AgentPLM: Agentic Protein Language Models with Reasoning-Augmented Decoding for Protein Sequence Design
Sahil Rahman, Maxx Richard Rahman
Abstract
Protein language models (PLMs) are passive oracles: they generate sequences in a single forward pass with no mechanism to consult external biophysical feedback or redirect generation when a candidate violates thermodynamic or structural constraints. We introduce AgentPLM, which addresses this by equipping a pre-trained PLM with i) Reasoning-Augmented Decoding (RAD), which interleaves autoregressive generation with tool calls (ESMFold, FoldX, AutoDock Vina), and ii) Contrastive Agent Policy Optimisation (CAPO), a trajectory-level extension of direct preference optimisation that trains the policy end-to-end to learn when oracle feedback is informative rather than merely imitating high-fitness sequences. We evaluate AgentPLM on benchmark tasks spanning de novo enzyme design, antibody optimisation, thermostability, PPI interface design, and zero-shot fitness prediction with standardised oracle APIs and controlled sequence-identity splits. AgentPLM achieves state-of-the-art results with a gain in antibody top-10% hit rate over the strongest passive baseline, providing mechanistic evidence of online error correction without explicit backtracking.
AI Impact Assessments
(1 models)Scientific Impact Assessment: AgentPLM
1. Core Contribution
AgentPLM proposes converting a pre-trained protein language model (ESM-2) from a "passive oracle" into an agentic system that interleaves autoregressive sequence generation with external biophysical tool calls (ESMFold, FoldX, AutoDock Vina). The two key technical contributions are: (i) Reasoning-Augmented Decoding (RAD), which augments the PLM vocabulary with tool-call tokens and uses a Tool Context Encoder (TCE) and Trajectory Memory Buffer (TMB) to integrate oracle feedback mid-generation; and (ii) Contrastive Agent Policy Optimisation (CAPO), a trajectory-level extension of DPO that trains the model end-to-end to learn *when* oracle calls are informative rather than simply imitating high-fitness sequences.
The conceptual framing — protein design as a POMDP over joint sequence-tool-call space — is intellectually appealing and well-articulated. The paper correctly identifies that standard PLMs cannot observe biophysical constraint violations during generation and that this fundamentally limits their design capability.
2. Methodological Rigor
Strengths in formulation: The POMDP formulation is rigorous, the augmented action space is cleanly defined, and the distinction between sequence-advancing and non-advancing actions (tool calls) is principled. The structural self-consistency (SSC) score as a "safety net" for mandatory tool calls is a thoughtful design choice.
Concerns about evaluation:
3. Potential Impact
If the approach generalizes beyond computational surrogates, AgentPLM could meaningfully advance protein engineering by providing a principled framework for integrating heterogeneous biophysical feedback into generative models. The key ideas — tool-augmented generation for scientific domains, trajectory-level preference optimization for multi-step scientific reasoning — have broad applicability beyond proteins (materials design, drug discovery, chemical synthesis).
The learned temporal division of labor among oracles (Figure 4) is an interesting emergent behavior suggesting the model discovers a rational design strategy. If validated experimentally, this would be a compelling demonstration of AI-driven scientific reasoning.
However, practical impact is contingent on several factors: (a) wet-lab validation of designed sequences, (b) scalability to longer proteins and more complex design objectives, and (c) the computational cost of oracle calls during inference (acknowledged as ~4.1s/sequence, bottlenecked by Vina).
4. Timeliness & Relevance
The paper addresses a genuine and timely gap. The protein design community has recognized that PLMs encode evolutionary constraints but struggle with out-of-distribution design targets. The agentic AI paradigm (ReAct, tool-augmented generation) is mature in NLP but nascent in computational biology. This paper is well-positioned at the intersection of these trends.
The CAPO objective is timely given the rapid adoption of DPO and related preference optimization methods. Extending these to trajectory-level, multi-step tool-use settings is a natural and important direction.
5. Strengths & Limitations
Key Strengths:
Key Limitations:
6. Additional Observations
The paper's writing quality is high and the presentation is clear. The figure design effectively communicates key results. However, the claim of "2.79× improvement in antibody top-10% hit rate" in the abstract appears to reference a different number than what Table 2 shows (52.41% vs 27.38% ≈ 1.91×), suggesting inconsistency between abstract claims and reported results.
The ZeroShot-Fitness results (Spearman ρ = 0.61) are notably modest compared to specialized methods on ProteinGym, suggesting CAPO fine-tuning may degrade the evolutionary prior for pure fitness prediction tasks.
Generated Jun 2, 2026
Comparison History (24)
Paper 1 presents a novel, methodologically rigorous integration of LLM agents and biophysical tools for protein design, offering direct, transformative applications in drug discovery and biotechnology. While Paper 2 offers a valuable evaluation framework for AI reliability, Paper 1 introduces a tangible algorithmic advancement (CAPO) that solves a critical bottleneck in computational biology, likely leading to more immediate and profound real-world scientific breakthroughs.
Paper 1 likely has higher scientific impact due to stronger real-world applicability and cross-disciplinary reach: integrating PLMs with structural/energetic/docking oracles targets therapeutics and enzyme engineering with direct experimental and industrial pathways. The RAD+CAPO framework is a novel, generalizable agentic paradigm for biology that could influence protein design workflows broadly. Paper 2 is timely and rigorous, with major benchmark gains in formal math, but its near-term practical impact is narrower (primarily theorem proving/formalization) and may be more benchmark-driven. Overall, AgentPLM’s translational potential and breadth favor higher impact.
Paper 2 likely has higher scientific impact due to clearer, near-term real-world applications (protein/enzyme/antibody design) and broader translational relevance (biotech, therapeutics). Its methodological contribution (tool-augmented decoding plus an end-to-end policy-optimization objective) is novel and can generalize to other scientific design problems. The evaluation appears more rigorous: multiple benchmark tasks, standardized oracle APIs, and controlled identity splits. Paper 1 is highly novel for mechanistic interpretability at scale and important for AI safety, but it explicitly notes major limitations in completeness and faithfulness evaluation, which may temper immediate impact.
Paper 1 introduces a highly innovative paradigm for protein design by integrating biophysical tool-use directly into the model's generation loop via a novel optimization strategy. This overcomes a major limitation of current passive protein language models. Its potential real-world applications in drug discovery, enzyme engineering, and synthetic biology offer immense, tangible scientific and societal value. While Paper 2 addresses an important and timely AI safety issue, Paper 1's methodological leap in computational biology is likely to catalyze broader, more transformative breakthroughs in the life sciences.
Paper 1 introduces a highly novel agentic approach to protein language models, integrating external biophysical feedback and tool usage. This addresses a major bottleneck in protein design and has immense real-world applications in drug discovery and synthetic biology. In contrast, Paper 2 provides a useful but incremental evaluation of baseline search algorithms, offering significantly less novelty and a narrower scope of impact.
AgentPLM introduces a novel paradigm of augmenting protein language models with agentic reasoning and external tool use, addressing a fundamental limitation of current PLMs. The approach combines multiple innovations (RAD, CAPO) with demonstrated state-of-the-art results across multiple important protein design benchmarks. Its breadth of impact spans computational biology, drug design, and AI methodology. Paper 2, while interesting, reports largely negative results on a narrow BCI application, with the frontal alpha asymmetry signal failing to reliably distinguish emotional states, limiting its immediate scientific impact.
Paper 1 provides a fundamental theoretical result (kernel obstruction theorem) proving why LLMs inherently fail at causal discovery, which has broad implications across all fields using LLMs for scientific reasoning. It offers both a rigorous impossibility result and a constructive solution (A-CBO) with provable convergence guarantees. The theoretical depth, breadth of impact across AI and causal inference, and the paradigm-shifting nature of proving intrinsic limitations of popular training methods (SFT, DPO, ICL) give it higher potential impact than Paper 2, which, while impactful in computational biology, is more domain-specific and primarily engineering-driven.
AgentPLM introduces a highly novel intersection of agentic AI and structural biology. By enabling protein language models to consult external biophysical tools during generation, it addresses a major limitation in computational biology. Its applications in enzyme and antibody design offer profound real-world impacts in medicine and drug discovery, significantly exceeding the more incremental algorithmic improvements to LLM preference optimization presented in Paper 1.
AgentPLM introduces a fundamentally novel paradigm—agentic protein language models with reasoning-augmented decoding and tool-calling during generation—addressing a clear limitation of existing PLMs. It demonstrates state-of-the-art results across multiple protein design benchmarks with broad applications (enzyme design, antibody optimization, drug discovery). The methodological innovation of integrating biophysical feedback into autoregressive generation via CAPO is highly novel and transferable. Paper 1 addresses bilayer material property prediction with multimodal learning, which is useful but more incremental in scope and methodological novelty.
AgentPLM introduces a novel paradigm shift by transforming passive protein language models into agentic systems with reasoning-augmented decoding and tool-calling capabilities, addressing a fundamental limitation in protein design. Its broad evaluation across multiple protein engineering tasks (enzyme design, antibody optimization, thermostability, PPI design) demonstrates wide applicability in a high-impact domain. The combination of LLM-style agentic reasoning with biophysical oracles is timely and novel, bridging AI agents and computational biology. Paper 2, while theoretically rigorous with strong bounds for causal Bayesian optimization, addresses a more niche problem with narrower immediate applications.
While Paper 2 presents a massive engineering achievement in formalizing mathematics, Paper 1 introduces a highly novel agentic approach to protein design with immediate, high-stakes applications in drug discovery and synthetic biology. Integrating biophysical tool-use directly into the decoding process of Protein Language Models addresses a critical bottleneck, likely yielding broader and more immediate scientific and societal impact.
Paper 2 addresses a highly critical challenge in structural biology and drug discovery by integrating agentic AI with biophysical tools for protein design. The ability to use external feedback during sequence generation represents a significant methodological leap with broad real-world implications in biotechnology and medicine. While Paper 1 offers a rigorous statistical approach to emotion modeling in NLP, the profound life-sciences applications and timeliness of generative protein design give Paper 2 substantially higher potential for widespread scientific impact.
Paper 1 offers a concrete, methodologically rigorous advancement in protein design with immediate real-world applications in drug discovery and biotechnology. While Paper 2 presents an ambitious vision for a general scientific operating system, Paper 1 provides stronger empirical validation against standardized benchmarks, making its tangible scientific impact and widespread adoption more probable.
Paper 2 likely has higher impact: it introduces an agentic framework for protein design that tightly couples PLMs with biophysical tools and a novel training objective (CAPO) to learn when/why to query feedback. This is highly timely for AI-driven biology and can translate directly to real-world protein engineering (enzymes, antibodies, PPIs), influencing multiple communities (ML, structural biology, drug discovery). Paper 1 is a clever, efficient architectural fusion for LMs, but its primary impact is within model design/efficiency and may face fast-moving competition from adjacent hybrid attention/SSM methods.
AgentPLM introduces a novel paradigm of agentic protein language models that integrates external biophysical tools during decoding, addressing a fundamental limitation of current PLMs. The combination of Reasoning-Augmented Decoding with a new training objective (CAPO) across multiple challenging protein design benchmarks represents significant innovation with direct applications in drug discovery, enzyme engineering, and therapeutic antibody design. Paper 2, while solid in improving test-time inference for planning via classical search integration, addresses a narrower problem with less transformative potential. The biological applications and methodological novelty of Paper 1 suggest broader and deeper scientific impact.
Paper 2 likely has higher impact due to greater novelty (agentic decoding with tool-augmented feedback plus a new training objective, CAPO), stronger and broader real-world applications (protein/enzyme/antibody design with direct biotech relevance), and wider cross-field reach (ML, structural biology, drug discovery). The methodology appears more rigorous via multiple standardized benchmarks, controlled splits, and tool-based evaluation. Paper 1 is timely and useful for agent safety/monitoring but is narrower in scope, validated on a small labeled dataset (68 diffs) and targets a specific trait-measurement setting.
Paper 2 is more novel and broadly impactful: it introduces an agentic framework for protein sequence design that integrates external biophysical tools during decoding plus an end-to-end policy-optimization method (CAPO) to learn when feedback is useful. This has strong real-world applications across drug discovery, enzymes, antibodies, and materials, and its agent/tool paradigm is timely and transferable beyond proteins. Paper 1 is incremental (DenseNet+LSTM with off-policy RL) and reports modest metric gains, with narrower domain impact and more limited methodological novelty.
Paper 1 likely has higher scientific impact due to stronger novelty and cross-disciplinary significance: it turns PLMs into tool-using, feedback-driven agents for protein design, integrating structure/energy/docking oracles with a new training objective (CAPO) and demonstrating gains across high-value biological tasks (enzymes, antibodies, PPIs, thermostability). This directly targets real-world therapeutic and industrial protein engineering. Paper 2 is timely and useful for systems/agent deployment, but is more incremental (memory/critic framework without parameter updates) and its impact is narrower and more application-engineering oriented than a broadly enabling method for biomolecular design.
AgentPLM introduces a fundamentally new paradigm for protein design by integrating agentic reasoning with external biophysical tools into protein language models. This bridges AI and computational biology with immediate real-world applications in drug design, enzyme engineering, and therapeutics. The novel CAPO training method and tool-augmented decoding represent significant methodological contributions. TriLens is a solid contribution to hallucination detection using internal model signals, but it is more incremental—applying logit-lens analysis in a refined way. AgentPLM's cross-disciplinary impact and practical applications in biotechnology give it substantially higher potential impact.
AgentPLM introduces a novel paradigm of combining protein language models with agentic reasoning and external tool calls, addressing a fundamental limitation of current PLMs. The framework spans multiple high-impact application domains (enzyme design, antibody optimization, protein stability), offers a concrete new training algorithm (CAPO), and demonstrates state-of-the-art results. While Paper 2 provides valuable theoretical insights on multi-model self-consuming loops, its contributions are more incremental—extending known model collapse analyses to multi-model settings. Paper 1's direct applicability to drug discovery and protein engineering gives it broader real-world impact potential.