PaperlayerView exact revision on arXiv

Source-linked AI summary

From Retrieval to Reasoning: Agentic Mechanism Prediction from Cell Painting Profiles

Jiayuan Chen, Botao Yu, Tianyu Liu, Thai-Hoang Pham, Meng Wu, Ping Zhang

arXiv:2609.36406v1cs.AI

TL;DR

Cell Painting-based MOA prediction commonly uses morphological similarity to retrieve likely labels, although such similarity is indirect and retrieved neighbors can be unreliable. PhenoAIR treats neighbors as uncertain evidence, integrates it through candidate memory and controlled refinement, and outperforms representation-matching and LLM-based baselines across controlled, realistic, and open-world settings.

  • Problem

    Cell Painting MOA methods commonly infer labels from similar reference perturbations, although morphological similarity alone does not establish reliable or mechanistically consistent evidence.

    Source: §1 · pp. 1–2

  • Method

    PhenoAIR calibrates retrieved evidence and uses phenotype and mechanism agents, shared candidate memory, and controller-guided refinement to produce eligibility-checked MOA predictions.

    Source: §1 · pp. 2–3

  • Results

    Across controlled, realistic, and open-world settings, PhenoAIR outperforms representation-matching and LLM-based baselines.

    Source: §1 · pp. 2–3, §5 · p. 10

  • Takeaways & Limitations

    The results suggest that structured evidence reasoning is most useful when similarity-based evidence is noisy, incomplete, or ambiguous.

    Source: §5 · p. 10

    Self-evolving refinement is restricted to MoA-Novel because its gains are non-monotonic and it does not improve accuracy on MoA-Extended, where it can degrade performance.

    Source: §3.3.2 · p. 9

Abstract

from arXiv · show

Cell Painting is a high-content morphological profiling assay widely used for phenotype-based biological inference, with mechanism of action (MOA) prediction as a central application. Existing approaches largely formulate Cell Painting-based inference as representation matching, assigning predictions from nearby reference perturbations in morphological feature space. However, retrieved neighbors are often noisy and partially misleading evidence due to batch effects, non-specific cytotoxicity, phenotypic convergence, and source-dependent variability. We reformulate Cell Painting-based MOA prediction as a calibrated evidence reasoning problem, where retrieved neighbors are treated as uncertain observations that must be evaluated, compared, and sometimes rejected before supporting a mechanistic conclusion. We propose PhenoAIR, a reliability-aware multi-agent framework that maintains a candidate-centric evidence memory and performs controller-guided refinement over phenotype- and mechanism-side evidence. PhenoAIR uses offline reference-set calibration to weight evidence by source reliability, phenotype stability, and mechanism-level confusion. We evaluate PhenoAIR on a benchmark constructed from JUMP Cell Painting profiles and annotations, covering controlled, realistic, and discovery-oriented open-world MOA prediction settings. PhenoAIR outperforms representation-matching and LLM-based baselines across all settings.

1 Introduction

Cell Painting-based MOA prediction is framed as evidence reasoning because morphological similarity is indirect and retrieved neighbors may be unreliable. PhenoAIR evaluates evidence before making predictions and outperforms representation-matching and LLM-based baselines across three benchmark settings.

1 Introduction

  • Morphological similarity is an indirect signal of shared mechanism, so retrieved neighbors must be evaluated, calibrated, compared, or rejected as evidence.Similarity alone does not establish biological informativeness, reliability, or mechanistic consistency.Source: §1 · pp. 1–2
  • Fixed retrieval depth can miss relevant mechanisms, while source-dependent variability and confusable MOAs make single-profile retrieval insufficient.Cross-source agreement or disagreement can itself inform evidence reliability, and pharmacological context may help distinguish mechanisms with similar morphologies.Source: §1 · p. 2
  • Generic biomedical agents and debate-style systems are not designed to track uncertain Cell Painting neighbors as persistent, candidate-level evidence.PhenoAIR connects evidence calibration, hypothesis assignment, and controlled evidence-seeking actions.Source: §1 · p. 2
  • Across controlled, realistic, and open-world settings, PhenoAIR outperforms representation-matching and LLM-based baselines.The benchmark uses MoA-Verified, MoA-Extended, and MoA-Novel to assess classification, robustness to noisy retrieval, and recognition of unsupported known mechanisms.Source: §1 · pp. 2–3
  • PhenoAIR replaces free-form consensus with a shared candidate memory, offline reliability calibration, and controller-guided refinement over auditable evidence states.Phenotype and mechanism agents update candidate evidence, and an arbiter produces a canonical or novel MOA.Source: §1 · pp. 2–3, Figure 1 · p. 3

2 Method

PhenoAIR replaces neighbor-label aggregation with calibrated evidence integration over candidate MOAs. Its components define the task, prepare reliability resources, update shared candidate states, and refine evidence collection before an eligibility-checked prediction.

2.1 Problem Formulation

  • In closed-set prediction, the true MOA is assumed to be in the known label space; open-set prediction may instead return novel when no known candidate is adequately supported.The standard retrieval formulation aggregates labels from nearest reference compounds, while open-set inference distinguishes unsupported known mechanisms from supported ones.Source: §2.1 · p. 3

2.2 Method Overview

  • PhenoAIR integrates evidence over candidate MOA hypotheses using calibrated retrieval, role-separated agents, controller-selected actions, and an arbiter with deterministic eligibility checks.This replaces nearest-neighbor label aggregation and free-form agent consensus with persistent candidate states.Source: §2.2 · p. 4

2.3 Externalized Reliability Calibration

  • An offline reliability layer annotates reference-set evidence for cross-source consistency, anchor replicability, phenotype stability, MOA separation, generic morphology risk, and recurring confusions.It also constructs a mechanistic family ontology for group-level reasoning when evidence cannot resolve a fine-grained canonical label.Source: §2.3 · p. 4

2.4 Candidate-Centric Evidence Reasoning

  • Candidate memory tracks each MOA’s phenotype support and contradiction, mechanism verdicts, status, and reliability metadata; phenotype and mechanism agents provide distinct evidence updates.The phenotype agent uses Cell Painting evidence, while the mechanism agent evaluates structure-side evidence; their updates communicate through memory rather than free-form dialogue.Source: §2.4 · pp. 4–5

2.5 Controller-Guided Inference Refinement

  • A rule-based controller selects auditable evidence actions and stopping time; the arbiter’s output passes eligibility checks, with novel returned when no canonical candidate has adequate support.For open-set inference, offline self-evolving refinement updates controller rules and arbiter heuristics from development trajectories, with selected artifacts fixed before test evaluation.Source: §2.5 · p. 5

3 Experiment

The experiments evaluate MOA prediction across controlled, realistic, and discovery-oriented settings, using a benchmark built from Cell Painting profiles and curated drug annotations. PhenoAIR outperforms retrieval and LLM baselines, while component tests examine candidate memory, reliability calibration, and iterative refinement.

3.1 Benchmark Construction

  • The benchmark combines public Cell Painting profiles with curated drug MOA annotations to test prediction across three progressively challenging regimes.It is designed to assess robustness to reference quality, source heterogeneity, and incomplete mechanism coverage.Source: §3.1 · pp. 5–6, Table 1 · p. 6

3.1.1 Data Curation

  • Five sources cover approximately 115k unique compounds before annotation filtering; curated multi-label MOAs are supplemented with representative labels for retrieval baselines.The sources include overlapping and source-specific profiles, and one source uses 0.625 μM rather than the 10 μM standard used elsewhere.Source: §3.1.1 · p. 6

3.1.2 Evaluation Settings

  • The benchmark tests high-confidence closed-set, heterogeneous noisy-reference, and open-world settings, with half of MoA-Novel test compounds assigned mechanisms absent from the reference labels.Closed-set metrics include Top-1 and Any-match accuracy; MoA-Novel uses overall accuracy and novel-detection F1.Source: §3.1.2 · p. 6

3.1.3 Baselines

  • Baselines span CellProfiler and CellCLIP retrieval, molecular-fingerprint retrieval, structure-only LLM inference, fixed phenotype-then-mechanism workflows, and single-agent tool use.The single-agent baseline isolates the effects of reasoning decomposition, shared candidate memory, and controller-guided refinement.Source: §3.1.3 · p. 7

3.1.4 Implementation Details

  • LLM-based methods use GPT-5.1 and Claude 4 Sonnet with consistent tool interfaces, curated external sources, and five-run averages for stochastic experiments.Trivial outputs that directly reveal ground-truth labels are excluded to reduce information leakage.Source: §3.1.4 · p. 7

3.2 Main Results

  • PhenoAIR performs best across all three settings and both LLM backbones, while retrieval-only methods remain limited and fixed workflows do not consistently improve over structure-only inference.In MoA-Novel, PhenoAIR improves overall accuracy and novel-detection F1; retrieval baselines are inapplicable because they must assign an existing label.Source: §3.2 · pp. 7–8, Table 2 · p. 8

3.3.1 Component Ablation

  • Adding the multi-agent workflow improves over direct prompting; candidate memory adds consistent gains, while reliability calibration helps most on MoA-Verified and less on MoA-Extended.The smaller MoA-Extended gain is consistent with weak or deeply ranked phenotype evidence limiting local retrieval.Source: §3.3.1 · p. 8, Table 3 · p. 8

3.3.2 Effect of Self-Evolving Refinement

  • Self-evolving refinement improves MoA-Novel results for both LLM backbones, but gains are non-monotonic and MoA-Extended accuracy does not improve and can degrade.The authors restrict refinement to MoA-Novel to avoid development-set-specific rule tuning for closed-set classification.Source: §3.3.2 · pp. 8–9, Figure 2 · p. 9

4 Analysis and Discussion

PhenoAIR’s gains depend on correctly matched phenotype evidence and persist across two morphological feature spaces. Its remaining errors arise from both limitations in the evidence and failures in reasoning.

4.1 Confidence Calibration

  • Shuffling only phenotype evidence reduces accuracy by 26.4 points to near the structure-only agent, while shuffling structure-side evidence leaves PhenoAIR at 63.2% accuracy, indicating complementary evidence streams.Source: §4.1 · p. 9

4.2 Robustness Across Image Features

  • PhenoAIR reaches 60.0% accuracy with CellProfiler and 62.6% with CellCLIP, versus 9.5% and 14.0% for kNN, respectively.Calibration raises Recall@10 from 0.545 to 0.570 for CellProfiler and from 0.575 to 0.605 for CellCLIP; Figure 3 compares these metrics across feature spaces.Source: §4.2 · p. 9, Figure 3 · p. 10

4.3 Limitations and Failure Modes

  • Most MoA-Extended errors reflect pathway or toxicity convergence and retrieval misses, while others involve mechanism interpretation, candidate selection, misleading evidence, or arbitration.The authors identify stronger mechanism grounding, candidate validation, and arbitration under conflicting evidence as future-work directions.Source: §4.3 · p. 10

5 Conclusion

The paper frames Cell Painting-based MOA prediction as evidence reasoning rather than direct representation matching, and presents PhenoAIR as an approach for evaluating retrieved evidence. It situates this claim within prior profiling and agentic-reasoning work and summarizes supplementary materials and inference steps.

5 Conclusion

  • PhenoAIR evaluates the reliability and consistency of retrieved phenotypic evidence before making mechanistic predictions, and structured reasoning is most useful when similarity evidence is noisy, incomplete, or ambiguous.The conclusion suggests combining improved representations with explicit evidence-evaluation and decision-making mechanisms.Source: §5 · p. 10

Appendix

  • The appendix covers prior work, supplementary quantitative and qualitative analyses, component details, benchmark construction and evaluation, and agent prompts and tool schemas.These materials are described in the appendix overview passages.Source: Appendix · p. 13

Algorithm Summary

  • At test time, PhenoAIR uses fixed calibrated retrieval, phenotype and mechanism agents, a controller, and an arbiter; its reliability memory and related artifacts are precomputed from reference and development data.Open-set controller rules and arbiter heuristics may be selected through offline self-evolving refinement, without using test outcomes.Source: Algorithm Summary · p. 13

A.1 Cell Painting Profiling

  • Cell Painting profiles support biological inference, while current pipelines commonly derive conclusions from nearest-neighbor retrieval or embedding-space aggregation despite confounding variation and heterogeneous neighbors.The paper contrasts this representation-matching paradigm with reasoning over uncertain and potentially conflicting phenotypic observations.Source: §A.1 · pp. 13–14

A.2 LLMs for Biological Reasoning

  • Biological LLM systems use retrieval, tools, and agent collaboration, but explicit coordinated reasoning over noisy and conflicting Cell Painting evidence remains largely unexplored.Prior systems include drug-discovery agents, collaborative scientific teams, workflow automation, and multimodal biological models.Source: §A.2 · p. 14

B.1.1 Full Benchmark Results

Across two backbone LLMs, PhenoAIR improves over ablated variants and non-agentic baselines, with its largest gains on MoA-Novel. Its final canonical predictions require phenotype support even when chemical structure provides a strong closed-set control.

Full benchmark results

  • Across GPT-5.1 and Claude 4, PhenoAIR improves over ablated variants and non-agentic baselines, with the most pronounced gains on MoA-Novel.Table 5 reports overall accuracy and novel-MOA precision, recall, and F1; Table 6 reports accuracy and any-match hit rate on MoA-Verified and MoA-Extended.Source: §B.1.1 · p. 14, Table 5 · p. 15, Table 6 · p. 15
  • Structure-only inference is a strong closed-set control, but final canonical predictions must be supported by Cell Painting-derived phenotype evidence.Mechanism-side evidence may propose or prioritize candidates, but does not replace phenotype grounding.Source: §B.1.1 · p. 15

B.1.2 Robustness Across Image Feature Representations

PhenoAIR’s relative gains persist across handcrafted and deep-learning feature backbones, while retrieval preprocessing and aggregation effects vary by benchmark setting. Batch-effect correction improves retrieval results across both settings.

Robustness Across Image Feature Representations

  • Batch-effect correction consistently improves retrieval for handcrafted and learned representations, while deep-learning representations generally outperform CellProfiler features.CellCLIP has the strongest overall retrieval performance across MoA-Verified and MoA-Extended.Source: §B.1.2 · pp. 15–16, Table 7 · p. 16, Table 8 · p. 16
  • Per-source voting improves several Top-5 and overlap metrics in MoA-Verified, but its gains are less consistent in heterogeneous MoA-Extended.The reported difference is associated with more reliable references in MoA-Verified and noisier retrieval neighborhoods in MoA-Extended.Source: §B.1.2 · p. 16
  • PhenoAIR’s relative gains persist across handcrafted and deep-learning feature backbones, indicating that its effectiveness is not tied to a particular embedding space.Stronger representations improve overall performance, but agentic inference gains remain across feature types.Source: §B.1.2 · p. 16

B.1.3 Runtime and API Usage

PhenoAIR incurs greater total inference cost because it uses multiple reasoning rounds, but its compute adapts to query difficulty. Table 10 reports per-query calls, tokens, and wall time across methods using the same base model and hardware.

B.1.3 Runtime and API Usage

  • Table 10 reports average LLM calls, cumulative input and output tokens, and wall time per query, with all LLM methods using the same model and hardware.Source: §B.1.3 · p. 16, Table 10 · p. 17
  • 170k cumulative input tokens across 14 calls (about 12k per call) make PhenoAIR costlier overall, while per-call prompts remain comparable to fixed workflows’ 11k.The higher total cost reflects the number of reasoning rounds rather than inflated single-call prompts.Source: §B.1.3 · p. 17
  • Easy queries typically use approximately 90k cumulative input tokens and 3.5k output tokens, terminating within half the average compute.The controller stops when candidate memory has stable, high-confidence support and continues refinement when evidence is weak or conflicting.Source: §B.1.3 · p. 17

B.2.1 Case Study: AZD8931

The AZD8931 and IT1t cases show PhenoAIR iteratively evaluating phenotype and mechanism evidence: localized phenotype support distinguishes an RTK subtype, while mechanism evidence can introduce a candidate for phenotype verification.

B.2.1 Case Study: AZD8931

  • Refinement moved from noisy PI3K and JAK attractors to EGFR after calibration penalized unstable retrieval anchors while preserving locally visible EGFR evidence.Source: §B.2.1 · p. 18
  • PhenoAIR correctly predicted EGFR inhibitor for AZD8931, resolving an RTK-family ambiguity that target and structural evidence alone could not settle.Source: §B.2.1 · pp. 17–18
  • Mechanism evidence supported RTK inhibition but not subtype; localized calibrated top-20 phenotype support singled out EGFR, which arbitration used as the tie-breaker.Source: §B.2.1 · p. 18

B.2.2 Case Study: IT1t

  • The arbiter accepted the mechanism-introduced candidate with low confidence after phenotype reinspection found one calibrated supporting reference and no contradictory evidence.Source: §B.2.2 · pp. 18–19
  • Calibration reduced confidence in noisy initial phenotype candidates while leaving the sparse CC chemokine signal inspectable for later verification.Source: §B.2.2 · p. 19
  • PhenoAIR correctly predicted CC chemokine receptor antagonist for IT1t after mechanism evidence introduced a CXCR4 hypothesis and phenotype reinspection found sparse support.Source: §B.2.2 · p. 19

B.2.3 Case Study: CH5132799

This section examines PhenoAIR’s evidence-refinement process and its application to the CH5132799 case, where evidence localizes the mechanism to the PI3K–mTOR family without distinguishing two closely related labels. It covers the agents, shared candidate memory, controller-guided refinement, final eligibility checks, and offline development of bounded inference artifacts before test evaluation.

B.2.3 Case Study: CH5132799

  • PhenoAIR predicted mTOR inhibitor but retained PI3K inhibitor, the ground-truth label, as an active alternative within the correctly identified PI3K–mTOR family.Both phenotype and mechanism evidence supported the two pathway siblings, which calibrated retrieval flagged as difficult to separate.Source: §B.2.3 · pp. 20–21

B.3 Self-evolving Case Studies

  • The self-evolving scaffold proposes, evaluates, disables, and refines controller and Arbiter heuristics using development-set trajectories and cross-round changes.It revises controller rules and decision heuristics rather than updating the prediction model directly.Source: §B.3 · p. 21, Table 11 · p. 21

C.1 Externalized Reliability Calibration

  • PhenoAIR’s offline reliability layer precomputes reusable evidence-quality annotations over the reference set to condition interpretation of retrieved Cell Painting neighbors at inference.Its components are reference memory construction, a mechanistic family ontology, and calibrated retrieval.Source: §C.1 · p. 22

C.1.1 Reference Memory Construction

  • An offline LLM-assisted pipeline analyzes reference data at source, MOA, anchor, and MOA-pair levels, producing structured annotations of reliability and failure patterns.The LLM plans and interprets reference-set analyses; it is not used to predict query MOAs.Source: §C.1.1 · p. 22

C.1.2 Mechanistic Family Ontology

  • The mechanistic family ontology groups canonical MOAs into biologically coherent families to support family-level reasoning without replacing the benchmark’s canonical prediction labels.It also helps distinguish plausible intra-family ambiguity from arbitrary retrieval noise.Source: §C.1.2 · pp. 23–24

C.1.3 Calibrated Retrieval

  • Calibrated retrieval conservatively attenuates neighbors flagged as unreliable while preserving strong local evidence and guarding against rank demotion and redundant penalties.Source-level reliability and intra-family ambiguity inform agents but are not multiplicative factors in the deterministic similarity calibration.Source: §C.1.3 · pp. 24–25

C.2 Candidate-Centric Evidence Reasoning

  • PhenoAIR integrates phenotype and mechanism agents through shared candidate memory, where each contributes structured evidence rather than overwriting the other’s conclusions.Candidate-level states make cross-agent influence explicit.Source: §C.2 · p. 25

C.2.1 Candidate Memory Schema

  • Candidate memory tracks phenotype and mechanism evidence, reliability context, eligibility, and audit history so final decisions depend on accumulated candidate states rather than one agent or neighbor.Mechanism-supported alternatives without phenotype grounding require phenotype support before becoming eligible; reliability metadata alone cannot create or finalize candidates.Source: §C.2.1 · pp. 25–26

C.2.2 Phenotype Agent

  • The phenotype agent evaluates calibrated Cell Painting neighborhoods, records candidate support and contradiction, and flags uncertainty without using structure-side evidence.Strong phenotype contradictions trigger validation rather than automatic removal, and group-level checks stay within the requested candidate group.Source: §C.2.2 · p. 26

C.2.3 Mechanism Agent

  • The mechanism agent independently assesses candidates using structure-side evidence, but its proposals require phenotype grounding before final selection.It can propose a canonical MOA or a broader mechanistic family when a fine-grained label is not reliable.Source: §C.2.3 · pp. 26–27

C.2.4 Candidate Memory Update Rules

  • Candidate memory integrates evidence asymmetrically: phenotype initializes and grounds candidates, mechanism critiques or proposes alternatives, and reliability metadata contextualizes evidence quality.Hard suppression requires both strong phenotype contradiction and strong mechanism rejection; neither agent can unilaterally delete a plausible candidate.Source: §C.2.4 · pp. 27–28

C.3.1 Controller Action Space; C.3.2 Controller Decision Logic

  • The controller selects evidence-gathering or termination operations, prioritizing candidate validation, broader phenotype evidence, family-level checks, or reference inspection as needed.Its signal-driven rules stop unproductive cycles and do not directly edit predictions or candidate labels.Source: §C.3.1 · p. 28, §C.3.2 · p. 29

C.3.3 Arbiter Logic

  • The arbiter selects from candidates with auditable memory evidence, while a deterministic gate rejects ineligible choices and returns novel when no known MOA is adequately supported.A candidate must be in the task label space, unsuppressed, evidence-supported, and free of unresolved strong contradiction.Source: §C.3.3 · p. 29

C.3.4 Self-Evolving Procedure

  • Offline self-evolution analyzes development-set failures and proposes schema-constrained updates to controller rules or arbiter heuristics, with cross-round checks for regressions.The process does not update LLM parameters or use test outcomes; selected artifacts are fixed before final test evaluation.Source: §C.3.4 · p. 30

D.1 Benchmark Construction

The benchmark combines JUMP Cell Painting profiles with curated Drug Repurposing Hub MOA annotations and defines controlled, heterogeneous, and open-set evaluation settings. It applies label curation and leakage controls while prioritizing reference quality over dataset size.

D.1 Benchmark Construction

  • Profiles come from JUMP cpg16 across five laboratory sources, while Drug Repurposing Hub annotations are matched by JCP2022 identifiers and profiles are aggregated by source and across sources.Treatment-level profiles use per-source medians across replicate wells and plates; the global consensus uses a cross-source median.Source: §D.1 · p. 31
  • Concatenated MOA strings are parsed into atomic labels, and compounds with hierarchical parent-child label redundancy are removed to avoid inflated multi-label evaluation.For Top-1 and retrieval-quality analyses, the first-listed remaining label is designated representative; the paper states this choice is inconsequential to primary metrics after redundancy removal.Source: §D.1 · p. 31
  • MoA-Verified, MoA-Extended, and MoA-Novel respectively test reproducible cross-source evidence, broader but noisier coverage, and recognition of mechanisms absent from the reference labels.Verified and Extended retain classes with at least two reference compounds; Novel uses the Verified reference set and excludes query labels represented in its label space.Source: §D.1 · p. 31
  • The benchmarks are smaller than the full JUMP compound collection because reliable annotations, non-redundant labels, and sufficient reference anchors take priority over maximizing size.The authors identify scaling to larger MOA ontologies and additional split constructions as future work.Source: §D.1 · p. 32
  • Reference and test compounds are disjoint, and additional controls screen novel queries, refinement data, and external-tool outputs for specified forms of label leakage.Novel MOAs are checked against represented mechanistic families; self-evolving refinement uses a 30-compound development split separate from test, and tool outputs are filtered for direct ground-truth exposure.Source: §D.1 · p. 32

D.2 Evaluation Protocol

Evaluation uses Top-1 accuracy and any-match hit rate for MoA-Verified and MoA-Extended, with additional open-set metrics for MoA-Novel. An LLM judge assesses semantic label equivalence, with reported agreement against human annotations.

D.2 Evaluation Protocol

  • MoA-Verified and MoA-Extended use Top-1 Accuracy for the representative ground-truth MOA and Hit Rate for any semantically equivalent predicted and ground-truth labels.Hit Rate counts primary or secondary predictions matching any label in the ground-truth set.Source: §D.2 · p. 32
  • MoA-Novel additionally reports overall accuracy and novel-class precision, recall, and F1, treating novel detection as binary classification over in-scope and out-of-scope queries.Source: §D.2 · p. 32
  • 98% agreement with human annotations was observed for the LLM judge on a randomly sampled subset of prediction-ground-truth pairs.The judge evaluates semantic equivalence for Top-1 and Any-match predictions, including free-text MOA descriptions that differ lexically from canonical labels.Source: §D.2 · p. 32

D.3 Implementation and Reproducibility

This section describes PhenoAIR’s implementation and reproducibility, including model, feature, tooling, and inference settings; leakage controls; the separation of global reference calibration from query-time evidence refinement; and the roles and implementation details of LLMs. It also summarizes the experimental and compute information provided, notes that code was not fully released at submission, and cautions that model-generated mechanisms are not validated findings.

D.3 Implementation and Reproducibility

  • Implementation combines pretrained Cell Painting encoders, cheminformatics tools, and capped iterative inference with leakage-controlled access to external knowledge.CA-MAE and CellCLIP use ViT-Base encoders pretrained for 100 epochs on approximately 3.3 million images; refinement is capped at Tmax = 8. Query compounds are restricted to structure-derived inputs, while external annotations support reference-compound characterization and offline calibration.Source: §D.3 · pp. 32–33

E Prompt Templates

  • The Phenotype Agent ranks candidates from calibrated retrieval evidence, while the Mechanism Agent evaluates those candidates using structure-driven evidence without treating either source as standalone proof.The phenotype role assesses neighborhood reliability and ambiguity and avoids relying on the top neighbor alone; the mechanism role may support, reject, or remain uncertain about candidates and may return a candidate group when evidence does not justify a specific label.Source: E Prompt Templates · pp. 33–34
  • The Arbiter calibrates confidence across the refinement trajectory, retains alternatives when evidence is ambiguous, and makes the final canonical-label or novel-MOA decision.Confidence depends on retrieval reliability, mechanism support, and candidate stability; the Arbiter should lower confidence for oscillating rankings or unresolved conflicts and reserve novel-MOA decisions for cases without a sufficiently supported allowed canonical candidate.Source: E Prompt Templates · pp. 35–36
  • Global preparation builds source-specific profiles, a reliability memory and mechanistic-family ontology, then fixes a calibrated retrieval rule that is not recomputed per query.The reliability memory analyzes source-, MOA-, reference-anchor-, and MOA-pair reliability using offline LLM-assisted code execution.Source: E Prompt Templates · p. 37
  • At query time, PhenoAIR calibrates retrieved neighbors, creates a reliability context and empty candidate memory, then iteratively updates it with phenotype and mechanism evidence.The trajectory log records the refinement process; the supplied procedure also runs an arbiter over the final memory and trajectory before a deterministic eligibility gate.Source: E Prompt Templates · p. 37
  • Controller actions can request additional evidence, including expanded phenotype neighborhoods, candidate or group validation, anchor inspection, and mechanism-group mode.The procedure updates available evidence for the next round and applies an arbiter and deterministic final eligibility gate.Source: E Prompt Templates · p. 37

NeurIPS Paper Checklist

  • The checklist says Sections 3 and 4 disclose benchmark construction, baselines, metrics, implementation details, ablations, and experimental settings needed to reproduce and interpret the main results.The listed settings include benchmark splits, LLM configurations, and evaluation metrics.Source: NeurIPS Paper Checklist · pp. 39–40
  • Experiments use publicly available Cell Painting profiles and curated MOA annotations, but the authors did not provide a full code release at submission.The checklist says benchmark construction and the evaluation protocol are described.Source: NeurIPS Paper Checklist · p. 40
  • Stochastic LLM-based methods are averaged over multiple runs, while deterministic retrieval baselines are reported once; the evaluation protocol and variability sources are described.The checklist also says Section 4 and the Appendix report the compute environment and approximate runtime, including LLM inference settings and retrieval/evaluation workloads.Source: NeurIPS Paper Checklist · pp. 40–41
  • The paper identifies potential benefits for drug mechanism discovery and warns that model-generated mechanistic hypotheses could be over-interpreted without experimental validation.The checklist says the work uses public biological profiling data and curated annotations and does not involve human subjects or private personal data.Source: NeurIPS Paper Checklist · pp. 41–42
  • The paper declares LLMs as a core component and describes their roles in phenotype analysis, mechanism evaluation, and final arbitration, along with model configurations and prompting/evaluation protocol.Source: NeurIPS Paper Checklist · p. 44
Loading 2609.36406v1…