DoAtlas-2

At a glance

What DoAtlas-2 is built on

Figures as of 22 September 2026. Participant totals are source-reported and not deduplicated across resources.

What makes it different

It tests its own hypotheses, and the results decide what it studies next

Most AI-for-science systems stop once a hypothesis has been generated. DoAtlas-2 carries each one through to a test in human population data and keeps the outcome, whatever it is.

How it works

One research round, then the next

DoAtlas-2 keeps three things in step: the causal evidence it currently holds, the frontier of open questions, and the evidence it still needs. A round changes all three together.

1. AskA causal discovery agent reads the evidence graph for gaps, conflicts and unresolved paths, and proposes mechanism questions with the earlier results that motivated them.
2. Freeze the designEach selected question is turned into a study: population, measurement windows, estimand and identification conditions, fixed before outcomes are seen.
3. Test in human dataResearch agents write and check the analysis, run it on cohort and public data, and report every prespecified outcome with diagnostics and falsification checks.
4. ReviseA scientific learning agent compares outcomes with prior evidence and updates what is supported, where, and what to study next.

↻ The revised state is the starting point for the next round.

What each kind of result leads to

Data foundation

Populations, organ systems and molecular scales in one place

The data landscape brings together 771 research resources covering more than 720,000 participants in 48 countries. Measurements run from population surveys, longitudinal cohorts and clinical records to continuous physiological signals, medical imaging, and eight molecular layers: genomics, epigenomics, transcriptomics, proteomics, metabolomics, lipidomics, microbiomics and immunology assays.

The Human Phenotype Project (HPP) is the core longitudinal cohort. It links long-term follow-up, deep phenotyping and multi-omics at the participant level, and it is where candidate mechanisms are tested and the evidence state is revised.

Inside the Human Phenotype Project

Evidence for the loop

What three rounds of research produced

From a research space of 8,724 candidate questions, DoAtlas-2 has evaluated 2,031 end to end, producing 88,731 result estimates. In the Human Phenotype Project the catalog of pathway questions grew from 175 in round one to 1,079 after round two and 4,014 by round three (3,941 distinct pathways), across vascular, early-glycemic and hepatic-metabolic systems. Of the first 1,079 questions screened, 756 received statistical support.

Key findings

Blood pressure links body fat, the liver and lipids to the arteries

Representative analyses cover 15 pathways with complete serial estimates, forming a cross-system network of 19 phenotypes and 29 connections. Three results stand out.

1. Blood pressure is a convergence node

Waist-to-hip ratiobody-fat distribution →ALTliver enzyme →Diastolic blood pressureintermediate →Pulse-wave velocityarterial stiffness

Liver and lipid branches from adiposity converge on diastolic blood pressure and lower-limb pulse-wave velocity. The standardized path association for the pathway above is 0.01073 (95% CI 0.007575 to 0.01389). In an independent cohort of 912 adults, metabolic-to-pressure routes to brachial–ankle pulse-wave velocity replicate (BMI–triglycerides–diastolic pressure, GGT–triglycerides–systolic pressure, ALT–diastolic pressure and triglycerides–diastolic pressure). In a second cohort of 879 adults, all serial estimates share the same sign.

2. Liver and lipid markers come first

ALT → diastolic pressure → pulse-wave velocity and triglycerides → diastolic pressure → pulse-wave velocity are both supported under strict temporal ordering, consistent with a metabolic-to-hemodynamic sequence. Among ApoB-related sequences, visceral fat → ApoB → systolic pressure → pulse-wave velocity differs by sex, and weight → ApoB → diastolic pressure → pulse-wave velocity is weaker at older ages.

3. One pathway mostly dissolves under adjustment

The waist-to-hip ratio → inflammation (GlycA) → systolic pressure → pulse-wave velocity association falls from 0.01218 to 0.002022 after joint adjustment for BMI and smoking. The pathway reflects overall adiposity and smoking more than fat distribution: a result that narrows a mechanism instead of extending it.

A fourth example shows why the conditioning set matters: higher nocturnal oxygen saturation goes with lower systolic pressure, until waist circumference is in the model, when the association turns weakly positive. Waist circumference is a shared determinant of both.

Why causal

The discovered network is the predictive model

Black-box models give answers without reasons. DoAtlas-2 keeps a traceable causal path and the evidence behind every claim. The vascular network above is used directly as a predictive model: every prediction is an explicit closed-form expression, and every term belongs to a named variable or a source relationship. We see this as a way beyond black-box Transformer and agent-based foundation models, toward causally interpretable AI.

For researchers

How DoAtlas-2 supports research

DoAtlas-2 is meant to work alongside researchers, not replace their judgment. It takes on the part of the research cycle that scales poorly by hand: enumerating what could be tested, fixing how it will be tested, running the analyses, and keeping an honest record of what happened.

Who it is for

• Epidemiologists and clinical researchers: candidate mechanisms with designs and robustness checks to take into your own cohort for replication or refutation.
• Causal inference and statistics researchers: a large, fully recorded set of designed analyses to audit, benchmark methods on, and improve.
• AI-for-science researchers: an end-to-end reference for evaluating discovery systems on whether their hypotheses survive contact with data, not only on whether they sound plausible.
• Cohort and biobank teams: a view of which measurements, populations and follow-up would most change current scientific judgment.

Work with us

We welcome collaborations, in particular replication in independent cohorts, new data resources and methods review. Write to imran.razzak@mbzuai.ac.ae. If you would like to join the team, see the pages for prospective students and research positions.

Open source

Release status

DoAtlas will be open source and available to researchers soon. The first release includes 1,000 research questions with 30,000 corresponding experimental results. Links will be added here when it is published.

Reading the results responsibly

• Estimates are observational. Pathways are candidate mechanisms tested against human population data, not established causes or clinical guidance.
• Participant totals are as reported by each source and are not deduplicated across resources.
• Screening-level support is unadjusted. A relation's state is set only by primary records from full, prespecified designs.
• The interpretable predictive model reported so far is the first, disease-specific one, built on the vascular network.

Technical report

DoAtlas-2: A Foundation for Self-Evolving Causal Biomedical Discovery

Yulong Li*, Rong Xia*, Yuxuan Zhang*, Jianxu Chen, Xiwei Liu, Haochen Xue, Maosheng Li, Yuhang Liu, Yibo Yuan, Yutong Xie, Chong Li, Jionglong Su, Hagai Rossman, Eran Segal†, Imran Razzak†

DoAtlas Team · MBZUAI. * Equal contribution. † Corresponding authors. arXiv:2609.35107, submitted 28 September 2026.

Read on arXivPDFDoAtlas Family

@article{li2026doatlas2,
  title   = {DoAtlas-2: A Foundation for Self-Evolving Causal Biomedical Discovery},
  author  = {Li, Yulong and Xia, Rong and Zhang, Yuxuan and Chen, Jianxu and Liu, Xiwei and Xue, Haochen and Li, Maosheng and Liu, Yuhang and Yuan, Yibo and Xie, Yutong and Li, Chong and Su, Jionglong and Rossman, Hagai and Segal, Eran and Razzak, Imran},
  journal = {arXiv preprint arXiv:2609.35107},
  year    = {2026}
}

Updates on what DoAtlas is discovering are coming. Stay tuned.