Source-linked AI summary
Kosmos: An AI Scientist for Autonomous Discovery
Ludovico Mitchener, Angela Yiu, Benjamin Chang, Mathieu Bourdenx, Tyler Nadolski, Arvis Sulovari, Eric C. Landsness, Daniel L. Barabasi, Siddharth Narayanan, Nicky Evans, Shriya Reddy, Martha Foiani, Aizad Kamal, Leah P. Shriver, Fang Cao, Asmamaw T. Wassie, Jon M. Laurent, Edwin Melville-Green, Mayk Caldas, Albert Bou, Kaleigh F. Roberts, Sladjana Zagorac, Timothy C. Orr, Miranda E. Orr, Kevin J. Zwezdaryk, Ali E. Ghareeb, Laurie McCoy, Bruna Gomes, Euan A. Ashley, Karen E. Duff, Tonio Buonassisi, Tom Rainforth, Randall J. Bateman, Michael Skarlinski, Samuel G. Rodriques, Michaela M. Hinks, Andrew D. White
TL;DR
Existing AI research agents remain limited in the number of coherent actions they can take, constraining the depth of automated data-driven discovery. Kosmos addresses this gap with a structured world model coordinating parallel literature search and data analysis over extended cycles. It produced seven discoveries, with three reproducing findings not accessed at runtime and four making novel contributions, while collaborators estimated a 20-cycle run as equivalent to 6.14 months of research.
Problem
Existing AI agents for scientific research remain limited in the number of actions they can take before losing coherence, restricting automated discovery depth.
Method
Kosmos uses a structured world model to coordinate parallel data-analysis and literature-search agents across iterative discovery cycles and produce traceable scientific reports.
Results
Seven discoveries were reported: three reproduced findings from manuscripts not accessed at runtime, while four made novel contributions to the scientific literature.
Takeaways & Limitations
Kosmos autonomously conducted extended investigations across multiple scientific fields, with discoveries validated by domain experts and claims linked to code or primary literature.
Takeaways & Limitations
Kosmos is limited to datasets of approximately 5GB, does not excel on raw data, and cannot autonomously access external public data for reference or orthogonal validation.
Abstract
from arXiv · showhide
Data-driven scientific discovery requires iterative cycles of literature search, hypothesis generation, and data analysis. Substantial progress has been made towards AI agents that can automate scientific research, but all such agents remain limited in the number of actions they can take before losing coherence, thus limiting the depth of their findings. Here we present Kosmos, an AI scientist that automates data-driven discovery. Given an open-ended objective and a dataset, Kosmos runs for up to 12 hours performing cycles of parallel data analysis, literature search, and hypothesis generation before synthesizing discoveries into scientific reports. Unlike prior systems, Kosmos uses a structured world model to share information between a data analysis agent and a literature search agent. The world model enables Kosmos to coherently pursue the specified objective over 200 agent rollouts, collectively executing an average of 42,000 lines of code and reading 1,500 papers per run. Kosmos cites all statements in its reports with code or primary literature, ensuring its reasoning is traceable. Independent scientists found 79.4% of statements in Kosmos reports to be accurate, and collaborators reported that a single 20-cycle Kosmos run performed the equivalent of 6 months of their own research time on average. Furthermore, collaborators reported that the number of valuable scientific findings generated scales linearly with Kosmos cycles (tested up to 20 cycles). We highlight seven discoveries made by Kosmos that span metabolomics, materials science, neuroscience, and statistical genetics. Three discoveries independently reproduce findings from preprinted or unpublished manuscripts that were not accessed by Kosmos at runtime, while four make novel contributions to the scientific literature.
1 Introduction
Kosmos is designed to automate data-driven discovery across scientific disciplines by coordinating iterative literature search, hypothesis generation, and data analysis. Its structured world model supports extended, traceable investigations that reproduce existing findings and generate novel contributions.
- Kosmos automates data-driven discovery across a wide range of scientific disciplines.
- The system performs parallel data analysis, literature search, and hypothesis-generation cycles from an open-ended objective and dataset.
- A structured world model shares and synthesizes information between agents, enabling 42,000 lines of code and 1,500 papers per run.
- Kosmos reports seven discoveries: three reproduce findings from inaccessible preprinted or unpublished manuscripts, while four make novel contributions.
2.1 Kosmos system and architecture
Kosmos uses a structured world model to coordinate parallel analysis and literature-search agents across iterative cycles, then synthesizes traceable scientific reports. Evaluations indicate substantial expert-equivalent research output, alongside accuracy differences across statement types and discoveries spanning reproduction, methodological development, and novel findings.
- Kosmos system and architecture: Kosmos updates a structured world model after each cycle to propose subsequent literature-search and data-analysis tasks.Each cycle executes up to ten literature-search and analysis tasks before updating the model.
- Kosmos system and architecture: Each report statement and figure cites either a literature source or a Jupyter notebook created by Kosmos.
- Performance evaluation: 79.4% of report statements were accurate overall, including 85.5% of data-analysis statements, 82.1% of literature-review statements, and 57.9% of synthesis statements.
- Expert-equivalent research time: 4.1 expert-months of research were estimated for each Kosmos run, while academic groups estimated 6.14 months for a 20-cycle run.The first estimate used counts of analyses and papers; the second was an orthogonal estimate from collaborating academic groups.
- Discovery outcomes: The seven studies include reproductions of inaccessible existing discoveries, additional support for existing discoveries, a new analytical method, and a novel clinically relevant discovery.The categories comprise two inaccessible reproductions, one independently reasoned reproduction, two runs adding novel support, one new method, and one clinically relevant discovery.
2.2 Kosmos replicates human findings from different fields
Kosmos reproduced findings across metabolomics, materials science, and neuroscience using open-ended objectives and domain-specific datasets. Its analyses recovered established patterns while extending some findings with mechanistic interpretations and experimentally relevant relationships.
- Metabolomics: Kosmos identified nucleotide-salvage activity as the likely metabolic response associated with hypothermic neuroprotection.Precursor bases and nucleosides decreased while phosphorylated nucleotide products increased; correlations did not support substrate-driven de novo synthesis, though the conclusion was not definitive.
- Materials science: Kosmos corroborated thermal-annealing humidity as the dominant single environmental factor affecting perovskite solar-cell efficiency.It identified a failure boundary above ~60 g/m3 absolute humidity within a critical temperature window above ~70 °C, and found a previously unreported linear decrease in JSC with increasing DMF solvent partial pressure.
- Neuroscience: Kosmos found strong positive scaling relationships among neuronal wire length, synapse count, and degree across connectome datasets.The correlations supported conserved neuronal-property scaling across species, while the analyses reproduced the underlying preprint’s major quantitative findings.
- Neuroscience: Neuron-level synapse counts and degree were best fit by log-normal distributions, while synapse count versus neurite length showed robust power-law scaling in seven of eight datasets.The fitted log-normal means showed high concordance with the preprint, including Pearson’s r = 0.77 for Synapses and r = 0.46 for Degrees.
- Neuroscience: Kosmos connected empirical neuronal distributions to multiplicative processes in neurodevelopment and reproduced the source preprint’s major quantitative results.This interpretation was presented as a proposed neuroscientific underpinning of the observed log-normal distributions.
2.3 Kosmos adds additional support to existing findings with novel methods
Kosmos applied autonomous analyses to myocardial fibrosis and Type 2 Diabetes datasets, reproducing human findings while proposing causal mechanisms and prioritization strategies. These results demonstrate additional support for existing findings through independent computational approaches.
- Discovery 4: SOD2 as a driver of myocardial fibrosis in humans: Kosmos autonomously executed a Mendelian randomization pipeline to identify candidate proteins and propose mechanisms for myocardial fibrosis.The provided dataset included myocardial T1 GWAS summary statistics and cis-acting plasma pQTLs.
- Discovery 4: SOD2 as a driver of myocardial fibrosis in humans: SOD2 was identified by both Kosmos and human MR analyses as the primary causal protein, with concordant protective effect estimates.Kosmos reported β = −0.231 (p = 4.23×10−13), while human analysis yielded β = −0.258 (p = 1.22×10−22).
- Discovery 4: SOD2 as a driver of myocardial fibrosis in humans: 32 proteins met the Bonferroni-corrected MR significance threshold, with 31 identified by both approaches and effect sizes showing Pearson r = 0.9991.The shared proteins had consistent directionality, with R2 = 0.9983.
- Discovery 4: SOD2 as a driver of myocardial fibrosis in humans: Kosmos proposed that rs4555948 in the SOD2 3′ UTR may disrupt hsa-miR-222-3p binding, but current prediction databases do not place a binding site at that variant.Literature supports miR-222 regulation of SOD2 expression without specifying the exact binding position.
- Discovery 5: Cis-regulation of SSR1 by a protective GWAS variant: For Type 2 Diabetes, Kosmos independently prioritized likely causal protective mechanisms using an integrated variant-level multi-omic dataset.The inputs combined GWAS fine-mapping, TFBS, ATAC-seq, eQTL, and pQTL annotations.
- Discovery 5: Cis-regulation of SSR1 by a protective GWAS variant: Kosmos created the Mechanistic Ranking Score, combining fine-mapping probability, QTL concordance, and experimental evidence to rank candidate mechanisms.MRS = PIP ×(1+Concordance Score+Experimental Evidence Score).
- Discovery 5: Cis-regulation of SSR1 by a protective GWAS variant: The rs9379084 locus received the highest MRS, and independent enhancer-target and TWAS data supported SSR1 as the relevant regulated gene.SSR1 was the only gene at the locus with transcriptome-wide significant association with T2D, with TWAS |Z| > 5.
- Discovery 5: Cis-regulation of SSR1 by a protective GWAS variant: This discovery showed that Kosmos can prioritize causal mechanisms and formulate specific hypotheses supported by independent data sources and literature.The result extended autonomous analysis beyond following a predefined analytical pipeline.
2.4 Kosmos independently develops new methods
Kosmos developed an unconventional method to determine when molecular pathways change along a disease continuum. In Alzheimer’s disease data, it used continuous pseudotime analysis and breakpoint modeling to identify and validate temporal ECM decline.
- Discovery 6: Temporal ordering of disease-related events: Kosmos proposed a data-science-driven approach to pinpoint when a cellular process was affected across a disease continuum.The approach addressed research objectives requiring a temporal sequence of disease-related events.
- Discovery 6: Temporal ordering of disease-related events: The Alzheimer’s disease analysis compared proteomes from tau-positive and tau-negative neuron minipools collected from 10 cases.Each minipool contained 20 neurons.
- Discovery 6: Temporal ordering of disease-related events: Early-versus-late pseudotime analysis identified extracellular matrix proteins as enriched among down-regulated proteins.Kosmos then selected ECM proteins to calculate a composite abundance score along pseudotime.
- Discovery 6: Temporal ordering of disease-related events: Kosmos hypothesized that ECM decline followed a nonlinear pattern and used segmented regression to estimate a breakpoint in the disease timeline.The figure reports a breakpoint at 0.58 pseudotime units.
- Discovery 6: Temporal ordering of disease-related events: The proposed breakpoint was supported across independent computational trials, bootstrapping, and an orthogonal sliding-window correlation analysis.Both alternative approaches produced identical conclusions.
- Discovery 6: Temporal ordering of disease-related events: An independent transcriptomic dataset also showed ECM score decline between tau-negative and tau-positive Alzheimer’s disease neurons.The comparison had adjusted p = 2.75×10−92 in the figure’s pairwise Mann–Whitney U test.
- Discovery 6: Temporal ordering of disease-related events: The analysis demonstrated that Kosmos can propose unconventional biological methods beyond standard differential expression and pathway enrichment analyses.The authors state that segmented modeling had not previously been applied in this biological context to investigate molecular events along a cellular timeline.
2.5 Kosmos makes novel clinical discoveries not identified by human researchers
Kosmos uncovered a clinically relevant mechanism of entorhinal cortex neuron vulnerability in aging that the original dataset researchers had not identified. Iterative transcriptomic analysis, literature review, hypothesis testing, and independent validation linked age-related flippase loss to microglial phagocytic activity.
- Discovery and objective: Kosmos uncovered a clinically relevant mechanism of entorhinal cortex neuron vulnerability in aging that the original dataset researchers had not identified.The study compared vulnerable ENT neurons with resilient CTX neurons using mouse single-nuclei RNA-sequencing data.
- Discovery and objective: Aging remains a major dementia risk factor, but the molecular mechanisms underlying vulnerability remain poorly understood.Entorhinal cortex vulnerability is relevant to early Alzheimer’s disease pathology.
- Iterative analysis: Nine additional flippase family members were concurrently downregulated, supporting a broader age-related collapse of flippase activity in ENT neurons.Kosmos proposed that reduced flippase activity increases phosphatidylserine exposure on neuronal membranes, creating an “eat-me” signal.
- Validation and relevance: Independent dataset validation and repeated computational replication confirmed joint flippase downregulation and upregulation of the microglial pro-phagocytosis axis.In human Alzheimer’s disease cases, decreased flippase expression in entorhinal supragranular neurons coincided with tau pathology at Braak stage II.
- Validation and relevance: The proposed mechanism may contribute to early loss of entorhinal-to-hippocampal connectivity in Alzheimer’s disease and indicate potential intervention areas.The authors connect reduced flippase activity with phosphatidylserine exposure and microglial clearance.
3 Discussion
Kosmos combines structured context management with extensive parallel exploration to conduct traceable, multi-field scientific investigations. Its limitations center on human dependence for evaluation, interpretive accuracy, dataset constraints, and stochastic or objective-sensitive behavior.
- Discussion: A structured world model coordinates hundreds of agent rollouts, tens of thousands of code lines, and thousands of papers within a single research objective.The system combines closed-loop literature search, data analysis, and world-model updates to support extended investigations.
- Discussion: Kosmos supports reproducibility and discovery across metabolomics, materials science, connectomics, statistical genetics, proteomics, and transcriptomics.Reports support statements with code or primary-literature citations, enabling independent verification or replication.
- Discussion: Human scientists remain responsible for dataset curation, result interpretation, and critical evaluation because input quality affects discoveries and evaluators observed overly strong claims and unexpected trajectories.The authors describe the intended role of Kosmos as augmenting rather than replacing scientists.
- Limitations and Future Work: 85% of data-analysis statements were accurate, but interpretation statements were 57% accurate, and the analyses selected for novelty or interest were not evaluated.The authors attribute interpretive weakness partly to conflating statistical significance with scientific value.
- Limitations and Future Work: Meaningful-discovery identification remains time-intensive and expert-dependent, while the system is limited to approximately 5GB datasets, raw-data analysis, external data access, consistent convergence, objective phrasing, and intermediate interaction.These constraints limit the scalability and flexibility of current deployments.
4 Methods
The methods combine domain-specific datasets and validation analyses with repeated Kosmos runs across metabolomics, materials science, connectomics, statistical genetics, and neurodegeneration. Validation compares Kosmos outputs with preprints, figures, fit parameters, independent Mendelian randomization, and repeated trajectories.
- Evaluation: Expert evaluators assessed 102 statements from three reports as supported or refuted based on reproducibility or primary-literature evidence.Statements were categorized by origin in scientific literature, data analysis, or interpretation between the two.
- Metabolomics: Metabolomics experiments used adult KOR-Cre mice assigned to control, hypothermic-POAKOR+, or normothermic-POAKOR+ conditions after chemogenetic manipulation of the medial preoptic area.The resulting normalized metabolite peak-intensity matrix was analyzed with Kosmos, and group comparisons used unpaired t-tests.
- Metabolomics: Metabolomic profiling combined UHPLC-Orbitrap mass spectrometry, polar and lipid separations, dual ionization modes, and computational peak processing before Kosmos analysis.Processing used XCMS, Compound Discoverer, and Skyline for detection, alignment, and identification.
- Materials Science and Connectomics: Materials-science results were validated against preprinted figures and reported trends, while connectomics results were validated against preprinted figures and fit parameters.The connectome data integrated publicly available single-neuron-resolution reconstructions across multiple species.
- Statistical Genetics: Statistical-genetics results used cis-pQTL exposure data and independently applied inverse-variance-weighted Mendelian randomization with Bonferroni-corrected significance thresholds.The workflow recorded the number of SNP instruments for each protein–outcome pair.
- Neurodegeneration: A neurodegeneration discovery was generated from a dataset run for 35 iterations, consolidated from iteration 8 after manual curation, and supported by 4 of 5 independent human Alzheimer’s disease trajectories.The clinical-relevance test compared early and late affected regions across Braak stages, focusing on the Braak 0-to-II comparison.
6 Statement of contributions
The project was conceived, designed, supervised, engineered, and evaluated by a multi-person team whose members also contributed to research, software development, infrastructure, and optimization.
- Statement of contributions: L.M. and A.Y. conceived and designed the overall project, while L.M. led the core system’s design and engineering.L.M. and M.M.H. supervised the project.
- Statement of contributions: B.C. led the discovery review system and evaluations of Kosmos system performance.The contribution statement also lists contributions to investigating system behavior and performance.
- Statement of contributions: Multiple contributors supported research, software development, code optimization, and platform infrastructure.The passage names contributors across these activities without assigning every activity to a single individual.
7 Funding information
The funding statement lists support from U.S. and U.K. biomedical research organizations, charitable funding, and a CIFAR catalyst award for different contributors’ work.
- Funding information: Work by R.J.B., E.C.L., A.K., and L.P.S. was supported by an NIH National Institute of Neurological Disorders and Stroke award to E.C.L.The passage identifies award R01 NS133365.
- Funding information: Work by K.E.D., M.F., and M.B. was supported by the UK Dementia Research Institute and Cure Alzheimer’s Fund.The UK DRI support was provided through UK DRI Ltd and principally funded by the Medical Research Council.
- Funding information: N.E. and T.B. acknowledged support from a CIFAR catalyst award.
Supplementary Information
The supplementary material lists seven Kosmos discovery topics and provides datasets, objectives, and processing instructions spanning neuroscience, materials science, metabolomics, myocardial fibrosis, diabetes, Alzheimer’s disease, and neural vulnerability.
- Seven discovery topics cover nucleotide metabolism, perovskite solar-cell failure, neural-network connectivity, myocardial fibrosis, type 2 diabetes, Alzheimer’s disease, and neuron vulnerability in aging.
- The connectome study aggregates neuronal degree length, synapse count, and local synaptic density across experimentally mapped connectomes.
- The perovskite study examines solvent partial pressure, absolute humidity, and ambient temperature across spin coating and thermal annealing.
- The hypothermia metabolomics study asks which metabolic adaptations support cerebroprotection during KOR+-induced torpor-like hypothermia and hypometabolism.
- The myocardial-fibrosis analysis seeks causal proteins using colocalization, fine mapping, Mendelian randomization, and linkage-disequilibrium information.
- The type 2 diabetes study ranks SNP–gene mechanisms using QTL validation, directional concordance, experimental evidence, and biological relevance.
- The Alzheimer’s disease analysis compares tau-positive and tau-negative neuronal proteomes while accounting for metadata and possible cofactors.
- The aging study compares entorhinal and cortical neuron subclasses to investigate transcriptional regulation, entropy, noise, vulnerability, and resilience.
Supplementary Information 3: Example Instructions for Expert Evaluations (Discovery 1)
The evaluation instructions define separate procedures for testing Kosmos claims through data analysis, literature review, and interpretation, using discovery reports and datasets as context.
- Dataset Information: The discovery dataset comprises untargeted mouse-brain metabolomics from control, hypothermia, and normothermia groups.
- Supplementary Information 3: Example Instructions for Expert Evaluations (Discovery 1): Evaluators first read the discovery report, then familiarize themselves with the dataset from which the analyses and discoveries were derived.
- Supplementary Information 3: Example Instructions for Expert Evaluations (Discovery 1): Claims may derive from dataset analysis, literature review, or synthesis across claims and conclusions, and should be evaluated according to claim type.
- Evaluation Rubrics: Data-analysis evaluations require attempting replication in Python and judging whether the central claim is supported, refuted, or unsure.
- Evaluation Rubrics: Literature-review evaluations assess whether scientific literature supports or refutes the claim, marking obviously incorrect claims as refuted.
- Evaluation Rubrics: Evaluators provide a categorical answer and a free-response explanation backed by specific papers or evidence, with notebooks supplied for data-analysis assessments.
- Evaluation Rubrics: Interpretation evaluations judge whether a claim is a logically reasonable inference from the report and supplied context without conducting new data analysis.
Supplementary Information 4: Examples of Kosmos evaluations across statement types
The examples show that Kosmos evaluations can reproduce some quantitative claims, refute others, and distinguish supported mechanisms from overstated interpretations.
- The betaine–carnosine correlation claim was reproduced exactly after controlling for D-xylitol, including r=0.910, r=0.744, and p=0.149.
- The claimed strong negative PUFA-PC–carnosine correlation in normothermia was refuted because corrected analysis found a nonsignificant relationship.
- ASPN-mimic peptide attenuation of SMAD2/3 phosphorylation was supported across multiple breast-cancer and cardiac-tissue models.
- Carnosine’s transition-metal chelation was supported, but evidence did not support specialized bilayer localization or scavenging of non-aldehyde reactive species.
- Because both treatment groups received KOR+ activation, the hypothermia-specific betaine–carnosine coupling supports a temperature-dependent metabolic program rather than activation alone.
- The modest PUFA-PC increase in hypothermia was insufficiently supported as evidence for preserved membrane function or protective lipid-mediator production.