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
TL;DR
Existing multi-agent systems often rely on coarse personas that do not capture scientists’ individual research trajectories. INDIBATOR grounds agents in publication and molecular histories, then has them debate through proposal, critique, and voting; across molecular discovery tasks, it consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance.
Problem
Existing agent frameworks use coarse role-based or keyword-based personas that oversimplify scientists’ research trajectories and distinctive scientific DNA.
Method
INDIBATOR constructs individualized scientist profiles from publication histories and molecular histories, using them in proposal, critique, and voting phases.
Results
Across diverse molecular discovery tasks, INDIBATOR consistently outperforms vanilla debating systems and achieves competitive or state-of-the-art performance compared with other baselines.
Takeaways & Limitations
Capturing fine-grained, fact-grounded scientific individuality is supported as a key component of molecular discovery in the evaluated tasks.
Takeaways & Limitations
The study uses public academic records, and agent outputs do not represent the actual opinions, unpublished insights, or endorsements of the cited scientists.
Abstract
from arXiv · showhide
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
Multi-agent systems address limitations of single-agent perspective by enabling collaborative automated discovery, but generic personas oversimplify scientists’ distinctive research trajectories. INDIBATOR instead grounds molecular-discovery agents in individualized scientific DNA derived from those trajectories.
- Motivation: Multi-agent systems support automated discovery by providing collaborative intelligence and multiple perspectives beyond single-agent systems.Single-agent systems can face bounded context windows and limited perspective diversity.
- Limitations of Existing Personas: Prior systems differentiate agents with generic role-play personas such as planners, verifiers, and reviewers or with keywords.These approaches primarily shape conversational output style.
- Limitations of Existing Personas: Generic roles and keywords oversimplify scientists’ individuality, which reflects cumulative experiences and domain-specific intuitions.The paper calls this distinctive research trajectory a scientist’s “scientific DNA.”
- Chemical Scientific DNA: Chemists exhibit distinctive preferences for scaffolds, functional groups, and reaction motifs based on their research trajectories.A study identified which of 1,815 chemists synthesized a molecule with 60% top-5 accuracy from structure alone.
- INDIBATOR: INDIBATOR grounds molecular-discovery agents in individual research trajectories rather than predefined roles or keywords.Its agent profiles use informative sources that encode researchers’ expertise.
1. Scientist selection
The task specifies designing a molecule to optimize TYK2 binding affinity while satisfying additional constraints.
- Task Objective: The target objective is to optimize a molecule’s binding affinity to TYK2.The task also includes a seed molecule and additional constraints.
2. Individuality-grounded profile construction
INDIBATOR constructs individualized scientist agents from publication and molecular histories, then uses them in a debate process for molecular candidate generation. The design aims to provide diverse, fact-grounded reasoning.
- Profile Construction: Scientist profiles combine publication history for literature-derived knowledge with molecular history for structural priors.Molecular history includes previously developed molecules, while publication history captures research knowledge and preferences.
- Profile Construction: INDIBATOR uses profiles to give agents distinct individuality that mirrors researchers’ cumulative knowledge and inductive biases.
- Debate Process: The framework’s debate has proposal, critique, and voting phases for generating molecular candidates.Agents propose candidates, critique proposals, and score them according to their expertise.
- Evaluation: The framework is evaluated on protein-conditioned generation, bioactivity-guided generation, and goal-directed lead optimization.These tasks cover multiple molecular discovery settings.
- Evaluation: INDIBATOR’s results show consistent advantages over vanilla debating systems and competitive or state-of-the-art performance across benchmarks.The paper presents scientific DNA as a fundamental driver of molecular design within the evaluated scope.
2. INDIBATOR
INDIBATOR creates scientist agents from real researchers’ publication and molecular histories, then runs iterative proposal, critique, and voting. These profiles are intended to increase diversity and fact-grounding in molecular discovery.
- INDIBATOR: INDIBATOR conditions each agent on a distinct real-world profile derived from prior publications and historical molecular discoveries.
- Diversity: Each agent’s uniquely constructed system prompt is intended to prevent redundant reasoning across the ensemble.
- Fact-grounding: Agent reasoning is grounded in papers and discovered molecules from real-world scientist profiles.
- Profile Construction: Expertise profiles contain publication histories retrieved from PubMed and molecular histories represented as SMILES strings.Publication histories encode research focus and methodological preferences, while molecular histories establish structural priors.
- Scientist Selection: The supervisor uses RAG over PubMed literature to identify relevant papers and extracts first and last authors as scientist agents.The resulting scientist set contains N agents, where N is a hyperparameter.
- Profile Construction: Each scientist agent receives an expertise profile Ei = {Pi, Mi}, with publications selected by task-relevant keywords and molecules retrieved by structural similarity when a seed is available.
- Debate Process: The debate iterates through proposal, critique, and voting until a round limit or sufficient candidate count is reached.
- Proposal: During proposal, profile-conditioned agents generate molecular candidates with rationales linked to their prior knowledge.
3. Downstream Task Evaluation
INDIBATOR is evaluated across protein-conditioned generation, bioactivity optimization, and goal-directed lead optimization against debate, LLM, and structure-based baselines. Across these settings, individuality-grounded agents improve molecular generation, while remaining competitive in constrained lead optimization.
- 3. Downstream Task Evaluation: The evaluation covers protein-conditioned generation, bioactivity-guided generation, and goal-directed lead optimization using the Deepseek-V3.2 backbone.These tasks use distinct molecular objectives, including binding affinity, bioactivity, and constrained optimization from seed molecules.
- 3.1. Protein-conditioned Molecule Generation: INDIBATOR consistently beats vanilla and keyword debate across all protein targets and metrics, improving both binding affinity and molecular diversity.The baselines frequently suffer from mode collapse, whereas INDIBATOR generates many structurally distinct clusters without compromising optimization quality.
- 3.1. Protein-conditioned Molecule Generation: Coarse-grained keyword profiles provide negligible improvement over vanilla debate, whereas fine-grained profiles adapt search strategies to distinct biological interactions.The comparison spans diverse protein families.
- 3.2. Bioactivity-guided Molecule Generation: 17.4% (DRD2) to 123.5% (JNK3) are INDIBATOR’s performance margins over Genetic GFN across the evaluated targets.The method also outperforms other optimization baselines by significant margins.
- 3.3. Goal-directed Lead Optimization: INDIBATOR generates more optimized leads than VANILLADEBATE in constrained goal-directed optimization, although it does not uniformly surpass trained state-of-the-art baselines.The task maximizes docking score while requiring QED ≥0.6, SA ≤4, and sim ≥0.6 across five target proteins with three seed molecules each.
4. Analysis
The analysis examines how INDIBATOR’s individuality, diversity, fact-grounding, collaborator count, and debate components affect molecular discovery. Across qualitative and ablation analyses, detailed real-world profiles and diverse collaboration are associated with stronger performance and more efficient scaling.
- Qualitative Case Study: Individual publication and molecular histories produce distinct, chemically plausible reasoning trajectories during JNK3-guided molecule generation.Agents propose and critique candidates using prior publications, scaffolds, and therapeutic-indication knowledge.
- Effect of Granularity: Increasing profile granularity consistently improves performance, with INDIBATOR significantly outperforming role- and keyword-persona baselines in performance and diversity.INDIBATOR uses full publication and molecular histories rather than generic roles or extracted keywords.
- Effect of Diverse Agents: INDIBATOR outperforms massive single-profile agents, indicating that diverse perspectives contribute beyond simply aggregating more knowledge.Single-profile agents perform worse than VANILLADEBATE in most metrics, while heterogeneous agents can cross-examine proposals and generate diverse candidates.
- Effect of Fact-Grounding: Fact-grounded profiles outperform LLM-generated and random-profile baselines across benchmarks, while hallucinated expertise can introduce noise into reasoning.The random-profile baseline contains factual but task-irrelevant histories; LLM-generated profiles may contain hallucinations.
- Ablation Study: Increasing the number of scientists improves INDIBATOR performance, whereas VANILLADEBATE degrades as fewer debate rounds remain.As N increases, both systems conclude in one round, but INDIBATOR’s diverse initial proposals cover chemical space more effectively.
- Ablation Study: Integrating proposal, critique, voting, and individuality yields progressive performance gains, with individuality providing the largest increment.The ablation evaluates all three bioactivities and identifies individuality as the dominant factor in INDIBATOR’s success.
5. Related Work
Related molecular-discovery systems have expanded from single-agent frameworks toward multi-agent collaboration and finer-grained personas. INDIBATOR addresses their reliance on coarse roles or ungrounded profiles by using individual research trajectories to support distinct expertise.
- LLM-based Multi-agent Systems: Multi-agent systems extend LLM-based discovery by using collaborative intelligence and distinct personas to simulate complex scientific interactions.Prior applications span scientific discovery and molecular discovery.
- Limitations of Existing Approaches: Keyword-based personas provide more detail than generic roles but lack the granular expertise required for real-world scientific debate.In molecular discovery, literature-derived knowledge alone is also insufficient without the context of prior molecules and research trajectories.
- INDIBATOR: INDIBATOR grounds each agent in actual publications and historical molecular discoveries, replacing generic role-play with collaboration driven by distinct scientific DNA.The framework targets the limitation of single-agent or coarse-grained systems that lack grounded individual expertise.
6. Conclusion
INDIBATOR grounds scientist agents in individual research trajectories through publication and molecular histories, enabling multi-turn, knowledge-grounded debate. Across diverse molecular discovery tasks, this individuality-based approach consistently outperforms vanilla debating systems and reaches competitive or state-of-the-art performance.
- INDIBATOR constructs each agent’s profile from publication and molecular history, creating a distinct “scientific DNA” for knowledge-grounded reasoning.Publication history supplies literature-derived knowledge, while molecular history contributes structural priors.
- INDIBATOR consistently outperforms vanilla debating systems across diverse molecular discovery tasks.
- INDIBATOR achieves competitive or state-of-the-art performance compared with other baselines.
- Evaluations validate granularity, diversity, and fact-grounding as benefits of individuality-based agents for high-quality scientific discovery.The framework captures nuanced inductive biases associated with individual researchers.
- The framework provides a foundation for incorporating broader modalities, including conversation records, into domain-specific agents.
Impact Statement
INDIBATOR uses multi-agent debate grounded in individual research trajectories for molecular discovery. The framework aims to accelerate drug design and improve the factual reliability of AI-driven scientific discovery, while recognizing the need for safety guardrails.
- INDIBATOR simulates scientific debate with agents grounded in individual research trajectories for molecular discovery.
- The framework aims to accelerate the drug design pipeline and improve the factual reliability of AI-driven scientific discovery.
- INDIBATOR currently optimizes drug-likeness and synthetic accessibility.
- Future open-source releases or deployments should include safety guardrails to prevent targeted design of harmful compounds.
Ethical Consideration
The paper acknowledges ethical considerations in constructing scientist profiles from public academic records and in generating agent debates. It states that these outputs do not represent the cited scientists’ actual opinions or private reasoning, and calls for safeguards against misuse.
- Scientist expertise profiles are constructed from publicly available academic records, including PubMed titles, abstracts, and molecular discoveries.
- Agent proposals, critiques, and votes are probabilistic LLM outputs rather than the actual opinions, unpublished insights, or endorsements of cited scientists.
- The use of specific scientist profiles is limited to validating individuality grounding in molecular discovery.
- The system could be misused to automate creation of toxic or harmful molecules.
- The paper recommends future safeguards and ethical guidelines to mitigate misuse risks.
B. Experimental settings
The experiments configure a multi-agent molecular-discovery debate with fixed sampling and proposal settings, then examine how scientist diversity and exploration depth affect performance. Results indicate that more scientists improve discovery performance, while deeper exploration can partly compensate for fewer scientists.
- The system uses N = 50 scientist agents, generates k = 30 candidate molecules per iteration, and permits up to 20 rounds at temperature 0.7.
- N × k = 1,500 fixes the total proposal budget while varying scientist diversity against exploration depth.
- Increasing the number of scientists consistently enhances discovery performance, while larger proposal depth can partially compensate for limited expertise diversity.
C.4. Additional PMO tasks
Additional PMO-1K results show that INDIBATOR generally improves performance over vanilla debate on multi-property optimization and molecule rediscovery tasks. These tasks are assessed with top-10 AUC but are less indicative because they are arithmetic structural puzzles.
- INDIBATOR achieves consistent performance improvements in most additional PMO cases compared with vanilla debate.
- The additional PMO tasks were excluded from the main text because they are arithmetic structural puzzles.