is implemented as autonomous AI agents that coordinate through structured interfaces, automating progression from target specification to lab-ready candidates.

The pipeline begins with Prot42, a 1.1 billion parameter autoregressive protein language model [STN+25] instruction-tuned on binding pairs. Given a target protein and hotspot residues defin- ing the binding epitope, Prot42 samples binder sequences encoding epitope specificity such that sequences interact with designated binding sites while avoiding off-target regions. These seeds enter a multi-objective optimization loop combining multiple loss terms1: Prot42 contributes a negative log-likelihood loss encouraging target binding, Boltz-2 (a diffusion-based structure predictor) adds contact, pTM, and pLDDT losses, and ProteinMPNN provides sequence recovery probability from predicted structures ensuring designability. Optimization proceeds via simplex-constrained proximal gradient descent with temperature annealing, followed by quality filters rejecting homopolymers, repeated motifs, and low-entropy sequences

Optimized candidates undergo structure prediction with Boltz-2 and Rosetta interface analysis before entering DyRA (Drugability and Yield Risk Assessment), a hierarchical multi-agent system with four GPT-5.2-powered agents calibrated against Elite binders to different targets. The Structural Agent (15% weight) evaluates structure prediction confidence. The BLI Agent (70% weight) assesses surface properties predicting biolayer interferometry success. The Interface Agent (15% weight) analyzes Rosetta energetics. Each specialist produces a score (0–100) and risk tier (Low ≥55, Medium 40–54, High <40).

The Coordinator Agent orchestrates the pipeline, computing weighted scores and applying deterministic qualification: a candidate qualifies if no agent assigns High Risk and at least one assigns Low Risk. GPT-5.2 generates natural language justifications while scores and decisions remain deterministic.

The framework exhibits key agentic properties: autonomy, specialization across structural biology, surface chemistry, and binding energetics, structured communication via JSON interfaces, and interpretability through metric-backed reasoning. This modular architecture enables independent agent updates, provides human-readable audit trails, and extends readily to additional specialists. By calibrating against lab-validated patterns, the framework produces qualification decisions that predict experimental outcomes, reducing the experimental burden in therapeutic protein binder development

References [STN+25] Mohammad Amaan Sayeed, Engin Tekin, Maryam Nadeem, Nancy A. ElNaker, Aahan Singh, Natalia Vassilieva, and Boulbaba Ben Amor. Prot42: a novel family of protein language models for target-aware protein binder generation, 2025