A protein's job is not a fixed property of its sequence. The same molecule can suppress tumor growth in one tissue and drive it in another: S100A11 acts as an oncogene in lung and liver but a tumor suppressor in bladder and esophagus The Context Dependent Duality of Cancer Signaling, and INPP4B, classically a tumor suppressor, promotes leukemic survival specifically in NPM1-mutated AML by switching from the AKT pathway to SGK3 Hongjun Jin et al.. Every computational model we have for this makes one passive, feed-forward prediction from a static embedding and cannot tell you which missing piece of context would change the answer — a limitation stated explicitly by the contextual model PINNACLE Contextual AI models for single-cell protein biology, which treats context as a discrete label, and by the query-conditioned retriever ProtQueSt ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation, which is bounded by what its database already contains.
Dialogue research solved a structurally identical problem decades ago: a spoken-dialogue system keeps a probability distribution over what the user might want, updates it softly as noisy evidence arrives, and asks the question that best splits the remaining uncertainty Spoken Dialogue Management Using Probabilistic Reasoning, A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management through Dynamic Stochastic State Evolution. This proposal transfers that machinery where it has never been: deciding a protein's function in a specific cellular context becomes a sequential interrogation in which an agent chooses among cheap database lookups, expensive structural tools, and direct questions to the biologist, each ranked by how much uncertainty it is expected to remove per unit of cost.
Core Claim
The claim is that context-dependent protein function can be resolved as a sequential Bayesian experimental design problem, and that doing so requires two things no existing system provides. First, a single cost-aware action space that mixes internal tool calls and external clarification questions, all scored by expected information gain per cost. Second, a shared probabilistic belief state — a distribution over functional hypotheses — as the communication medium between specialized agents, replacing the hard condition lists and free-text handoffs that current multi-agent systems use.
The closest existing systems each stop one step short. SAGE-Agent Structured Uncertainty guided Clarification for LLM Agents computes the expected value of information for clarification questions, but only over pre-specified API argument slots, as a single agent, with no cost tradeoff between cheap and expensive evidence sources, and only in software and utility domains. GOAgent GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation orchestrates real bioinformatics tools for protein annotation, but its tool order is a policy trained on final-label match, with no belief state and no notion of a tool being not worth its cost; it also evaluates only on static, context-free CAFA5 targets. CA-BED CA-BED: Conversation-Aware Bayesian Experimental Design does soft Bayesian updates with LLM-estimated likelihoods and multi-step lookahead, but only over 111 common objects and 100 murder mysteries, with no tools and no decomposition. Our proposal is the union: BED-style belief tracking, applied to functional hypotheses conditioned on context, driving a heterogeneous tool-and-question action space, executed by specialized agents that read and write only the belief state.
Research Program
The system, which we call an interrogator, maintains a belief $b(m \mid c)$ over functional modes $m$ (for example, promotes-survival versus suppresses-growth, or a conditioned set of Gene Ontology terms) given a partially observed context $c$ — mutation background, tissue, localization state, microenvironmental variables. Evidence $e$ arrives from tool executions or biologist answers and updates the belief softly, following the CA-BED pattern of LLM-estimated likelihoods $P(e \mid m)$ rather than hard accept/reject, with the Rao-Blackwellized correction from BED-LLM BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design to absorb the LLM's own uncertainty about likelihoods. Hard pruning is exactly the failure mode the dialogue literature documents: Wu et al.'s deterministic baselines collapsed under speech-recognition noise while soft top-candidate updates held an 86.7% success rate where random questioning fell to 61.7% A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management through Dynamic Stochastic State Evolution, and biological observations are no less graded than misheard words.
At each step the interrogator selects an action $q$ from the pooled space — a PROSITE motif scan, a GO hierarchy lookup, a Rosetta structural metric, a single-cell atlas query, or a question to the biologist ("is KRAS mutated in this sample?", "is the protein membrane-bound or shed?") — by
$$ q^{*} \;=\; \arg\max_{q \in \mathcal{Q}} \; \frac{H(b) - \sum_{e} P(e \mid q)\, H(b \mid q, e)}{\mathrm{cost}(q)}, $$
where $H$ is Shannon entropy over the belief and $\mathrm{cost}(q)$ is a per-action budget in wall-clock or proxy units. This bisection principle — pick the query that splits remaining probability mass roughly in half — is the established core of optimal question selection Sequential Bayesian Search, Active Embedding Search via Noisy Paired Comparisons; the new part is that the "questions" now include expensive computations, so the value must be normalized by cost, and that the hypothesis space is a biological function-in-context rather than a named entity.
The multi-agent decomposition follows the roles that have earned their keep elsewhere, but rewired through the belief state. A hypothesis proposer (the Thinker role from MACM MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems) generates candidate functional modes and candidate queries; a likelihood critic estimates $P(e \mid m)$ and audits the proposer for confirmation bias, in the prover-verifier pattern whose value MALADE quantified — its critic agents corrected 44.5% of primary-agent outputs MALADE: Orchestration of LLM-powered Agents with Retrieval Augmented Generation for Pharmacovigilance; tool executors wrap the bioinformatics utilities as used by GOAgent; and the selector implements the equation above. The crucial design difference from MACM is that its Judge makes binary accept/reject decisions on conditions — precisely the hard pruning that CA-BED shows fails under noisy evidence — whereas here every agent's output is a likelihood that nudges a shared distribution. Because context variables can be probed as queries, the scheme also begins to address the categorical-context limitation of PINNACLE and CDBridge CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling: the interrogator can actively ask about the continuous or perturbed variables (mutational burden, mechanical compression of the kind that bidirectionally modulates Wnt signaling through AXIN phase separation Volumetric compression regulates the phase separation of AXIN and acts as an operational amplifier to bidirectionally modulate Wnt signaling in organoids) that categorical models cannot represent.
Proposed Experiments
The first experiment builds the evaluation that the field lacks: a benchmark of literature-documented context-dependent function switches, drawing cases from the duality literature — INPP4B in NPM1-mutated AML Hongjun Jin et al., S100A11's tissue-disparate roles The Context Dependent Duality of Cancer Signaling, PP2A-B56α aberrant EGFR signaling under oncogenic KRAS Activation of PP2A-B56α leads to aberrant EGFR signaling and proliferative phenotypes in PDAC, OAS3's dual tumorigenic and immunological roles Pan-cancer multi-omics profiling of OAS3 reveals its immunological and prognostic associations across human cancers, ICAM-1's membrane-bound versus soluble contrast Zhiyuan Qiu et al., spatially compartmentalized PGE2 signaling Amrita Banerjee et al., and ZFAND5's cancer-type-contrasting survival correlations Yijia Wang et al., with contexts grounded in the 156 cell types and 24 tissues of Tabula Sapiens as used by PINNACLE Contextual AI models for single-cell protein biology. Each item presents a protein with a partially specified context; the system must reach the correct context-conditioned call, and the missing variables are exactly what it may query. No reported result exists on Lacuna for any model on this task, which is the point of building it; given the small case count, we pre-register the case list and report interval estimates rather than single accuracies.
The second experiment measures cost-efficiency on existing benchmarks. On CAFA5, where GOAgent's 8B agent reports a Molecular Function score of 0.50 GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation, we plot accuracy against cumulative tool cost and ask whether the interrogator matches that accuracy at a substantially smaller tool budget, or exceeds it at equal budget. On the out-of-distribution retrieval benchmark Prot-Inst-OOD, where ProtQueSt reports Entity-BLEU of 48.79 and an LLM-as-judge score of 6.40/10 ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation, we run with the human channel disabled and compare against those reported numbers. The third experiment runs the full interactive setting with a simulated biologist whose answers carry calibrated noise, measuring number of queries to a correct high-confidence call; the reference dynamics come from the dialogue side, where entropy-minimizing questioning reached a target in 3.3 turns under ideal conditions and survived 86.7% of the time under noise A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management through Dynamic Stochastic State Evolution, and where uncertainty-aware planning gives LLMs markedly more stable question-selection boundaries than self-assessed confidence Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in LLMs.
Baselines
The naive reference point is a zero-shot frontier LLM given the sequence and context description and asked to output the functional call and, if it wants more information, to ask for it. Its analogue has been measured in deduction settings and fails badly — direct prompting scored 27.6% on CA-BED's Detective Cases versus 46.0% for belief-tracking, and BED-LLM gained an average of 37.4 percentage points over direct prompting on 20 Questions CA-BED: Conversation-Aware Bayesian Experimental Design, BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design — but no reported protein-domain number exists on Lacuna, so we will run it ourselves on identical inputs. It is the weakest competitor and calibrates the floor.
The strongest existing competitor on the tool-orchestration question is GOAgent GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation, which reports a CAFA5 Molecular Function score of 0.50 with an 8B backbone and the same tool family we use. To make the comparison fair, we run it with the same backbone (Qwen3-8B), the same wrapped tools, the same training-rollout budget for its GRPO policy, and the same evaluation harness, counting tool invocations under a single published cost table for both systems. Beating it means matching or exceeding its accuracy on CAFA5 while cutting measured tool cost by a stated factor, and beating it decisively on the context-conditioned split, where its context-free training gives it no mechanism to excel. On the interactive track the strongest competitor is a SAGE-Agent-style baseline we implement ourselves — slot-level expected-value-of-information clarification over context variables, single agent, no tool channel Structured Uncertainty guided Clarification for LLM Agents — run on the same simulator and tuning budget; its published numbers (coverage 59.73% with 1.39 questions per task on ClarifyBench; When2Call accuracy lifted from 36.5% to 65.2% by uncertainty-weighted training) establish that this machinery works, which is why it is the bar to clear rather than a strawman.
The key ablation isolates the scheduler: the identical multi-agent scaffold, identical tools, identical belief state, but action selection replaced successively by a GRPO-trained tool-ordering policy in the GOAgent style, by greedy unweighted entropy maximization in the EMDM style, and by a fixed tool sequence. If the cost-aware expected-information-gain scheduler does not dominate these on accuracy-per-cost, the central claim is wrong and we will say so. A second ablation swaps soft updates for hard hypothesis pruning, and a third removes the likelihood critic so the proposing model also scores its own evidence.
Why not just…?
Why not just prompt a big LLM with good instructions? The evidence against this is consistent across every relevant comparison in the literature: single-instance prompting scored 54.68% on Level-5 MATH where the multi-agent Thinker-Judge-Executor decomposition reached 76.73%, and beat Tree-of-Thought by 25 points on the 24-Point Game MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems; direct prompting lost by 37.4 points on average to Bayesian questioning in BED-LLM's own experiments BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design; and prompted agents without structured uncertainty either hallucinate missing parameters or over-clarify trivial ones Structured Uncertainty guided Clarification for LLM Agents. Still, we run this baseline ourselves on our benchmark because none of those results is in the protein domain.
Why not just port SAGE-Agent's EVPI machinery and skip the multi-agent part? This is the most serious alternative, and we treat it as a real threat rather than arguing around it. SAGE-Agent already does POMDP belief tracking and expected-value-of-information questioning; one could apply it to context variables as slots. Its documented limits are rigid API schemas, a single agent with no separation between hypothesis generation and likelihood estimation, and no cost accounting across heterogeneous evidence sources — all three of which our design addresses. But the empirical question is whether those differences matter, so our SAGE-style baseline is exactly this port, and the decisive check is whether schema-slot EVPI degrades when the action space includes expensive tools with noisy outputs and an open hypothesis set. If the single-agent port matches us, the honest conclusion is that the decomposition is unnecessary here, and we will report that.
Why not just use contextual embeddings and predict in one shot? PINNACLE and CDBridge show contextual modeling beats context-free baselines — PINNACLE beat random-walk target nomination in 100% of cell types, and CDBridge reached an $R^2$ of 0.387 on GTEx tissue-specific expression where Enformer scored 0.127 Contextual AI models for single-cell protein biology, CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling — but they are passive: they cannot identify which context variable is missing, cannot ask for it, and, like ProtQueSt, are bounded by existing annotations when the mechanism is uncharacterized. We scope our claim accordingly: for well-annotated, fully specified contexts a static contextual model may win, and we will include it as a passive baseline; our claim covers the ambiguous and under-specified regime, which is exactly where the duality cases live.
Why not just use hard binary pruning over the hypothesis set? Hard pruning is cheaper and simpler, and it fails for the documented reason: one noisy answer permanently eliminates the true hypothesis, which is why soft-update systems held an 86.7% success rate where deterministic questioning fell to 61.7% under noise A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management through Dynamic Stochastic State Evolution and why CA-BED attributed much of its margin over pruning baselines to avoiding premature elimination CA-BED: Conversation-Aware Bayesian Experimental Design. Biological assays and screens are exactly the kind of graded, occasionally wrong evidence this fails on. The pruning ablation will quantify the gap on our data rather than rely on the analogy.
Risks and Decisive Checks
Three risks are genuine. Likelihood miscalibration can corrupt expected-information-gain values into degenerate questioning cycles, a failure BED-LLM's estimator was designed to dampen BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design; we will measure belief calibration directly and treat it as a first-class result. The tool quality ceiling noted by GOAgent bounds what any orchestrator can achieve regardless of scheduling intelligence GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation; the cost-efficiency experiments are designed so the claim survives a low ceiling, since matching accuracy at lower cost is a win even when absolute accuracy is capped. And the duality benchmark is small by construction, since literature-documented context switches number in the dozens; we mitigate with pre-registration, interval estimates, and a held-out split. The falsifiable predictions are concrete: the interrogator should dominate every scheduler ablation on accuracy-per-cost, reach correct context-conditioned calls on the duality benchmark where one-shot contextual baselines cannot, and ask fewer, better-targeted questions than prompted or slot-EVPI alternatives under matched simulator noise. If the single-agent EVPI port matches it, the multi-agent claim dies and the sequential-design claim survives on its own — an outcome we would report as written.
References
- The Calcium Binding Protein S100A11 and Its Roles in Diseases
- INPP4B promotes cell survival via SGK3 activation in NPM1-mutated leukemia
- Contextual AI models for single-cell protein biology
- ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation
- Spoken Dialogue Management Using Probabilistic Reasoning
- A Probabilistic Framework for Representing Dialog Systems and Entropy-Based Dialog Management Through Dynamic Stochastic State Evolution
- Structured Uncertainty guided Clarification for LLM Agents
- GOAgent: Tool-Orchestrating Language Agents for Protein Function Annotation
- CA-BED: Conversation-Aware Bayesian Experimental Design
- BED-LLM: Intelligent Information Gathering with LLMs and Bayesian Experimental Design
- Sequential Bayesian Search
- Active Embedding Search via Noisy Paired Comparisons
- MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems
- MALADE: Orchestration of LLM-powered Agents with Retrieval Augmented Generation for Pharmacovigilance
- CDBridge: A Cross-omics Post-training Bridge Strategy for Context-aware Biological Modeling
- Volumetric compression regulates the phase separation of AXIN and acts as an operational amplifier to bidirectionally modulate Wnt signaling in organoids
- Activation of PP2A-B56α leads to aberrant EGFR signaling and proliferative phenotypes in PDAC
- Pan-cancer multi-omics profiling of OAS3 reveals its immunological and prognostic associations across human cancers
- Roles of intercellular cell adhesion molecule-1 (ICAM-1) in colorectal cancer: expression, functions, prognosis, tumorigenesis, polymorphisms and therapeutic implications
- mPGES-1-Mediated Production of PGE2 and EP4 Receptor Sensing Regulate T Cell Colonic Inflammation
- ZFAND5 Is an Independent Prognostic Biomarker of Perihilar Cholangiocarcinoma
- Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in LLMs