Accelerating scientific discovery with Co-Scientist

Co-Scientist overview

• Co-Scientist combines specialized Gemini agents, asynchronous resource allocation, tournaments, scientific debate, evolution, self-critique, and external tools to generate and refine hypotheses.Its natural-language interface lets scientists specify goals and constraints, suggest ideas, and provide feedback throughout exploration.Source: Introduction · pp. 2–3

Introduction

• The system was validated across drug repurposing for cancer, novel treatment-target discovery for liver fibrosis, and mechanistic explanation of antimicrobial resistance.Reported applications included AML candidates, anti-fibrotic epigenetic targets in human hepatic organoids, and an independently recapitulated cf-PICI host-range mechanism.Source: Introduction · pp. 3, 5, 6

System analysis and evaluation

• Ablations showed that external search for Reflection reduced implausible hallucinations, scientific debate improved Ranking and reduced positional bias, and Evolution substantially improved hypothesis quality.These findings support the contribution of specialized agent components and prompting strategies to robust scientific reasoning.Source: System analysis and evaluation · p. 7
• Co-Scientist eventually significantly surpassed frontier language and reasoning models in Elo rating on 15 expert-curated biomedical goals as test-time compute increased.Across 203 research goals, both maximum and top-10 average Elo increased over sequential computation, with no observed saturation; these auto-evaluated trends do not establish objective ground truth.Source: System analysis and evaluation · p. 8, Figure 2 · p. 9
• Co-Scientist outputs were most preferred across 11 expert-evaluated research goals, with average preference rank 2.36 and novelty and impact ratings of 3.64 and 3.09 out of 5.These evaluations were subjective expert assessments, not objective ground truth.Source: System analysis and evaluation · p. 11

Real-world validations

• In AML repurposing experiments, Binimetinib, Pacritinib, and Cerivastatin inhibited cell viability, while Binimetinib reached an IC50 as low as 2 nM in most AML cell lines.Binimetinib showed much higher IC50 in the non-AML TK6 line, and the experiments were intended as an initial viability check rather than replacement for preclinical or clinical validation.Source: Real-world validations · pp. 13–14
• An 18-fold IC50 separation between KIRA6-sensitive KG-1a cells and TK6 controls indicated a potential selective therapeutic window, while other AML subtypes were less sensitive.KIRA6 inhibited viability across KG-1a, MOLM-13, HL-60, and NOMO-1 AML lines; Nanvuranlat and Leflunomide had limited effects on MOLM-13.Source: Real-world validations · pp. 15–16
• Clinical translation remains uncertain because limited in vitro models cannot capture patient heterogeneity, pharmacokinetics, off-target effects, tumor microenvironments, or systemic resistance mechanisms.Even expert review, preclinical rationale, and strong in vitro experiments do not guarantee in vivo efficacy or clinical success.Source: Real-world validations · p. 17
• Structured translational analysis identified frail, heavily pretreated AML patients as a clinical niche for Binimetinib, whose UGT1A1 pathway may avoid azole-associated CYP3A4 interactions.The framework incorporated demographics, ELN2022 risk, molecular features, preclinical activity, and safety and PK/PD data.Source: Real-world validations · p. 17
• Co-Scientist-proposed AML combinations showed strong synergy in MOLM-13, whereas KG-1a responses varied between synergy and antagonism across seven tested regimens.JNJ-64619178 plus Selinexor and JQ1 plus Olaparib plus MSA2 illustrate the validated dual- and triple-combination responses.Source: Real-world validations · p. 17, Figure 4 · p. 18
• In hepatic organoids, two of three AI-identified epigenetic modifiers showed significant anti-fibrotic activity without cellular toxicity, including FDA-approved Vorinostat as a repurposing candidate.Experts selected three top-ranked target hypotheses and corresponding drugs for testing in severe liver fibrosis.Source: Real-world validations · p. 19
• With minimal background information, Co-Scientist independently proposed that cf-PICIs interact with diverse phage tails to expand host range, matching a contemporaneous genomic discovery.The hypothesis was generated in two days and recapitulated the primary finding of an independent study before peer review.Source: Real-world validations · pp. 19, 23

Discussion

• The system iteratively refines hypotheses through generation, self-debate, tournament selection, and evolution, using test-time compute and persistent context to progressively improve scientific reasoning.Its context memory supports synthesis of information, identification of knowledge gaps, and a more coherent internal representation of the research process.Source: Discussion · p. 24, Methods · p. 32
• Co-Scientist generated experimentally tractable hypotheses across oncology, liver fibrosis, and microbiology, including AML drug candidates with in vitro efficacy and a recapitulated unpublished bacterial gene-transfer mechanism.For liver fibrosis, it proposed epigenetic targets leading to validation of several anti-fibrotic compounds, including one FDA-approved drug.Source: Discussion · p. 24
• The framework is model-agnostic and portable across frontier language models, allowing improvements in underlying models to enhance hypothesis generation and long-horizon scientific tasks without retraining the agentic framework.The study used Gemini 2.0 models, while the framework is described as portable to other similar models or combinations.Source: Discussion · p. 24, Methods · p. 32
• Co-Scientist remains limited by incomplete literature access, missing negative results, contradictory source quality, model hallucinations, imperfect factuality, and preliminary validation.Future provenance capabilities should trace claims to specific figures or data to reduce propagation of unreliable findings.Source: Discussion · pp. 24–25

Conclusion

• Responsible deployment requires rigorous verification, peer review, safeguards against bias and homogenized research directions, and human scientific reasoning to prevent low-quality artifacts from worsening reproducibility.Future work includes stronger fact-checking, broader literature and database access, multimodal reasoning, human and experimental feedback, expanded evaluations, and laboratory automation.Source: Discussion · p. 25, Conclusion · p. 25

Methods

• Co-Scientist’s architecture combines natural-language scientist interaction, asynchronous task execution, specialized agents, and persistent context memory around configurable criteria for alignment, plausibility, novelty, testability, and safety.Scientists can define goals, constraints, preferences, and feedback while the Supervisor allocates resources across worker agents.Source: Methods · pp. 29–30
• Specialized Generation, Reflection, Ranking, Proximity, Evolution, and Meta-review agents collectively explore literature, critique hypotheses, conduct Elo-based tournaments, refine proposals, and synthesize research overviews.The agents use complementary strategies including deep verification, observation and simulation reviews, similarity graphs, hypothesis combination, simplification, and divergent exploration.Source: Methods · pp. 31, 33, 34, 35, 36
• The Meta-review agent enables feedback propagation without back-propagation by appending synthesized review insights to later prompts, allowing iterative improvement as compute scales.It identifies recurring issues in reviews and debates, improving future reviews and hypothesis generation while avoiding over-fitting to individual critiques.Source: Methods · pp. 32, 35, 36
• Scientists steer Co-Scientist by defining and refining goals, reviewing or contributing hypotheses, and selecting ranked proposals; AML setup took under one hour and final selection about three hours.The system also uses web search, open databases, private repositories, and specialized models such as AlphaFold; computational evaluations used n = 203, n = 15, and n = 11 goals, while five AML cell lines were tested in independent biological triplicates, with successful replication and blinded human evaluation.Source: Methods · pp. 37–39
• The multi-faceted Generation agent’s diverse strategies contributed to correct hypotheses across different benchmarks, while the Proximity agent’s similarity scores correlated strongly with actual quality-score differences.Together, these findings support broad search and semantic comparison across the problem space.Source: Methods · p. 38
• Search integration reduced novelty errors from 6.14 to 2.38 and increased correctness from 7.4 to 8.46, while GPQA AUC rose from 0.643 to 0.651.These results used Gemini 2.0 Flash as the agents’ base foundational LLM and evaluated published non-novel ideas and known-correct ideas.Source: Methods · p. 38
• Scientific-debate prompting improved Ranking-agent accuracy for high-quality hypotheses and reduced positional bias.This comparison was made against a simple comparison prompt.Source: Methods · p. 38
• Evolution’s iterative refinement increased GPQA precision from 70.9% to 75.4% and research-hypothesis quality from 4.7 to 5.6.The quality scores were auto-evaluations on a constructed dataset.Source: Methods · p. 38
• Meta-review improved correctness-review AUC from 0.521 to 0.597 on the constructed dataset and from 0.629 to 0.634 on GPQA diamond.The ablation evidence supports the contribution of specialized agents and prompting strategies to reliable hypothesis evaluation.Source: Methods · p. 38

28. Jones, N. OpenAI’s ‘deep research’ tool: is it useful for scientists? Nature

• Reference 29 cites Megat et al.’s integrative genetic analysis of ALS heritability and risk genes.The citation spans Nat. Commun. 14, 342 (2023).Source: §28 · p. 39
• Reference 30 cites Elo’s book on rating chess players, published by Ishi Press in 2008.Source: §28 · p. 39
• Reference 31 cites LeCun’s theoretical framework for back-propagation from the 1988 Connectionist Models Summer School proceedings.The cited chapter appears on pages 21–28.Source: §28 · p. 39
• Reference 32 cites work on the impact of intransitivity on the Elo rating system by Hamilton, Kalenkova, and Roughan.The citation is to PLoS One 20, e0338261 (2025).Source: §28 · p. 39
• References 33 and 34 cite extensions and applications of Elo ratings to Go move patterns and margin of victory.They cite Coulom’s ICGA Journal work and Kovalchik’s International Journal of Forecasting work.Source: §28 · p. 39
• Reference 35 cites Khan et al.’s study of whether debating with more persuasive LLMs leads to more truthful answers.The work appears in the Proceedings of the 41st International Conference on Machine Learning, volume 235, pages 23662–23733 (2024).Source: §28 · p. 39
• Reference 36 cites fast lexically constrained decoding with dynamic beam allocation for neural machine translation.The work is from the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1314–1324.Source: §28 · pp. 39–40
• Reference 37 cites the Gemini Team’s preprint describing Gemini as a family of highly capable multimodal models.Source: §28 · p. 40

Code availability

• The full Co-Scientist source code is not publicly available because of proprietary infrastructure, computational demands, and safety implications.The authors therefore cannot provide broad access immediately.Source: Code availability · p. 40
• Scientists can request experimental access subject to computational resources, with broader access via Google APIs and bespoke interfaces planned over time.Pseudocode for multi-agent orchestration and tournament logic and the exact system prompts are provided in Supplementary Notes 8 and 9.Source: Code availability · p. 40
• Gemini, the foundational model used by Co-Scientist, is publicly available through APIs, Google AI Studio, the Gemini App, and other surfaces.Additional Gemini technical details are described in technical reports 8,14,37.Source: Code availability · p. 40
• The system and analyses used Python 3.11.7 with pandas 2.1.4, numpy 1.26.4, seaborn 0.12.2, and matplotlib 3.8.0, alongside GraphPad Prism 10.6.0 and Julius AI.GraphPad Prism supported dose-response and IC50 estimation, while Julius AI analyzed single-agent and combination-treatment dose-effect data.Source: Code availability · p. 41

Competing interest declaration

• The study was funded by Alphabet Inc. and/or a subsidiary, and many authors were Alphabet employees who may own company stock.EDV and BL are Sequome employees, and BL founded Dendra Therapeutics.Source: Competing interest declaration · p. 42

Supplementary Information

• Supplementary materials cover terminology, Co-Scientist evaluation, ablations, expert assessment, drug-repurposing analyses, and wet-lab validation.They also include AML rationale, clinical-variable guidance, and detailed information on validated repurposing candidates.Source: Supplementary Information · p. 43
• Additional supplementary sections provide safety and ethical analysis, agent pseudocode, specialized prompts, examples of inputs and outputs, AlphaFold tool use, and related works.The supplement also includes a detailed Co-Scientist output for a validated AML repurposing candidate.Source: Supplementary Information · p. 43

1 Glossary of terminology and concepts

• A novel repurposing candidate is an existing, safety-profiled drug proposed for a disease or condition outside its current approval or common use.Drug repurposing differs from traditional discovery of new compounds designed to bind disease-relevant targets.Source: §1 · p. 44
• A novel target is a gene, protein, pathway, or other biological entity not previously recognized as a therapeutic target for a specific disease.Source: §1 · p. 44
• A novel mechanistic explanation is a newly proposed pathway, interaction, or biological process explaining phenomena such as disease progression or antimicrobial resistance beyond prior literature.Source: §1 · p. 44

2.1 The Elo rating is concordant with high quality Co-Scientist results

• 78.4% top-1 accuracy was achieved on GPQA diamond by selecting each question’s highest-Elo Co-Scientist result, with Elo concordant with accuracy.Accuracy was averaged across Elo-rating buckets, with Gemini 2.0 responses providing a question-matched reference baseline.Source: §2.1 · p. 45, Figure 1 · p. 46

2.2 Co-Scientist augmented expertise through iterative refinement

• Human-seeded hypotheses progressively improved with scaled test-time compute and eventually surpassed autonomously generated solutions after tournament-based critique and evolutionary refinement.The experiment used 15 highly complex biomedical research goals with experts’ initial hypothetical solutions.Source: §2.2 · p. 46

2.3 LLM-as-a-judge preference auto-evaluation

• Co-Scientist outputs were consistently preferred by all four LLM evaluators, and the rankings closely matched human clinical experts’ subjective assessments.The evaluation compared Co-Scientist with Gemini 2.0 Flash Thinking Experimental, Gemini 2.0 Pro Experimental, and OpenAI o1 across 15 expert-curated research goals.Source: §2.3 · p. 47

2.4 Safety evaluation of Co-Scientist using adversarial research goals

• Co-Scientist passed all preliminary adversarial safety checks across 1,200 dangerous and over 2,000 safe research goals spanning 40 biomedical and scientific topics.The checks assessed rejecting dangerous or unethical goals while preserving safe research, using a dataset available upon request but not publicly released.Source: §2.4 · p. 47

3 Quantitative ablation and sensitivity analysis

• Ablations across Co-Scientist’s agents and external tools demonstrate measurable benefits for the overall multi-agent architecture on GPQA Diamond and the constructed paper QA benchmark.The paper QA benchmark contains 200 recent bioRxiv examples filtered to prevent access to ground-truth answers; summarized findings appear in Supplementary Table 1.Source: §3 · p. 48, Table 1 · p. 48

Ranking agent bias

• Scientific-debate prompting significantly improves high-quality idea comparisons and reduces positional bias, producing more stable and reliable pairwise rankings.The debate prompt yielded better performance for high-Elo ideas, while score differences became more symmetric and the second-player advantage was nearly eliminated.Source: Ranking agent bias · pp. 49, 53, Figure 4 · p. 53, Figure 5 · p. 54
• 75.4% precision versus 70.9% for baseline hypotheses on GPQA, while paper-QA quality rose from 4.7 to 5.6 through iterative refinement.On GPQA, the absolute precision increase was 4.5% [95% CI: -2%, 11%]; on paper QA, the average maximum score improvement was 0.9 points [95% CI: 0.05, 1.7].Source: Ranking agent bias · pp. 49, 54, 55
• Meta-review improved correctness-prediction AUC from 0.629 to 0.634 on GPQA and from 0.521 to 0.597 on paper QA.On paper QA, correlation with idea quality scores also increased from 0.101 to 0.226.Source: Ranking agent bias · pp. 49, 55
• External search grounded reflection by lowering novelty scores from 6.14 to 2.38 and raising correctness confidence from 7.4 to 8.46 on the paper QA benchmark.With Flash 2.0 on GPQA, search also increased correctness-classification AUC from 0.643 to 0.651.Source: Ranking agent bias · p. 50, Figure 2 · p. 51
• Generation strategies made distinct contributions without one dominating, while Elo ratings, ranks, and proximity scores supported accurate and diverse hypothesis selection.Examples include 13.8% from “using focus areas” and 13.0% from “using generate prompt” on GPQA; high proximity scores of 4-5 corresponded to small quality-score differences.Source: Ranking agent bias · pp. 49–50, Figure 3 · p. 52

Clinical expert evaluation of drug repurposing proposals in NIH Specific Aims format

• Clinicians pre-screened proposals for mechanistic incompatibility, unfavorable pharmacokinetics, toxicity, confounding observational benefits, and insufficient preclinical evidence before formal evaluation.Pamidronate was excluded because apparent benefits in bone-metastasis studies primarily reflected fewer skeletal-related events rather than disease-modifying antitumor activity.Source: §4 · pp. 57–58
• Nine board-certified hematologists and oncologists from one institution evaluated 78 unique repurposing hypotheses presented in NIH Specific Aims format.The panel included six domain-specific oncologists and three general hematologists and oncologists, with an average of 6.7 years of clinical experience.Source: §4 · p. 57

Expert rating

• The modified NIH rubric assessed 15 axes spanning research importance, innovation, rigor, feasibility, clinical relevance, and potential for clinical translation.Raters used a five-point agreement scale and were instructed to assess clinical relevance and translation potential rather than translational capacity or clinical-trial design.Source: §4 · pp. 57–58
• Spearman’s rho was 0.745 (p < 0.001), and both expert groups consistently rated Co-Scientist’s Specific Aims “Strongly Agree” or “Agree” across evaluation criteria.Supplementary Fig. 7 reports high-quality ratings across all rubric axes for the same 78 proposals reviewed independently by two expert groups.Source: §4 · p. 58, Figure 7 · p. 59
• Interpretation is constrained by single-center expert sampling, clinician pre-screening, and the absence of randomized phase III trials establishing efficacy or regulatory approval for the proposed new indications.The examples were generated without restricting the literature subset or filtering for existing preclinical evidence or prior multidrug-regimen evaluation.Source: §4 · p. 58, Expert rating · p. 69

Proposed Solution

• For AML, the givosiran proposal targeted ALAS1-dependent heme biosynthesis in MYCN-driven disease to reduce leukemic growth and survival.The Specific Aims also proposed testing oxidative stress, chemotherapy sensitivity, hepatic safety, and in vivo tumor-growth inhibition, with rationale grounded in heme-biosynthesis studies.Source: Proposed Solution · pp. 60–61, Specific Aims 1 · p. 61, Specific Aims 2 · p. 61
• For COAD, the Selinexor proposal aimed to inhibit XPO1, restore nuclear tumor-suppressor activity, suppress NF-κB signaling, and enhance apoptosis.Its aims included in vitro efficacy assessment and a patient-derived xenograft pilot evaluating tumor growth and biomarker changes.Source: Proposed Solution · p. 63, Specific Aims 1 · p. 64, Pilot Evaluation · p. 65
• A second COAD proposal positioned lapatinib for molecularly selected disease by inhibiting EGFR/HER2 signaling, with combination strategies intended to address KRAS-mediated resistance.The proposed studies examined proliferation, apoptosis, predictive biomarkers, MEK-inhibitor combinations, and a pilot in HER2-positive, KRAS-wild-type metastatic COAD.Source: Proposed Solution · pp. 66–67, Specific Aims 1 · p. 67, Specific Aims 2 · p. 67

4.2 Specific Aims count distribution

• Supplementary Table 2 reports the cancer-type count distribution among Co-Scientist’s Specific Aims drug-repurposing proposals.The supplied passage identifies the table’s scope but does not provide its cell values.Source: Table 2 · p. 70

4.3 Specific Aims evaluation rubric

• The evaluation framework was developed collaboratively with two oncologists and adapted from NIH Specific Aims criteria for AI-generated proposals.Because LLMs synthesize public information without human researchers’ tacit knowledge or access to unpublished findings, the framework emphasizes logical consistency, citation quality, literature integration, and factual accuracy.Source: §4.3 · p. 70
• Proposals were evaluated across 15 criteria spanning clinical significance and scientific rigor and methodology.Clinical significance covered unmet needs, therapeutic-landscape understanding, and repurposing rationale; scientific rigor covered hypothesis clarity, experimental design, endpoints, translational potential, and hallucinated or fabricated content.Source: §4.3 · p. 71
• Each criterion used a 5-point Likert scale, with evaluations conducted independently by two groups of expert hematologist-oncologists.The scale ranged from Strongly Disagree to Strongly Agree and enabled quantitative assessment across proposals.Source: §4.3 · p. 71
• The framework is a starting point rather than a validated substitute for comprehensive grant review.It omits areas such as statistical power, detailed study design, budgets, and investigator qualifications, and would require reliability, validity, and benchmark studies for formal validation.Source: §4.3 · p. 71
• The significance and innovation questions assessed unmet clinical needs, therapeutic gaps, evidence-based rationale, relevant prior studies, and avoidance of unsupported extrapolation.These questions addressed whether repurposing proposals were clinically important, scientifically grounded, and appropriately constrained by available evidence.Source: §4.3 · pp. 71–72
• The rigor and feasibility questions assessed aims, methods, translational pathways, endpoints, preclinical design, factual accuracy, evidence-grounded assumptions, originality, terminology, and organization.The criteria also evaluated whether proposals described clinically relevant efficacy and safety studies and avoided hallucinations, ambiguity, or generic statements.Source: §4.3 · p. 72

5.1 Co-Scientist suggests plausible drug repurposing candidates as rated by computational biology analyses and experts

• Co-Scientist searched 2300 approved drugs across 34 cancer types, generating each drug-cancer proposal with a hypothesis, review, and possible mechanism of action.Generation and Ranking prompts were modified for this constrained search space while the core Co-Scientist logic remained unchanged.Source: §5.1 · pp. 73–74
• Candidates were prioritized using Co-Scientist review scores and DepMap scores, retaining pairs with review score ≥4 and DepMap score ≥0.99 for expert review.Higher Co-Scientist scores correlated with higher DepMap scores, and expert oncologists selected promising candidates for in vitro validation.Source: §5.1 · p. 75

5.2 AML cell line selection rationale and mechanistic interpretation of drug sensitivities

• AML cell lines were selected to represent genetic and phenotypic diversity, including driver mutations, AML subtypes, isogenic controls where available, and clinically characterized molecular states.The strategy aimed to distinguish on-target from off-target effects, interpret pathway responses, and guide later validation in primary patient samples.Source: §5.2 · pp. 76–77
• KIRA6 sensitivity tracked AML differentiation state, with primitive CD34⁺ KG-1a cells more susceptible than differentiated MOLM-13 and HL-60 cells.The KG-1a IC50 was 18-fold lower than that of TK6; NOMO-1 showed intermediate sensitivity with IC50 = 144 nM, supporting differential dependence on IRE1α-mediated proteostasis.Source: §5.2 · p. 79

5.3 Details of the selected repurposed drugs

• The selected repurposed drugs targeted MEK1/2, JAK2/FLT3, Nrf2/NFκB, metabolic reprogramming and vesicular transport, PRMT5, and XPO1.Examples included binimetinib [5, 6], pacritinib [7, 8], dimethyl fumarate [9,10], statins [11], JNJ-64619178 [12,13], and selinexor.Source: §5.3 · pp. 80–81

5.4 In vitro wet-lab validation setup

• Single-agent efficacy was measured by IC50 across MOLM-13, KG-1a, HL-60, NOMO-1, and TK6 using serial dilution, viability assays, and dose-response curve fitting.TK6 served as a non-AML control for assessing AML-selective cytotoxicity and a potential therapeutic window.Source: §5.4 · pp. 82–83
• Combination treatments used fixed molar ratios and Chou-Talalay analysis, classifying interactions as synergistic when CI < 1, additive when CI ≈ 1, or antagonistic when CI > 1.Eight-point dilution series were tested in triplicate after single-agent dose-response measurements.Source: §5.4 · pp. 83–84
• The validation setup measured three-drug dose grids as fractional effects and assessed interactions using Bliss independence and Highest Single Agent benchmarks.Three-drug interactions were visualized with dose–dose heatmaps, using the maximum interaction across the third drug dose and annotating the dose producing that maximum.Source: §5.4 · p. 84

5.5 Additional wet-lab results

• MOLM-13 responses were predominantly synergistic, whereas KG-1a showed context-dependent synergy and antagonism across Co-Scientist-proposed drug combinations.Strong synergy included JNJ-64619178 + Selinexor, JNJ-64619178 + SNDX-5613, Palbociclib + Selinexor, and selected triple combinations in MOLM-13; KG-1a responses varied by combination and effect level. Extended Data Table 2 summarizes the combination findings.Source: §5.5 · pp. 84–85

5.6 Clinical variable analysis guideline for AML drug repurposing

• The clinical-variable framework evaluates repurposed drugs by patient characteristics, AML molecular and cytogenetic risk, preclinical models, safety, and PK/PD considerations.It incorporates factors including age, performance status, comorbidities, prior treatment, cell-line selection, off-target toxicity, achievable exposure, drug interactions, and response biomarkers.Source: §5.6 · pp. 85–89
• Binimetinib analysis identified older or frail patients with relapsed or venetoclax-resistant AML as a specific unmet-need context.The analysis linked this context to MAPK-pathway activation and polypharmacy challenges in venetoclax-resistant disease.Source: §5.6 · p. 90
• Binimetinib’s UGT1A1-mediated metabolism may avoid CYP3A4-dependent interactions with azole antifungals, supporting its suitability for frail or heavily pretreated patients.The conclusion specifically concerns settings where CYP3A4 inhibition limits the feasibility of other targeted regimens.Source: §5.6 · p. 90

Summary

• KIRA6 is proposed as a novel repurposing candidate for FLT3-ITD-positive AML, particularly MOLM-13 cells, with combination therapy intended to overcome resistance.The proposal targets a high-risk AML subtype and pairs KIRA6 with FLT3 inhibitors or chemotherapeutics.Source: Summary · p. 91, Areas worth exploring · p. 93

Molecular mechanism of action

• By inhibiting IRE1α, KIRA6 blocks adaptive UPR signaling, disrupts protein homeostasis, induces ER stress, and activates apoptotic pathways, with stronger effects expected in stressed FLT3-ITD cells.The proposed downstream effects include impaired XBP1 signaling, ISR activation, translational suppression, and accumulation of misfolded proteins.Source: Molecular mechanism of action · pp. 94–95

Impacted pathways

• KIRA6 is expected to suppress MYC, NF-κB, MCL-1, inflammatory signaling, and other resistance mechanisms while sensitizing AML cells to FLT3 inhibitors and chemotherapy.The proposed resistance effects include reduced production of short-lived survival proteins, efflux pumps, and DNA-repair factors.Source: Molecular mechanism of action · p. 95, Impacted pathways · p. 95, Effect on AML (MOLM-13) cell… · p. 95

Effect on AML (MOLM-13) cell proliferation inhibition

• KIRA6 is predicted to selectively inhibit MOLM-13 proliferation and reduce viability dose-dependently, with combination treatment potentially achieving greater cell death at lower drug concentrations.These effects are attributed to ER stress, ISR activation, apoptosis, and reduced inflammatory signaling, but remain predictions requiring validation.Source: Effect on AML (MOLM-13) cell… · pp. 95–96

IC50 assay concentrations

• The proposed IC50 assay tests KIRA6 from 100 nM to 10 µM for 48–72 hours, using viability assays, vehicle controls, gilteritinib normalization, and combination matrices.Promising combinations would receive additional apoptosis and cell-cycle confirmation.Source: IC50 assay concentrations · p. 96

Safety and toxicity

• KIRA6 has limited human safety data and may cause ER-stress toxicity in high-protein-synthesis tissues, so preclinical toxicity and combination-dose studies are required before clinical translation.Repurposing and dose reduction through combinations may reduce development risk and systemic toxicity, but these benefits require rigorous testing.Source: Safety and toxicity · pp. 96–97

Experimental plan

• The experimental plan measures ER stress, IRE1α activity, apoptosis, ISR markers, pathway proteins, cell cycle, efflux pumps, combination synergy, generalizability, xenograft efficacy, and pharmacokinetics.Testing expands beyond MOLM-13 to AML lines with varied mutations, primary AML cells, stromal co-cultures, and non-leukemic controls.Source: Experimental plan · pp. 97–98, Conclusion · p. 98

Novelty review

• The proposal is moderately novel because KIRA6 use in AML and its combinations with FLT3 inhibitors or chemotherapy appear underexplored, although IRE1α targeting in AML has been studied and key selectivity and synergy assumptions remain unvalidated.The recommendation is to pursue testing across diverse cell lines and patient samples while further investigating resistance mechanisms and comparing other IRE1α inhibitors.Source: Reasoning about assumptions · p. 100, Improvements to the idea · p. 100, Novel aspects · pp. 101–102

Reasoning about novelty and recommendation

• The recommendation is reasoned against related literature on AML signaling, FLT3 inhibition, unfolded-protein-response targeting, and approved-drug repurposing.The cited works provide context for assessing novelty and therapeutic rationale.Source: Reasoning about novelty and… · p. 103

Critiques

• KIRA6 requires preclinical toxicity and in vivo studies because its human safety profile is unknown and current evidence lacks clinical validation.The critiques also call for testing beyond MOLM-13, including AML cell lines with diverse genetic backgrounds and primary patient samples.Source: Critiques · p. 104
• The proposed efficacy, selectivity, drug synergy, resistance effects, IC50 range, and impacts on MYC, NF-κB, inflammatory pathways, and UPR require experimental validation and stronger comparative benchmarking.Reviewers specifically requested comparisons with other IRE1α inhibitors, deeper resistance studies, and investigation of whether KIRA6 could itself induce resistance.Source: Critiques · p. 104

7 Safety and ethical implications

• Co-Scientist’s acceleration of science creates distinct dual-use safety risks and ethical risks involving research that conflicts with disciplinary norms.The paper emphasizes that further research is needed to understand and mitigate both categories.Source: §7 · p. 105

Evolving ethics frameworks, policy and regulations for advanced AI use in scientific

• Advanced agentic AI requires updated ethics policy and regulation addressing capabilities, autonomy, user intent, domain risks, and broader impacts.Existing frameworks assess risks including persuasion, deception, cybersecurity, self-proliferation, self-reasoning, manipulation, and value alignment.Source: Evolving ethics frameworks, policy and… · p. 105
• Scientific AI safety requires adversarial testing across iterative reasoning components, unsafe queries, intermediate and final hypotheses, internal checks, and filters.Automated attacks can exploit prompt suffixes, few-shot demonstrations, in-context learning, multimodal inputs, and iterative interactions, while specialized human expertise remains necessary.Source: Evolving ethics frameworks, policy and… · p. 106
• A comprehensive safety approach combines threat modeling, defenses, red-teaming, rapid vulnerability response, continuous monitoring, access controls, logging, and manual review.The paper also calls for configurable safeguards, gradual rollouts, trusted testers, community feedback, and inclusive best-practice development.Source: Evolving ethics frameworks, policy and… · p. 106
• Co-Scientist rejects unsafe research goals and hypotheses, monitors research directions, provides reasoning traces, logs activity, and maintains continuous human expert oversight.Preliminary red teaming assessed unsafe-goal detection using 1,200 adversarial research goals across 40 topic areas.Source: Evolving ethics frameworks, policy and… · p. 107

Initialization

• Co-Scientist parses the scientist’s research goal into a structured ResearchPlan, saves it in SharedMemory, and queues an initial hypothesis-generation task.The workflow begins with StartCoScientist and creates a Generation-agent task for CreateInitialHypotheses.Source: §8 · p. 108

Evolution

• When hypothesis quality stops improving, the Evolution agent refines top ideas by combining, simplifying, or reimagining them, then routes each evolved hypothesis back for review.The procedure fetches the top 5 hypotheses and treats each evolved output as a new idea.Source: §8 · pp. 109, 111, 112

System feedback

• The system can synthesize feedback by gathering reviews and tournament debate transcripts from SharedMemory and asking a language model to analyze the critiques.This system-wide feedback is created as a queued Metareview task.Source: §8 · pp. 109, 112

Hypothesis generation and review

• Initial hypotheses are generated through literature search and simulated expert debate, saved in SharedMemory, and sent to Reflection for review.The review evaluates novelty, correctness, and scientific plausibility using retrieved evidence and separate verification.Source: §8 · pp. 109–110

Tournament ranking

• Reviewed hypotheses enter a tournament in which paired ideas are debated, a winner is selected, and both hypotheses receive updated Elo ratings in SharedMemory.The tournament selects pairs, prioritizes new or similarly rated hypotheses, and uses a language-model debate to determine outcomes.Source: §8 · p. 111

• The system summarizes recurring weaknesses and strengths as feedback for the whole system, then stores this summary in SharedMemory.Source: §8 · p. 112
• The final overview procedure retrieves the top 10 hypotheses from SharedMemory and asks a language model to synthesize them.Source: §8 · pp. 112–113
• The generated overview is returned to the scientist with the main research directions and their justifications.Source: §8 · p. 113

9.1 Prompts for the Generation agent

• The Generation agent formulates novel, robust hypotheses from literature reviews, existing hypotheses, analytical reasoning, and detailed domain-expert descriptions.The prompt supplies consulted articles with reasoning and requests a detailed proposed hypothesis covering entities, mechanisms, and anticipated outcomes.Source: §9.1 · p. 114
• After scientific debate, the Generation agent proposes three distinct hypotheses and iteratively evaluates, critiques, improves, and refines them collaboratively.The process includes clarifying uncertainties, assessing utility and practicality, identifying limitations, and concluding with a refined hypothesis.Source: §9.1 · p. 115
• The debate-based generation prompt requires critique against hypothesis criteria, including adherence, utility, practicality, detail, specificity, and potential weaknesses.Agents must propose concrete improvements and maintain a bold, creative, collaborative approach while prioritizing hypothesis quality.Source: §9.1 · p. 115

9.2 Prompt for the Reflection agent

• The Reflection agent extracts article observations and determines whether a hypothesis causally explains them, offers a novel explanation, or is contradicted.It performs individual and summary causal analyses, distinguishing established causes from potential missing explanatory pieces.Source: §9.2 · p. 116
• Reflection conclusions classify hypotheses as already explained, other explanations more likely, missing piece, neutral, or disproved.Observations expected regardless of the hypothesis are classified as neutral when they do not disprove it.Source: §9.2 · pp. 116–117

9.3 Prompts for the Ranking agent

• The Ranking agent compares two hypotheses and selects the superior one according to specified attributes, providing a concise rationale and final choice.Independent numerical review scores are explicitly disregarded because they may not be comparable across reviews.Source: §9.3 · p. 117
• In tournament-based simulated scientific debate, experts compare hypotheses using predefined criteria and critically evaluate each against the stated goal, including potential correctness/validity.Source: §9.3 · pp. 118–119
• The Ranking agent identifies the superior hypothesis and states its rationale using the required “better idea” output format.Source: §9.3 · p. 119

9.4 Prompts for the Evolution agent

• The Evolution agent improves a conceptual hypothesis’s practical implementability while retaining novelty, logical coherence, and specific articulation.Source: §9.4 · p. 119
• The feasibility-improvement prompt requires domain context, recent research findings, successful methodologies, precedents, and a reasoned account of how current technology enables realization.Source: §9.4 · p. 119
• The evolution prompt’s core contribution is a detailed, innovative, technologically viable alternative that emphasizes simplicity and practicality.Source: §9.4 · p. 120
• The out-of-the-box generation prompt creates one novel hypothesis from analogous elements in provided concepts rather than directly replicating or merely aggregating existing methods or entities.Source: §9.4 · p. 120
• Hypothesis generation is structured around domain introduction, recent findings, successful research, and promising exploratory avenues before articulating a detailed, original, specific hypothesis.Source: §9.4 · p. 120

9.5 Prompt for the Meta-review agent

• The Meta-review agent synthesizes provided reviews into a structured meta-analysis focused on recurring critiques and common issues.Source: §9.5 · pp. 120–121
• The meta-analysis should provide actionable insights for future proposals while avoiding evaluation of individual proposals or reviews.Source: §9.5 · p. 121

10.1 From research goal to research plan configuration

• Co-Scientist parses a natural-language ALS research goal into a plan emphasizing a novel, feasible, correct hypothesis with a detailed mechanism and experiment.The example goal concerns phosphorylation of an NPC nucleoporin and requires an experimentally testable explanation of ALS-related mechanisms.Source: §10.1 · p. 122, Scientist research goal · p. 122, Parsed research plan configuration · p. 122

10.2 Example output of the Generation agent

• The generated hypothesis proposes that cellular stress induces phosphorylation or O-GlcNAcylation of Nup98/Nup62, altering their interaction with TDP-43 and increasing its retention at the NPC.The proposed retention is intended to disrupt nucleocytoplasmic transport and provides a concrete, testable alternative to focusing directly on LLPS.Source: Hypothesis · p. 123, Rationale and specificity · p. 123

10.3 Example output of a complete novelty review by the Reflection agent

• The novelty review identifies the stress-induced PTMs of Nup98/Nup62 and their direct effects on TDP-43 interaction and NPC retention as the hypothesis’s novel elements.TDP-43 pathology, transport defects, NPC disruption, and FG-Nup barrier function were described as established, whereas the specific Nup PTM mechanism was characterized as new.Source: §10.3 · pp. 124–125

10.4 Example output of review critiques by the Reflection agent

• The review critiques note missing justification for motor-neuron specificity, unclear downstream consequences and timing, technical difficulty in validating PTM effects, and risks from a narrow, non-holistic focus.These critiques concern whether the mechanism precedes TDP-43 aggregation, how transport defects produce ALS pathology, and whether other Nups, PTMs, or cellular processes matter.Source: §10.4 · p. 125

10.5 Example output of a full review by the Reflection agent

• The full review lists assumptions spanning stress-induced Nup PTMs, altered Nup–TDP-43 binding, NPC retention, transport disruption, ALS pathology, and greater motor-neuron susceptibility.It judges the PTM and interaction steps plausible but requiring experimental validation, notes support for transport defects in ALS, and finds motor-neuron vulnerability insufficiently justified.Source: §10.5 · p. 127

10.6 Example output of deep verification by the Reflection agent

• Deep verification decomposes the proposal into testable claims, including stress-induced PTMs, Nup98/Nup62–TDP-43 interactions, NPC retention, transport disruption, downstream ALS pathology, and motor-neuron vulnerability.The verification also checks whether stress is sufficient to alter TDP-43 localization and whether TDP-43 localizes to the NPC.Source: §10.6 · p. 128

1. Cellular stress induces PTMs like phosphorylation and O-GlcNAcylation.

• ER stress activates the unfolded protein response, in which PERK and IRE1 phosphorylation pathways regulate translation, protein folding, and degradation.PERK phosphorylates eIF2α to attenuate translation, while IRE1 autophosphorylates and promotes XBP1 splicing.Source: §1 · pp. 128–129
• Cellular stress alters phosphorylation and O-GlcNAcylation, but O-GlcNAcylation can increase or decrease depending on stress conditions, cell type, nutrient availability, and target proteins.O-GlcNAcylation may protect cells by enhancing chaperone activity and protein folding, while oxidative stress, nutrient deprivation, and hypoxia also reshape PTMs.Source: §1 · pp. 128–129

10.7 Example output of deep verification via probing questions by the Reflection agent

• The Reflection agent’s AML example evaluates Reparixin as a CXCR1/2-inhibiting drug-repurposing candidate, requiring testable concentrations, AML selectivity, mechanism, pathways, and safety assessment.The research goal emphasizes inhibition of MOLM-13 proliferation, minimal toxicity in healthy cells, and no prior experimental testing for AML.Source: §10.7 · p. 129, [Research Goal] · pp. 129–130

[Output]

• The AML analysis treats single-agent Reparixin as a baseline for studying CXCR1/2 inhibition, resistance mechanisms, patient stratification, tumor-microenvironment reprogramming, and sequential combination therapy.It acknowledges that AML heterogeneity and compensatory pathways may make combination therapies necessary, rather than claiming universal single-agent efficacy.Source: [Output] · pp. 130–131

10.8 Example output of an observation review by the Reflection Agent

• The Reflection agent’s AMR example proposes the Modular Capsid-OMV Mediated Transfer hypothesis to explain how cf-PICIs spread across different bacterial species.The hypothesis combines modular capsid interactions with tail-dependent and tail-independent OMV-mediated transfer, supported by PICI distributions and distinctive att sites.Source: §10.8 · p. 131, [Research Goal] · p. 131, [Output] · p. 132

10.9 Example of a scientific debate match in the Ranking agent tournament

• In the ALS ranking debate, experts favored Idea 1 because it offered a more focused, experimentally tractable mechanism, while questioning Idea 2’s evidence for ISR as a trigger and its technical challenges.The discussion also called for rigorous tests of motor-neuron specificity, stress effects, and PTM-mediated interactions.Source: §10.9 · pp. 132–133

I. Core Hypothesis and Mechanism:

• Co-Scientist’s ALS hypotheses should establish whether mechanisms are initiating causes rather than consequences through longitudinal experiments, early-stage models, and targeted perturbations.Proposed drivers should be specifically knocked down to test whether they are necessary and sufficient, while addressing motor-neuron specificity, disease subtypes, molecular targets, cellular compartments, stress types, and timing.Source: §I · pp. 133–134
• ALS research designs should prioritize feasible, interpretable experiments with validated tools, quantitative outcomes, appropriate controls, replicates, sample sizes, statistics, and time-course interventions.Because iPSC-derived motor neurons may not capture in vivo interactions or aging, studies should acknowledge model limitations and, where possible, validate findings across multiple systems.Source: §I · pp. 134–135

V. General Advice Based on Common Critiques:

• The guidance favors mechanistically detailed, literature-grounded hypotheses with strong rationales, explicit assumptions, phased feasibility, specific controls, alternative explanations, and therapeutic relevance without overstating “key switch” claims.Critical assumptions should be tested in the order of the proposed mechanism, since failure of early steps may make later validation unnecessary.Source: §I · pp. 135–136, §V · p. 136

10.11 Example of a research overview by the Meta-review agent

• In the Meta-review example, Co-Scientist presents a research overview to scientists, demonstrating how generated hypotheses can be organized into interconnected research directions.The overview is described as the main output presented to the scientist, while the contact-identification example extends the workflow toward relevant domain expertise.Source: §10.11 · p. 136, §10.12 · p. 140

[Main Research Directions]

• The ALS research overview organizes investigation around mitochondrial dysfunction and oxidative stress, RNA processing and stress granules, and protein homeostasis with nucleocytoplasmic transport.These directions connect mitochondrial, RNA, proteostatic, and transport defects to inflammatory signaling, neurodegeneration, protein homeostasis, and neuronal survival.Source: [Main Research Directions] · pp. 136–137

Mitochondrial Dysfunction and Oxidative Stress

• The mitochondrial direction proposes that specific early vulnerabilities in mtDNA repair, calcium handling, and ROS regulation may initiate ALS, with OGG1, oxidized mtDNA, MCU, and cGAS-STING as testable elements.The protein-homeostasis direction examines UPS and autophagy defects, including TDP-43 and FUS degradation, cargo recognition, autophagosome formation, lysosome fusion, and cell-type-specific vulnerability.Source: Mitochondrial Dysfunction and… · pp. 137–139

Research Direction: Oxidative DNA Damage & Mitochondrial Base Excision Repair (BER) in ALS

• The BER research direction focuses on mitochondrial OGG1, oxidized mtDNA lesions, cytosolic mtDNA release, and cGAS-STING activation as possible links between mitochondrial damage and neuroinflammation.The overview also identifies potential domain experts studying mitochondrial OGG1, mtDNA release, neuroinflammation, and mtOGG1-overexpressing mice.Source: Mitochondrial Dysfunction and… · pp. 137–138, Research Direction: Oxidative DNA… · p. 140

11 An example of tool use in Co-Scientist with AlphaFold

• AlphaFold validates Co-Scientist protein-sequence proposals by assessing structural plausibility and feeding results back into iterative design, illustrated by a modified OCT4 sequence whose predicted stability and interaction metrics remained comparable.ESM-2, ESMFold, and RoseTTAFold provided independent supportive predictions, but the OCT4 example remains demonstrative and requires binding-affinity, off-target, and laboratory validation.Source: §11 · p. 141, Figure 9 · p. 142

12 Related works

• Co-Scientist scales test-time compute using scientific-method inductive biases in a multi-agent framework for scientific reasoning and hypothesis generation.Source: §12 · p. 143
• Unlike specialized AI tools such as AlphaFold 2 and AI systems for antibiotics, protein binders, and materials, Co-Scientist targets broader scientific discovery.Source: §12 · p. 143
• LLM research assistants have addressed manuscript feedback and literature synthesis, but Liang et al. 98 and PaperQA2 99 do not cover novel hypothesis generation through the full research cycle.Source: §12 · p. 144
• HypoGeniC 100 and data-to-paper 101 generate or refine research outputs, yet retrospective evaluation or publication recapitulation leaves truly novel, grounded hypotheses insufficiently established.Source: §12 · p. 144
• Virtual Lab 102 experimentally validated LLM-designed nanobody binders, but its generality remains unclear and its design differs substantially from Co-Scientist.Source: §12 · p. 144
• Unlike Coscientist 103, which focuses on autonomous chemical experimentation and laboratory hardware, Co-Scientist is broadly applicable, self-improving, and scientist-in-the-loop.Source: §12 · pp. 144–145
• Compared with The AI Scientist 96, Co-Scientist emphasizes test-time scaling and combines automated, human-expert, and wet-lab validation rather than fully automating discovery.Source: §12 · p. 145
• In biomedicine, Co-Scientist builds on general and specialized AI models while offering scalable predictions validated through expert evaluations and wet-lab experiments.Source: §12 · pp. 145–146