PaperlayerView exact revision on arXiv

Source-linked AI summary

INDIBATOR: Diverse and Fact-Grounded Individuality for Multi-Agent Debate in Molecular Discovery

Yunhui Jang, Seonghyun Park, Jaehyung Kim, Sungsoo Ahn

arXiv:2602.01815v1cs.AI

TL;DR

Generic role- and keyword-based personas simplify scientists’ distinctive research trajectories, while INDIBATOR grounds agents in publication and molecular histories. Across diverse molecular-discovery tasks, it consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance against other baselines.

  • Problem

    Generic role- or keyword-based personas simplify scientists’ distinctive research trajectories, leaving individuality-grounded expertise insufficiently represented in molecular-discovery agents.

    Source: §1 · p. 1

  • Method

    INDIBATOR profiles agents using publication histories for literature-derived knowledge and methodological preferences, and molecular histories for structural priors such as preferred scaffolds and functional groups.

    Source: §1 · p. 1, §2 · p. 2

  • Results

    Across diverse molecular discovery tasks, INDIBATOR consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance against other baselines.

    Source: §6 · p. 8

  • Takeaways & Limitations

    The results support individuality grounding as a basis for distinct, knowledge-grounded agent reasoning in molecular discovery.

    Source: §6 · p. 8

    Scientist-profile outputs are LLM-generated and do not represent the cited scientists’ actual opinions, unpublished insights, or endorsement.

    Source: Ethical Consideration · p. 9

Abstract

from arXiv · show

Multi-agent systems have emerged as a powerful paradigm for automating scientific discovery. To differentiate agent behavior in the multi-agent system, current frameworks typically assign generic role-based personas such as ''reviewer'' or ''writer'' or rely on coarse grained keyword-based personas. While functional, this approach oversimplifies how human scientists operate, whose contributions are shaped by their unique research trajectories. In response, we propose INDIBATOR, a framework for molecular discovery that grounds agents in individualized scientist profiles constructed from two modalities: publication history for literature-derived knowledge and molecular history for structural priors. These agents engage in multi-turn debate through proposal, critique, and voting phases. Our evaluation demonstrates that these fine-grained individuality-grounded agents consistently outperform systems relying on coarse-grained personas, achieving competitive or state-of-the-art performance. These results validate that capturing the ``scientific DNA'' of individual agents is essential for high-quality discovery.

1. Introduction

INDIBATOR grounds molecular-discovery agents in individual research trajectories rather than generic personas. The introduction motivates this design by arguing that scientists’ distinctive chemical preferences encode expertise that can support diverse collaboration.

1. Introduction

  • Multi-agent systems extend beyond single-agent planning to support automated scientific discovery, including molecular discovery (Yao et al., 2023; Schick et al., 2023; Du et al., 2024; Liang et al., 2024; Chan et al., 2024; Du et al., 2025; Kim et al., 2025).The introduction notes that single-agent systems face bounded context windows and limited perspective diversity.Source: §1 · p. 1
  • Role- or keyword-based separation shapes output style but often oversimplifies scientists’ individuality, which reflects fine-grained research trajectories, cumulative experiences, and domain-specific intuitions, limiting systems’ access to nuanced collaborative insights.Source: §1 · p. 1
  • 60% top-5 accuracy from molecule structure alone let models identify which of 1,815 chemists synthesized a molecule, revealing distinctive chemical styles.Blevins & Quigley (2025) frame the phenomenon as “Clever Hans” leakage; the introduction instead treats these styles as heuristics for navigating chemical space and representing expertise diversity.Source: §1 · p. 1
  • INDIBATOR grounds molecular-discovery agents in individual research trajectories, bridging generic personas and chemical reality.The passage identifies Figure 1 as illustrating the framework, but does not describe its profile sources in full.Source: §1 · p. 1

1. Scientist selection · 2. Individuality-grounded profile construction

INDIBATOR selects relevant scientists from publication authors identified through RAG, builds individualized profiles from publication and molecular histories, and uses those profiles in molecular-generation debate. The paper reports that INDIBATOR outperforms vanilla debating systems and achieves competitive or state-of-the-art benchmark performance.

1. Scientist selection

  • A supervisor agent selects relevant scientists by using RAG to identify authors of publications related to the task.This selection precedes construction of individualized scientist profiles.Source: Figure 1 · p. 2

2. Individuality-grounded profile construction

  • Each scientist’s profile combines publication history, capturing literature-derived knowledge and methodological preferences, with molecular history, capturing structural priors such as preferred scaffolds and functional groups.The two histories are drawn from the scientist’s search trajectory.Source: §2 · p. 2
  • Individualized profiles are intended to reduce redundant reasoning by giving agents distinct knowledge and inductive biases.The paper describes this as the diversity benefit of individuality-grounded design.Source: §2 · p. 2
  • Grounding agents in explicit publication and molecular records is intended to make their reasoning verifiable and evidence-based.The paper describes this as the fact-grounding benefit of data-driven profiles.Source: §2 · p. 2
  • Agents iteratively propose molecular candidates, critique proposals, and vote by assigning scores based on their specific expertise.These three phases are designed to mirror the collective intelligence of real-world scientist teams.Source: §2 · p. 2, Figure 1 · p. 2
  • INDIBATOR consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance across molecular-discovery benchmarks.The evaluation covers protein-conditioned molecule generation, bioactivity-guided molecule generation, and goal-directed lead optimization; the authors identify expertise-grounded individuality as critical for molecular discovery.Source: §2 · p. 2

2. INDIBATOR

INDIBATOR grounds scientist agents in publication and molecular histories to encourage diverse, fact-grounded reasoning. The agents iteratively propose, critique, and vote on molecular candidates.

2. INDIBATOR

  • INDIBATOR conditions each agent on a distinct real-world research trajectory to reduce redundant reasoning and ground arguments in empirical evidence from papers and molecules.The profiles draw on each scientist’s prior publications and molecular discoveries.Source: §2 · p. 2

2.1. Individuality-grounded Profile Construction

  • Profiles combine publication histories for literature-derived knowledge with molecular histories that provide structural priors.The supervisor retrieves task-relevant PubMed papers, identifies first and last authors as researchers and principal investigators, and initializes their agents with selected publications and molecules similar to a seed molecule when available.Source: §2.1 · pp. 2–3

2.2. Multi-agent Debating System

  • In each round, agents propose expertise-grounded candidates with rationales, critique their own and peers’ proposals, then score and vote on candidates.Scores reflect task relevance, synthetic feasibility, and novelty; aggregated votes rank candidates for the next round or final selection, and iteration stops at a round limit or when enough candidates accumulate.Source: §2.2 · p. 3

3. Downstream Task Evaluation

INDIBATOR is evaluated on protein-conditioned molecule generation, bioactivity-guided generation, and goal-directed lead optimization using a Deepseek-V3.2 backbone.

3.1. Protein-conditioned Molecule Generation

  • INDIBATOR beats both baselines across all protein targets and metrics, improving binding affinity and diversity while avoiding the mode collapse seen with baseline methods.KEYWORDDEBATE yields negligible improvement over VANILLADEBATE; evaluation covers eight kinase and non-kinase proteins with 1,000 candidates per protein.Source: §3.1 · pp. 3–4, Figure 2 · p. 4

3.2. Bioactivity-guided Molecule Generation

  • On PMO-1K, INDIBATOR improves performance across all three bioactivity tasks, outperforming MT-Mol, KEYWORDDEBATE, and Genetic GFN by 17.4% (DRD2) to 123.5% (JNK3).The benchmark reports average top-10 AUC over 1,000 generated molecules for GSK3β, DRD2, and JNK3.Source: Table 1 · p. 4, §3.2 · pp. 4–5

3.3. Goal-directed Lead Optimization

  • INDIBATOR generates more optimized leads than VANILLADEBATE and is competitive, but does not uniformly surpass state-of-the-art baselines.VANILLA generates no molecule satisfying all constraints; INDIBATOR uses inference-time profiles without task-specific training, unlike the trained optimization baselines.Source: Table 2 · p. 5, §3.3 · p. 5

4. Analysis

INDIBATOR’s analysis finds that detailed, factual, diverse scientist profiles improve molecular-generation performance and shape chemically informed debate. Individuality also supports scaling to more agents, while profile histories guide distinct proposals and critiques.

4.1. Qualitative Case Study

  • Agents’ distinct research histories produce chemically plausible reasoning trajectories, proposals, and targeted critiques during the proposal, critique, and voting phases.One agent transfers JNK1-inhibitor QSAR knowledge to propose an aminopyridine carboxamide scaffold; another suggests an indole core based on Aurora B inhibitors. A CNS-focused agent identifies sulfonamide as a BBB liability and suggests chloro or cyano replacements to reduce polarity.Source: §4.1 · p. 6

4.2. Effect of Granularity

  • INDIBATOR’s full publication and molecular histories significantly outperform all baselines in performance and diversity, while coarser profiles offer only marginal gains over VANILLADEBATE.The comparison spans VANILLADEBATE, generic role personas, and keyword personas; the authors attribute the benefit to fine-grained research trajectories.Source: §4.2 · p. 6, Table 3 · p. 7

4.3. Effect of Diverse Agents

  • INDIBATOR outperforms a comprehensive shared profile, showing that gains stem from diverse perspectives rather than knowledge volume alone.A homogeneous single profile often performs worse than VANILLADEBATE, whereas INDIBATOR’s diverse agents can cross-examine and propose diverse candidates.Source: §4.3 · p. 7

4.4. Effect of Fact-grounding Agents

  • INDIBATOR’s factual profiles outperform all benchmarks, while hallucinated profiles often fare worse than task-irrelevant but factual profiles.The authors conclude that real publication and molecular histories guide exploration toward biologically plausible regions.Source: §4.4 · p. 7

4.5. Ablation Study

  • Increasing scientist agents from 5 to 50 improves INDIBATOR but degrades VANILLADEBATE as debate rounds shrink, and individuality is the dominant component.INDIBATOR’s diverse agents can generate high-quality initial proposals that cover chemical space, substituting for iterative debate; integrating each component also improves performance across all three bioactivities.Source: §4.5 · pp. 7–8

5. Related Work

Prior multi-agent systems use generic or keyword-based personas, while molecular-discovery agents use tools or role specialization; INDIBATOR instead grounds individuality in publication and molecular histories.

LLM-based multi-agent systems

  • Role-play prompting creates generic personas that scale poorly to massive systems, while publication-derived keywords still lack the granular expertise needed for scientific debate.VirSci (Su et al., 2025) uses publication and collaboration-network keywords, but literature alone misses chemists’ structural priors, including scaffolds and functional groups.Source: §5 · p. 8
  • INDIBATOR profiles agents with publication and molecular histories to support diverse, fact-grounded collaboration reflecting individual scientific expertise.The profiles capture literature-derived knowledge and molecular structural priors that publication texts alone do not fully represent.Source: §5 · p. 8

LLM in molecular discovery

  • Molecular-discovery systems include molecule-generation methods LICO (Nguyen & Grover, 2025) and MOLLEO (Wang et al., 2025), plus tool-using agents ChemCrow (M. Bran et al., 2024) and MT-Mol (Kim et al., 2025).LICO and MOLLEO use structured embeddings or evolutionary search; ChemCrow combines chemistry tools with ReAct reasoning, while MT-Mol uses tool-guided reasoning and role-specialized agents.Source: §5 · p. 8
  • Existing molecular-discovery agents rely on single-agent architectures or coarse role-based personas, whereas INDIBATOR grounds each agent in publications and previously discovered molecules.These research trajectories provide the distinct inductive biases that generic role-play lacks.Source: §5 · p. 8

6. Conclusion

INDIBATOR grounds molecular-discovery agents in individual scientists’ publication and molecular histories, guiding knowledge-grounded reasoning. Across diverse tasks, it consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance against other baselines.

6. Conclusion

  • Across diverse molecular-discovery tasks, INDIBATOR consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance against other baselines.The evaluation also empirically validates the framework’s benefits of granularity, diversity, and fact-grounding.Source: §6 · p. 8
  • INDIBATOR constructs individual scientist profiles from publication and molecular histories, giving agents distinct “scientific DNA” that guides knowledge-grounded reasoning.The framework is presented as an improvement over coarse-grained generic role-playing.Source: §6 · p. 8
  • The evaluation validates granularity, diversity, and fact-grounding as benefits of individuality-based agents for high-quality scientific discovery.The authors identify researchers’ nuanced inductive biases as a critical component.Source: §6 · p. 8
  • The framework establishes a foundation for incorporating broader modalities, such as conversation records, to further enhance agent fidelity in domain-specific environments.Source: §6 · p. 8

Impact Statement

INDIBATOR aims to accelerate drug design and improve the factual reliability of AI-driven discovery through debate among agents grounded in individual research trajectories, while noting that harmful-compound design requires safety guardrails.

Impact Statement

  • The framework aims to accelerate the drug design pipeline and improve factual reliability by simulating scientific debate among agents grounded in individual research trajectories.It currently optimizes drug-likeness and synthetic accessibility.Source: Impact Statement · p. 9
  • Future open-source releases or deployments should include safety guardrails to prevent targeted design of harmful compounds.The current framework optimizes drug-likeness and synthetic accessibility.Source: Impact Statement · p. 9

Ethical Consideration · A. Prompts

This section documents INDIBATOR’s individualized debate prompts and the baseline and task prompts used in the experiments, including bioactivity-guided generation for GSK3B, DRD2, and JNK3 and seed-based lead optimization under drug-likeness and structural-similarity constraints. It also notes that profiles draw on public academic records, debate outputs are not scientists’ actual views, and outputs could be misused to create harmful molecules.

Ethical Consideration

  • The authors caution that profiles use public records, agent outputs are not scientists’ actual opinions or endorsements, and the system could be misused to automate harmful-molecule creation.They suggest future safeguards and ethical guidelines.Source: Ethical Consideration · p. 9

A.1. INDIBATOR

  • INDIBATOR prompts ground scientist agents in published expertise and molecular history, then guide molecule proposals, peer critiques, and candidate scoring.Proposals require valid SMILES and scientific rationales; critiques consider issues such as toxicity and synthesis difficulty, while voting weighs task relevance, feasibility, and novelty.Source: §A.1 · p. 13

A.2. Role persona

  • The role-persona baseline initializes agents with task-relevant expert roles and descriptions, then uses them in proposal, critique, and voting phases.Roles are generated for the task, with examples including medicinal chemists, computational biologists, and pharmacokinetics experts.Source: §A.2 · p. 15

A.3. Keyword persona

  • The keyword-persona baseline extracts 10–20 researcher-specific keywords from publication abstracts, covering methods, targets, compound classes, and disease areas.Agents are initialized with the scientist’s name and the extracted topics.Source: §A.3 · p. 16

A.4. LLM-generated profile

  • The LLM-generated profile baseline prompts for 5–10 fictional but plausible task-relevant publication titles and 5–10 relevant molecule SMILES attributed to a researcher.The prompts ask for diverse but coherent publications and valid, drug-like molecules.Source: §A.4 · p. 17

A.5. Vanilla

  • The vanilla baseline asks a generic molecular-design researcher to generate a batch of unique molecules as valid, nonduplicated SMILES.Its output is specified as a JSON array.Source: §A.5 · p. 17

A.6. Task prompts

  • The protein-target task prompt asks for valid SMILES maximizing predicted binding affinity, measured as log10(IC50) in µM, where lower values indicate stronger binding.It recommends hydrogen-bonding and hydrophobic groups, 300–600 Da molecular weight, aromatic rings for pi-stacking, and possible salt bridges.Source: §A.6 · p. 18
  • The task prompts require outputting only the SMILES string, without explanation.Source: §A.6 · pp. 18, 20
  • GSK3B prompts seek drug-like molecules with high predicted inhibitory activity, favorable ADMET properties, and no duplicates of examples or prior outputs.They also ask agents to consider structural features that enhance kinase binding affinity.Source: §A.6 · p. 18
  • DRD2 prompts maximize receptor-binding probability and score while requiring drug-like molecules and avoiding example-matching or repeated structures.The target is the dopamine D2 receptor.Source: §A.6 · pp. 18–19
  • JNK3 prompts seek drug-like molecules with high predicted inhibitory activity and favorable ADMET properties, while avoiding duplicate examples and prior outputs.They ask agents to consider structural features that enhance kinase inhibition.Source: §A.6 · p. 19
  • Lead optimization starts from a provided seed and requires improved predicted docking affinity while satisfying Tanimoto similarity ≥{sim threshold}, QED ≥0.6, SA Score ≤4, and nonidentity to the seed.Tanimoto guidance says to preserve the core scaffold and make small modifications; QED guidance lists molecular weight 200–500 Da, logP 0–5, H-bond donors ≤5, acceptors ≤10, rotatable bonds ≤10, and 1–4 aromatic rings. SA guidance recommends common building blocks and simple ring systems while avoiding large rings, bridgehead/spiro atoms, and multiple stereocenters.Source: §A.6 · pp. 19–20
  • Recommended edits include small substituent changes and bioisosteric replacements; prompts warn against large polycycles, more than two new rings, unusual groups, and removing the core scaffold.Examples include benzene↔pyridine and -COOH↔-CONH2 replacements.Source: §A.6 · p. 20

B. Experimental settings … C.3. Detailed protein-conditioned molecule generation results

Experiments found that increasing scientist count improved discovery performance, while the accompanying sections specify debate settings, computational resources, worked debate examples, and protein-conditioned generation evaluations.

B.1. Hyperparameters

  • The debate used 50 scientist agents, 30 candidate molecules per agent per iteration, up to 20 rounds, and a sampling temperature of 0.7.Source: §B.1 · p. 20

B.2. Computational resource

  • Boltz-2 protein-target molecule-generation predictions used a single NVIDIA RTX A5000 GPU.The passage attributes Boltz-2 to Passaro et al. (2025).Source: §B.2 · p. 20

C.1. Detailed examples

  • The detailed examples include publication-history records and molecule-history structures, followed by proposal, critique, and voting outputs.The proposal rationales describe structural design choices, critiques raise concerns about metabolism and hinge-binding motifs, and voting examples assign candidate scores.Source: §C.1 · pp. 21–22
  • One proposal rationale adapts aminopyridine-carboxamide JNK inhibitor features for JNK3, while an alternative replaces a sulfonamide with an acetamide.The rationales discuss a cyano group and cycloalkylamine substituents; the alternative retains the cyanoaminopyridine core and cyclopropylamine.Source: §C.1 · p. 22
  • Critiques flag sulfonamide metabolism and the hinge-binding motif, and recommend exploring alternative scaffolds or heterocyclic hinge binders.Source: §C.1 · p. 22

C.2. Number of scientists and proposals

  • Increasing the number of debating scientists consistently enhanced discovery performance across both experimental settings, while exploration depth partly compensated for low scientist diversity under a fixed 1,500-proposal budget.At low N, the system maintained relatively high AUC scores when k increased.Source: §C.2 · p. 23

C.3. Detailed protein-conditioned molecule generation results

  • Protein-conditioned molecule-generation results are presented for binding affinity in Table 5 and diversity in Table 6.The supplied passages give the tables’ scopes but no cell values or comparative outcomes.Source: Table 5 · p. 23, Table 6 · p. 24, §C.3 · p. 24

C.4. Additional PMO tasks

On additional PMO tasks, INDIBATOR improves consistently over vanilla debate in most cases, though the reported metrics are less indicative of its intended utility because the tasks do not require profile-grounded reasoning or broad chemical-space exploration.

C.4. Additional PMO tasks

  • INDIBATOR improves over vanilla debate in most PMO-1K tasks assessed by top-10 AUC.The paper reports results for multi-property optimization and molecule rediscovery in Table 7, with gains in most cases.Source: Table 7 · p. 24, §C.4 · p. 24
  • The additional PMO benchmark covers multi-property optimization and molecule rediscovery, which the authors excluded from the main text as arithmetic structural puzzles.The benchmark is attributed to Gao et al. (2022).Source: §C.4 · p. 24
  • These metrics are less indicative of INDIBATOR’s intended utility because the tasks require neither individual profile-grounded reasoning nor broad chemical-space exploration.The framework is designed to facilitate both capabilities.Source: §C.4 · p. 24

Related papers

Loading 2602.01815v1…