AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design 1. AgentPLM reframes protein sequence design from a one-shot “generate then hope it satisfies constraints” process into an agentic loop: the model can pause mid-generation, query biophysical oracles (ESMFold, FoldX, AutoDock Vina), and then continue with updated context to correct issues online. 2. Core idea: Reasoning-Augmented Decoding (RAD) expands the PLM’s action space beyond amino acids to include tool-call actions. Tool calls do not advance the sequence position; instead, they enrich the context at the same position and the model re-scores the next residue choice, enabling conditional correction without explicit backtracking. 3. A key mechanism in RAD is the Structural Self-Consistency (SSC) score: it measures whether oracle feedback is “surprising” relative to the model’s internal representation of the partial sequence. If SSC drops below a threshold, RAD can force a tool call (subject to min-gap and max-budget constraints), acting as an uncertainty-resolution safety net. 4. Architecture details that make tool-use practical: (i) tool-call tokens are added to the vocabulary and initialized using embeddings from short natural-language tool descriptions (to avoid random special-token training), (ii) a Tool Context Encoder (TCE) uses cross-attention to map heterogeneous oracle outputs (coordinates/pLDDT, scalar ΔΔG, docking vectors) into the PLM hidden space. 5. To avoid blowing up context length with tool outputs, AgentPLM uses a Trajectory Memory Buffer (TMB): it compresses up to Bmax tool-call embeddings into a fixed-size memory vector and injects it via layer-wise gated residual updates, giving O(1) additional memory cost while conditioning on the full tool-call history. 6. Training contribution: Contrastive Agent Policy Optimisation (CAPO) extends Direct Preference Optimisation to full trajectories (residue actions tool calls). It contrasts “winner” trajectories (high fitness with coherent oracle use) against “losers” (low fitness or contradictory oracle signals), teaching not just what sequences look good, but when oracle feedback is worth paying for. 7. CAPO dataset construction: generate 1,000 sequences per task with a reference policy (frozen ESM-2), evaluate with the oracle suite, pair top-10% vs bottom-10%. Tool-call positions are retrospectively annotated using SSC thresholds to mark where oracle feedback would have been informative. 8. Benchmarks span five settings with standardized oracle APIs and controlled splits: ThermoStab-75 (ΔTm with fold-class splits), AntibodyOpt-VH (KD, with 89 antigens fully withheld for test), EnzymeDesign-EC3 (kcat/Km under ≤30% identity splits), PPI-Interface (binding improvement with monomer stability constraint), and ProteinGym ZeroShot-Fitness (no tools at test time). 9. Results: AgentPLM leads across all tasks. Notable gains include AntibodyOpt top-10% hit rate 52.41% vs 27.38% (ProtAgent) vs 12.37% (passive ESM-2), EnzymeDesign normalized kcat/Km 1.89 vs 1.34 (ProtAgent) vs 0.43 (ESM-2), ThermoStab 7.64°C vs 5.19°C (ProtAgent), and stronger PPI improvements (more negative ΔGbind). 10. Mechanistic evidence: trajectory plots show stepwise fitness jumps aligned with oracle calls, consistent with online correction. Integrated-gradient attribution around destabilizing FoldX calls shows attention concentrating locally near the problematic region after the tool output is incorporated; across 500 trajectories, 87% of destabilizing calls trigger significant local attribution increases, while stabilizing calls do not. 📜Paper: arxiv.org/abs/2606.02386 #ProteinDesign #ProteinLanguageModels #ComputationalBiology #MachineLearning #AI4Science #AntibodyEngineering #EnzymeDesign #StructuralBiology #ReinforcementLearning #ToolAugmentedAI