01PhAI Labs · Toward real scientific discovery
Bring frontier questions into the loop of intelligence
Connect real scientific questions, data and experimental feedback with next-generation AI—and explore how AI can move from assisting research to participating in discovery.
The first cohort prioritizes life science and biomedicine, while remaining open to chemistry, materials and other frontier fields.
02Why this program
AI can already retrieve literature, generate code, process data and assist with parts of the research workflow. But discovery often begins with a question that is not yet well defined—and advances through hypotheses, evidence, tools, experiments and continuous revision.
That process cannot happen only on static benchmarks or closed tasks. AI must enter real research environments, confront incomplete data and constraints, and learn from experiments and reality.
Begin with the unknowns, bottlenecks and anomalies scientists actually face.
Subject model outputs to data, experiments, expert judgement and the physical world.
Turn questions, tools, data and verifiers into reusable infrastructure for the next discovery.
Scientists bring real problems, data and experimental feedback; PhAI brings models, tools and analysis; results are tested together in real experiments and go back to the scientist.
Move the cursor over the drawing
03What is DFM
DFM (Discovery Foundation Models) is neither a single model nor a software framework that forces every system through one pipeline. It is the research paradigm, capability definition and system organisation that PhAI Labs proposes around Discovery Intelligence: through formal definitions, process diagrams, system instances and real scientific cases, it turns “discovery” from a vague notion into a capability of model systems that can be learned, executed and evaluated.
The DFM Scientist Collaboration Program brings real scientific problems, data and verification conditions into this paradigm, so systems form their judgement in real research.
Three independent research and engineering directions
ScienceBuddy, ScienceIDE and JEPA-Anything are three fully independent modules and research directions: independent in product form, usage, papers and implementation, with no required calls or dependencies between them. Looking ahead, the three could form a composable research path: ScienceBuddy collects feedback and research trajectories from the interaction between scientists and agents; ScienceIDE organises that data and experience into reusable scientific environments for task execution, reinforcement learning and model training; JEPA-Anything explores unified state prediction and simulation across multiple scientific environments and interventions. That is a possible future combination, not an integration that exists today, and it does not change their independent standing.
An interactive Scientific Agent, exploring RSI in RSI in real research interaction
It focuses on an interactive form of Science RSI: a user poses a scientific question, the agent calls tools and answers; the scientist's follow-ups, edits, acceptances, rejections and re-runs become feedback that drives the research agent's workspace and harness to keep evolving.
Open ScienceBuddyReleasing 16 SeptA platform integrating scientific environments, code, tools and training pipelines
It turns real scientific code, research procedures, task goals and verification criteria into scientific environments that run, measure and reproduce, so that researchers can do agent execution, model training and reinforcement-learning research in real scientific settings.
ScienceIDE project pageReleasing 17 SeptA unified science world model for cross-domain state prediction, intervention simulation and scientific diagnosis
It explores how to learn, across vision, biology, the clinic, molecules, control and physics, latent representations that predict changes of world state, and to offer one interface for intervention prediction, state synthesis and scientific diagnosis.
JEPA-Anything paperReleasing 18 Sept04What we look for
We are most interested in questions with real scientific value, a consequential unknown, and a path to data or experimental feedback—not tasks already fully specified and solved by an existing workflow.
A question that could explain an unresolved phenomenon, reveal a new mechanism or pattern, or lead to a new way of doing research.
Current theory, models, data analysis or experiments do not provide a satisfying answer—and the failure itself is informative.
Some combination of data, literature, tools, experimental conditions or domain knowledge makes a first exploration possible.
A hypothesis can be supported, corrected or rejected through data, experiments, expert judgement or the real world.
Questions to begin with
05First cohort & who should join
The program is open to researchers in universities, institutes, laboratories and industrial R&D teams. Questions may be at the stage of early conception, data accumulation, methodological bottleneck or experimental validation.
These fields combine rich multimodal data, specialized tools and experimental verification with a large number of open, consequential scientific questions.
The collaborators we hope to explore with
You know which questions matter and where current methods stop, and you want foundation models, research agents or automated experimentation wired into how you already work. We start from the specific bottleneck, work out where AI actually helps, then design and test the approach together.
You have specialist data, research code, computational models or simulation environments, and want AI to extend how you analyse and explore. We look at how AI can run verifiable research tasks inside them.
You run experiments, collect observations and test hypotheses, and want to know whether AI can help design the experiment, read the result or choose the next direction. We close the loop from question to evidence together.
Researchers who want AI deep in the research process and are willing to work out a new paradigm with us. If you have a question you keep returning to, relevant groundwork, or a direction you want to explore together, tell us about it.
You are working on the science behind a product or a process, with business context, data or test conditions already in hand, but current methods cannot explain the key behaviour, or it takes too many trials to find something that works. We identify where AI fits and measure progress by computation, simulation or experiment, against your real requirements.
06Collaboration
Data & intellectual propertyBefore substantive work begins, the parties will document data scope, confidentiality, publication, authorship and IP arrangements for the specific project.
07Exploration cases
The first seed users are bringing real research questions into testing. With their permission, we will share representative exploration processes, intermediate outcomes and lessons.
Rather than display only a final answer, we want to show how a problem was framed, where AI participated, how scientists judged the output, and how the research moved forward.
Autonomous therapeutic discovery by an embodied AI scientist
Generalizable Agentic Laboratory Intelligence for Learning, Experimentation, and Optimization
GALILEO is the DFM idea made concrete in the life sciences: AI is no longer a tool that completes a single task but enters a real research environment, starts from a real question, keeps learning from experimental feedback and takes part in the whole process of discovery. The work is available as a bioRxiv preprint (June 2026).
Most AI systems in biomedicine stop at prediction: predicting targets, predicting molecular properties, without reaching real experimental intervention. GALILEO asks whether AI can move past prediction and close the loop autonomously in a dynamic membrane system, from candidate peptide discovery to wet-lab validation, and find therapeutic targets and molecules that intervene in tumour immunity.
GALILEO runs on an OTAS reasoning loop (Observation, Thought, Action, Summary) and executes autonomously:
The DFM viewGALILEO shows what DFM argues for: moving from completing research tasks to taking part in discovery. AI enters a real research environment, meets the questions scientists actually care about, works with incomplete data and real constraints, and keeps learning from experiments and reality, so each exploration becomes the starting point of the next discovery.
08Collaboration process
Describe the scientific question, current progress, key bottlenecks, and available data or verification.
Assess scientific value, the role for AI, verification feasibility and mutual resource fit.
Clarify the scientific context, collaboration goals and current research foundation together.
Define goals, workflow, contributions, data use and stage-gate criteria together.
Move into early validation or deeper joint research and iterate through real feedback.
09Frequently asked questions
We prioritize consequential scientific unknowns where AI can make a substantive contribution and where data, experiments or expert feedback can support verification. The question need not be fully formalized, but its importance and a viable starting point should be clear.
Yes. Identifying a valuable unknown and turning it into a researchable problem is itself part of the DFM agenda. Share the observation, the limits of current methods, and the evidence or conditions already available.
We prefer to begin with a bounded subproblem and test the problem formulation, data conditions, role for AI and verification path before considering deeper joint research. Scope and pace are co-designed.
We consider scientific importance, whether DFM or the Research Agent can create genuine research leverage, the feasibility of data and verification, and whether both sides can commit the right research and engineering resources.
Outcomes may include a validated hypothesis, reusable research workflow, tool or data asset, model or Agent improvements, and where appropriate, papers, open-source releases or other joint outputs.
Do not upload patient-identifiable, clinically sensitive, third-party restricted or otherwise unauthorized or confidential material. Appropriate data exchange and protection can be discussed later.
10Apply now
You do not need a complete project proposal. Tell us what you want to understand, why it matters, where the current work is blocked, and what data, tools or verification conditions already exist.
Apply now Feishu form
Please do not upload patient-identifiable, clinically sensitive, third-party restricted or otherwise unauthorized material in the initial submission.