AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design
Summary
AgentPLM introduces an agentic protein language model designed to overcome the limitations of passive PLMs by integrating external biophysical feedback. This model employs Reasoning-Augmented Decoding (RAD), which interleaves autoregressive sequence generation with calls to external tools like ESMFold, FoldX, and AutoDock Vina. It also utilizes Contrastive Agent Policy Optimisation (CAPO), a trajectory-level extension of direct preference optimisation, to learn when oracle feedback is truly informative. AgentPLM achieves state-of-the-art results across benchmark tasks including de novo enzyme design, antibody optimisation, thermostability, PPI interface design, and zero-shot fitness prediction. Notably, it demonstrates a gain in antibody top-10% hit rate over the strongest passive baseline, providing mechanistic evidence of online error correction without explicit backtracking.
Key takeaway
For research scientists focused on protein sequence design or antibody optimization, AgentPLM offers a significant advancement by integrating real-time biophysical feedback. You should consider this approach for tasks requiring online error correction and adherence to thermodynamic or structural constraints, as it demonstrably improves top-10% hit rates. Explore its Reasoning-Augmented Decoding and Contrastive Agent Policy Optimisation to enhance your model's ability to learn from external oracle tools.
Key insights
AgentPLM enhances protein language models by integrating external biophysical feedback and learning to effectively utilize it during sequence generation.
Principles
- Interleave autoregressive generation with tool calls for online error correction.
- Train policies to discern informative oracle feedback from mere high-fitness imitation.
Method
AgentPLM uses Reasoning-Augmented Decoding (RAD) to interleave generation with tool calls (ESMFold, FoldX, AutoDock Vina) and Contrastive Agent Policy Optimisation (CAPO) to train the policy on trajectory-level feedback.
In practice
- Design de novo enzymes with improved properties.
- Optimize antibodies for enhanced performance.
- Predict protein thermostability and PPI interfaces.
Topics
- Protein Language Models
- Agentic AI
- Protein Sequence Design
- Reasoning-Augmented Decoding
- Antibody Optimization
- ESMFold
Best for: AI Scientist, Research Scientist
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.