AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design
Researchers have developed AgentPLM, a novel protein language model designed for more effective protein sequence design. Unlike traditional models that generate sequences in a single pass, AgentPLM integrates external biophysical feedback through a process called Reasoning-Augmented Decoding. This allows the model to consult tools like ESMFold and FoldX, and learn when to use this feedback via Contrastive Agent Policy Optimisation. The model demonstrates state-of-the-art performance in various protein design tasks, including enzyme and antibody design, by exhibiting online error correction capabilities. AI
IMPACT Enhances protein design capabilities by enabling models to incorporate external feedback for improved accuracy and error correction.