AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design

Researchers introduce AgentPLM, a protein language model enhanced with real-time biophysical feedback and tool integration to generate optimized protein sequences. The system combines reasoning-augmented decoding with a novel training approach, achieving state-of-the-art performance on enzyme design, antibody optimization, and structural stability tasks.

AgentPLM represents a fundamental shift in how protein design systems approach sequence generation. Traditional protein language models operate as static predictors—they generate sequences in isolation without checking whether those sequences satisfy real-world physical constraints. This paper solves that limitation by introducing an agentic framework that interrupts generation to consult external tools like ESMFold for structure prediction, FoldX for energy calculations, and AutoDock Vina for binding affinity assessment. The system learns when to trust these feedback signals versus when to continue generation, a distinction that passive models cannot make.

The technical innovation—Contrastive Agent Policy Optimisation—trains the model end-to-end using trajectory-level preferences rather than simply imitating high-fitness sequences. This allows the agent to learn the strategic value of consulting external feedback at decision points, rather than memorizing which sequences tend to work. The evaluation spans multiple protein design challenges with controlled dataset splits to prevent overfitting, providing credible validation across enzyme design, antibody engineering, thermostability, and protein-protein interaction scenarios.

For the biotech and synthetic biology sectors, this approach accelerates the development cycle for engineered proteins by enabling real-time constraint satisfaction during generation. Rather than generating candidates and filtering them post-hoc, AgentPLM corrects course during generation itself. This compounds efficiency gains, particularly in applications where sequence space is vast but viable solutions are constrained. The framework's modularity—swapping different external tools—suggests broad applicability across protein engineering domains where computational feedback can guide design.

• →AgentPLM integrates real-time biophysical feedback into protein sequence generation, moving beyond passive prediction to active constraint satisfaction.
• →Contrastive Agent Policy Optimisation trains models to strategically use external tools rather than merely imitating high-fitness sequences.
• →State-of-the-art results on antibody optimization and enzyme design demonstrate measurable improvements over previous approaches.
• →The framework supports online error correction without explicit backtracking, reducing computational overhead compared to iterative redesign.
• →Modular tool integration enables deployment across diverse protein engineering applications with different constraint requirements.