Working draft

Propel

Propel would give scientists at a research institution six months to pursue bold biological questions with AI research agents. The program would support their work and turn what they learn into tools and worked examples, so future participants and other scientists could go further.

Working draft for discussion

November 9, 2026

The proposal was developed through discussions among Jackson Weir, Luca Pinello, Matteo Di Bernardo, Yasha Ektefaie, Arya Rao, and Shantanu Singh in 2026. They explored how a program could support scientists pursuing biological questions with AI research agents and share what they learned.

The draft is shared here for discussion. It does not announce an approved institutional initiative or solicit funding.

A new instrument for discovery

AI systems now prove theorems, design molecules that work in the wet lab, and carry out analyses that once took experts weeks. The same capability reaches into every part of scientific work, from the literature to data, code, and the tools of the laboratory.1 Propel would put scientists in charge of that new capability, directing it toward questions that matter to their own work. Their experience helps them recognize promising results and judge whether they hold up2, so they can build and direct work that was previously beyond their capacity.

Propel would begin with a small group given protected time over six months to pursue that work. Nothing here waits on a future breakthrough. Models already follow instructions and use scientific tools, and they improve faster than the institutions around them can adjust. The program would use what researchers learn from the first projects to build tools that help other scientists take on more ambitious questions.

Examples of research agents in use

The scientists who helped develop this proposal include inventors of the methods they use, who are now skilled at directing research agents. Their understanding of the data comes from years of working with it, often from generating it themselves. For scientists already working with research agents, results like these are becoming familiar.

A scientist spent two years studying melanoma with spatial genomics. He wanted to understand how neighboring cells help cancer cells evade immune attack.3 He gave an agent his notes and asked it to distill the analysis procedures and choices from that work. On a new dataset from 25 newly collected colorectal cancer samples, it prepared the data and ran the starting analyses in about a day.

A computational biologist leading a regulatory genomics lab has made AI agents part of how his lab uses and develops models that predict what a stretch of DNA does in a cell. With collaborators, he built an open source framework that lets agents run leading models from plain-language questions.4 Working hands-on with Claude, he also took a new model through implementation, computational testing, and a complete manuscript in about five months. This single model predicts regulatory activity across cell types, designs DNA for a chosen activity, and makes the DNA patterns behind both inspectable.

A postdoc coming from outside medicinal chemistry co-authored the study introducing Fleming, an agent for designing tuberculosis antibiotics. It combined molecular models with literature searches to design and refine candidates, weighing predicted activity, toxicity, and other drug properties. His collaborators synthesized six compounds from its designs, and all six inhibited the pathogen in the wet lab.5

The scientists were pursuing these projects months ago, with models less capable than those available today. If that is the new floor, where does the real work begin? The proposed first Propel projects would explore that question by tackling biological puzzles and testing how much more of the work scientists can entrust to agents.

The first questions

The following questions and project descriptions reflect the work we discussed while developing the proposal.

Spatial genomics Jackson Weir

How do tumors recruit immune cells, and can patterns in spatial data lead to a mechanism we can test?

Sources

Liver toxicity Shantanu Singh

Which cellular phenotypes signal liver injury, and can we detect them reliably in images?

Sources

Interview. Shantanu Singh on why this question.

Transcript

A medicine that is meant to treat a disease can also do harm, and designing a drug that does only the one thing you want is really hard. The organ hit the hardest is the liver, where drugs and most of what we're exposed to get metabolized. Now, today we test mostly in animals, which is expensive, it's slow, it's ethically fraught, and often a poor predictor of what happens in people.

In OASIS, we treat liver cells in a dish with hundreds of compounds and measure gene expression, proteins, and the detailed appearance of the cells using Cell Painting, a method our lab invented. We want to predict toxicity, but also understand what's happening inside the cell.

Cell Painting gives us a broad fingerprint of the cell, but one thing that has really bothered me is how hard it is to go back and understand what is driving it.

So here's what I want to try. Instead of the broad fingerprint, we'll measure specific phenotypes that biologists already link to liver injuries, such as fat building up inside liver cells. And that's a bias, since profiling is meant to be broad, but every readout would have a direct biological basis. With AI agents, we can build those measurements, I think, relatively quickly, where each one used to otherwise take a really long time. Then we find the compounds that are toxic in those ways and test them in a follow-up assay for that specific kind of toxicity.

Regulatory DNA across cell types Luca Pinello

Which DNA patterns control gene activity across cell types, and can we use them to design sequences for experimental testing?

Sources

Optical pooled screening Matteo Di Bernardo

Can an agent learn from an expert's records to automate the analysis of an optical pooled screen?

Sources

Rapid experimental loops Yasha Ektefaie

Can a scientist train an agent to carry out wet-lab work, using experimental results to pursue a question over repeated rounds?

Human adaptation Arya Rao

Which genetic variants drove recent human adaptation, and what can they teach us about the biological pressures that shaped our species?

Sources

Proposed program design

Participants and projects. Propel would explore what experienced scientists can accomplish with agents and protected research time. The proposal was developed around projects discussed with Jackson Weir, Shantanu Singh, Luca Pinello, Matteo Di Bernardo, Yasha Ektefaie, and Arya Rao. Participants would spend six months pursuing bold biological questions or developing new ways to use agents in experimental work. The six-month window is deliberately aggressive: research agents make it plausible to attempt projects that researchers would otherwise expect to take substantially longer. The projects would start from the researchers' own work and focus at first on data that are public or that they intend to make public. The scientists involved in developing the proposal are already highly skilled with research agents, including coding agents such as Claude Code and Codex equipped with scientific tools.

The project plan. Each participant would set out the question they would investigate, the result they would aim to obtain in six months, and how they would test it.6

The work. The proposed program would provide seed funding for each project. Researchers would decide how to organize and carry out the work, meeting only when useful. They should be free to change direction as they learn. The process need not be clean; the goal is to make discoveries and advance the science.

Participants would record the work so others could build on it. Research-agent sessions would capture much of the computational work automatically, including the scientist's guidance as the investigation develops.7 Participants would also keep records of their bench work and maintain a project repository. A small dedicated team would work alongside the researchers to address technical barriers and study what advances the science. The team would learn from the scientists' work and build tools that help them pursue new questions and approaches.

The outcome. After six months, each researcher would report the findings and explain how they would pursue the remaining scientific questions. Some projects might produce a computational result whose experimental test follows later. The larger scientific question may remain open.

Success in the first round. Each project would be judged against its recorded six-month result and validation test. For the round as a whole, the proposed measures of success are:

  • Science. Most projects reach their planned six-month result, and at least one produces a finding with an experiment underway to test it. Some projects falling short is expected.
  • Reach. At least one scientist outside the first group uses a worked investigation or tool from Propel to start their own work.
  • Learning. The team gives a clear account of what held the work back and fixes some of those barriers, so the next round is better set up.

Promising projects could receive support for experimental validation or another round of work.

What Propel would share

Scientific work rarely follows a path laid out in advance. It is more like an ocean crossing: only afterward do we see its wake as "a brief, glimmering trace on the waters."8

Research-agent logs let us reconstruct more of that scientific journey so other scientists can build on it.

Working with the scientists who made the decisions, Propel would turn that experience into two kinds of resources to help others pursue discoveries of their own. First, a library of worked scientific investigations. These would bring together the questions, analyses, code, and key decision points, including what failed and how results were checked, so other scientists and their agents could actually learn from and adapt them9. Second, reusable tools that address barriers these investigations run into, including tools that make scientific data easier for agents to access and analyze10. Where investigations demonstrate a need, these tools could extend to models, methods, or science sandboxes11.

Propel's team would build tools needed for its projects.

Each project would contribute reviewed and redacted session records for this analysis.1213 Propel would share the findings alongside reusable tools and selected records that let others follow the work. The team would also convene discussions with other scientists to help them apply and adapt lessons from the projects.

Perhaps the greatest lesson is simply knowing that a difficult thing can be done.14 Scientists are using research agents to pursue ideas they once set aside for lack of time or specialist help. Propel would show what scientists furthest along with research agents achieve, and how, so others could see which crossings are now within reach and begin their own.

Illustrative budget

The original planning budget assumed approximately $1.5 million in philanthropic support for the first round, including institutional costs, plus $300,000 for initial experimental work.15 All figures are preliminary planning estimates. Support would fund protected research time, follow-on experiments, and the dedicated team working alongside the researchers and developing shared tools and resources.

The proposed timetable allowed for concise written results after the six-month projects, followed by a synthesis and shared materials within nine months of launch.

  1. Research agents can serve as apprentices: collaborators that can take on increasingly complex scientific work across literature, data, code, and experimental tools while scientific judgment remains with the scientist. Coding agents such as Earendil's Pi, OpenAI's Codex, and Anthropic's Claude Code can serve as research agents when equipped with scientific tools and access to data. Other examples include Anthropic's Claude Science and more autonomous systems such as Edison Scientific's Kosmos and Google DeepMind's Co-Scientist. Scientific examples also include Coscientist, which carried out chemistry experiments through an automated laboratory, and Virtual Lab, which designed nanobodies under human guidance. Bunne and Regev discuss a broader role for research agents in connecting models of cells with experimental investigation: "Beyond Representation: AI in Cellular Discovery", Daedalus (2026). The apprentice framing draws on Licklider's vision of people and computers developing ideas together, with people setting goals and evaluating the results: "Man-Computer Symbiosis", IRE Transactions on Human Factors in Electronics (1960). Autor and Manyika make the case for designing AI to work with human expertise and strengthen it: "A Better Way to Think About AI", The Atlantic (2025).↩︎

  2. Greka emphasizes the scientific intuition needed to frame meaningful biological questions and validate AI-assisted findings: "Building the Drug Discovery Engine of the Future with AI-Empowered Nodal Biology", Daedalus (2026). Koller argues that faster agent-driven experimentation depends on choosing objectives that reflect the biology of human disease: "Drug Discovery Has No Magic Wands", Deep Phenotype (2026).↩︎

  3. Weir et al., "A bone repair external state program drives cancer cell dedifferentiation and immune evasion in melanoma", Cancer Research (2026) [conference abstract].↩︎

  4. Chorus. Sequence-to-function models predict molecular activity, such as chromatin accessibility, transcription factor binding, or gene expression, directly from DNA sequence. They help scientists investigate the effects of genetic variants, but using several models together often requires substantial computational expertise. Chorus brings leading models, including AlphaGenome, Enformer, Borzoi, and ChromBPNet, behind one interface that an AI agent can operate. A researcher asks a question in plain language, such as what a disease-associated variant might do in liver cells. The agent selects appropriate models, runs them, and compares their predictions. Worked examples show how analyses that previously required days of assembling separate pipelines can be completed in an afternoon through conversation. The predictions provide a starting point for experimental validation and depend on the cell types and assays the models cover. Penzar et al., "Chorus: chatting with genomic oracles", Genomics x AI (2026).↩︎

  5. Wei et al., "Fleming: An AI Agent for Antibiotic Design for Mycobacterium tuberculosis", bioRxiv (2026). In a separate screening evaluation, the authors report an 83 percent in vitro hit rate among compounds the model predicted to be active, compared with typical random-screening hit rates of 1 to 5 percent.↩︎

  6. Before beginning work, each project would record its expected six-month result and validation test, together with any experimental validation expected after the six-month projects. Final reports would compare the findings with these expectations and distinguish validated results from work still awaiting validation. Bench records would capture each round's question, action, result, and next decision. Each project would maintain a GitHub repository for code, notebooks, and outputs.↩︎

  7. Logs may also help improve how agents work. Meta-Harness uses prior execution traces and evaluation results to revise an agent's harness, the code that manages its context and interaction with tools. See Lee et al., "Meta-Harness: End-to-End Optimization of Model Harnesses", arXiv (2026). Zhang and Khattab argue that harnesses can help models tackle unfamiliar problems by breaking them into familiar parts: "Language Model Harnesses Are Compositional Generalizers" (2026). These approaches suggest ways to turn project experience into better tools; their value for Propel's scientific work would need to be tested.↩︎

  8. Whyte, "Ambition," Consolations (2015): "A life's work is not a series of stepping stones onto which we calmly place our feet, but more like an ocean crossing, where there is no path, only a heading, a direction, in conversation with the elements. Looking back, we see the wake we have left as only a brief, glimmering trace on the waters." Whyte contrasts ambition fixed on a destination with a calling that changes us through the work. He closes with generosity, passing on to others "the core artistry that made the journey a journey."↩︎

  9. One simple example comes from planning experiments for MelanomaBox, discussed above. A researcher needed to choose a concentration range for each drug. He pointed Codex to an earlier expert discussion about whether reported potency values described effects on a drug's target or on cell survival. Codex built a script that combines evidence from public databases and papers to propose concentration ranges, preserving what each assay measured. The saved session logs record the questions and corrections that shaped the analysis. Another scientist could give Codex those logs and the script alongside a new compound list and experimental design, then ask which parts apply and what needs to change. Propel could compare the quality of the experimental choices and the effort required with those of the scientist's usual approach. Propel would study this pattern of adapting earlier work as questions and evidence change, including in more complex investigations.↩︎

  10. The projects may reveal where agents need better ways to access and use existing scientific data resources. A recent evaluation of NCBI Virus shows how much a better interface can change agent performance: Luebbert, "Paving the way for agents in biology", Anthropic (2026). Through the human interface, agent accuracy ranged from 16.9% to 91.3% and varied across runs; with gget virus, every evaluated agent exceeded 90%, peaking at 99.7%, and run-to-run variability largely disappeared. Illustrative targets that would require discussion with resource owners: gnomAD (GraphQL), the Connectivity Map / L1000 (CLUE REST), DepMap (Breadbox API), the GeneGenie results API (REST; includes FinnGen R13, UK Biobank, MVP, and other cohorts), the GTEx Portal (REST), and file-based resources such as the JUMP Cell Painting data, MSigDB, and the PROSPECT tuberculosis chemical-genetics data. Depending on the resource, the right interface may be an API, a lightweight data file, or a reusable skill.↩︎

  11. Some difficulties encountered in these projects could become science sandboxes: controlled settings where agents investigate a system through repeated experiments, recording how their explanations change with the evidence. These would let us test whether better tools or guidance help agents overcome a particular difficulty. See Rao et al., "Science sandboxes measure the scientific capability of AI agents", arXiv (2026), and the accompanying blog post, Rao et al., "Measuring the frontier of scientific capability in the age of AI", Substack (2026).↩︎

  12. Providing Propel-related session records after appropriate review and redaction would be a condition of participation. Sessions containing unrelated work would be separated or excluded, and raw records would remain private under agreed institutional arrangements. The research engineer would develop a lightweight method for preparing readable records for analysis and sharing. Public materials would include reviewed project reports, a synthesis across projects, reusable code and guidance, and selected session examples with sensitive and unrelated material removed.↩︎

  13. Recent agent benchmarks find that episodic records can preserve useful information lost when agents distill experience into general guidance. See Zhang et al., "Useful Memories Become Faulty When Continuously Updated by LLMs", arXiv (2026), and Lei et al., "SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills", arXiv (2026). Their results do not yet establish which will work best for scientific investigations.↩︎

  14. Roger Bannister ran the first sub-four-minute mile on 6 May 1954. John Landy surpassed his record 46 days later. World Athletics, "Fab five: mile races" (2019).↩︎

  15. Illustrative budget. The original planning budget assumed approximately USD 1.5 million in philanthropic support, including institutional costs, plus USD 300,000 for initial experimental work across six projects. All figures are preliminary planning estimates. Approximately USD 800,000 of the philanthropic support would fund the work directly: USD 180,000 for protected research time, USD 200,000 for the central team, USD 60,000 for compute, model access, and storage, and USD 360,000 for follow-on experiments. The remaining approximately USD 700,000 is a provisional allowance for institutional costs. The first round would use existing laboratories and shared facilities, subject to agreement with their owners. Additional equipment needs would be specified in individual project budgets. Participants would devote between a quarter and half of their time to Propel during the six-month projects. The central-team estimate covers a full-time research engineer for nine months, an independent senior scientist at 10 percent effort over the same period, and coordination and preparation of shared materials. The engineer would work alongside participants to address technical barriers and develop reusable tools. Drawing on project records and conversations with the scientists, the team would examine what advanced the science and what held it back. Participants would check the resulting accounts for accuracy, and the senior scientist would review the methods and conclusions. The findings would guide the team's work in subsequent rounds. Follow-on funding would be allocated after the six-month projects, based on the results and plans for further experiments. The proposed timetable allowed for project findings after six months and a synthesis and shared materials within nine months of launch. Under the proposed governance, scientific decisions and publication would remain with the researchers and their institution. The planning estimates, including salaries, benefits, and institutional costs, would require further review before any program could proceed. Technology partnerships could contribute compute and model access.↩︎

  16. Luebbert, "Paving the way for agents in biology", Anthropic (2026). The gget virus tool provides a deterministic retrieval interface for NCBI Virus, improving accuracy and consistency when agents construct viral sequence datasets.↩︎

  17. MelanomaBox connects agent-proposed drug combinations to prospective experiments in melanoma cells and returns measured cell viability as feedback. Yasha Ektefaie is among the project leads.↩︎

  18. OttoSeq and Brieflow provide existing laboratory and computational workflows on which this effort can build. Kirby et al., "A one-week automated genome-wide optical pooled screen using OttoSeq", Genome Biology (2026). Di Bernardo et al., "Brieflow: An Integrated Computational Pipeline for High-Throughput Analysis of Optical Pooled Screening Data", bioRxiv (2025).↩︎

  19. Acknowledgments. The proposal was developed together with Jackson Weir, Yasha Ektefaie, Shantanu Singh, Matteo Di Bernardo, Luca Pinello, and Arya Rao. On the framing: Nodar Gogoberidze, Alan Munoz, Arya Rao, Eric Lander, Johan Haslum, Jess Ewald, David Dao, Jackson Weir, Yasha Ektefaie, and Zeming Lin. For thinking out loud: Aleksandr Zimin, Alex Kalinin, Ank Kumar, Anna Greka, Anne Carpenter, Aviv Regev, Beth Cimini, Blake Lash, Brendan Saloner, Bryan Hsu, Carolina Wahlby, Daphne Koller, David Autor, Deb Hung, Enoch Huang, Iain Cheeseman, Jan De Boer, Jason Swedlow, Jesse Boehm, Jessika Baral, John Arevalo, Jonne Rietdijk, Luca Pinello, Matteo Di Bernardo, Misha Belkin, Nabiha Saklayen, Nelson Repenning, Pardis Sabeti, Paul Blainey, Puneet Batra, Ramnik Xavier, Roby Bhattacharya, Sandeep Kambhampati, Soumya Raychaudhuri, Tim Treis, Wei Ouyang, Wolfgang Huber, and Yajit Jain.↩︎