Virtual patients move AI beyond early discovery—toward the clinical decisions where drug-development value is won or lost
BEIJING, China — DeepoMe and The Future Laboratory at Tsinghua University today announced the release of a joint preprint, “Toward a Self-Learning AI Agent for Drug Repurposing: Building Human-Scale Representations for Virtual Patients.” The work explores a critically under-modeled frontier in AI for science: human-scale modeling—the representations needed to reason about how human biological states change under intervention.
AI is rapidly transforming how researchers model proteins, generate molecules and simulate cellular perturbations. But the central opportunity is no longer only to model molecules or cells; it is to model human response to intervention.
A recent Nature Reviews Drug Discovery Perspective makes the economic case: reducing failure by 20% at Phase II—the first major test of whether a medicine delivers meaningful benefit in people—can save nearly $900 million per successful drug launch, while the same improvement at early candidate selection has a comparatively small effect on total development cost. The article also finds that clinical programs using biomarkers to identify likely responders can reduce the average cost per successful launch to roughly half that of unstratified programs. The next frontier is therefore human-scale modeling: representing which biological states matter, which patients may respond and what evidence should guide the next intervention decision.
DeepoMe is building Human Response Intelligence: infrastructure for connecting measured human biological state, structured interventions and longitudinally observed response. The company is developing this work first in longevity science and aging-related disease—settings that combine repeated measurement, multidimensional biological state, modifiable interventions and cross-disease relevance.
Human Response Intelligence starts with defining state: determining which biological dimensions best describe a person for a particular decision. DeepoMe organizes these dimensions in a human-scale biological map. Modules are reusable, evidence-linked units of that map—such as aging processes, immune state, organ function, nutrition and exposure history. A representation is the decision-ready combination of modules selected for a particular disease, intervention and response question.
SteeraMed Bench is the evaluation and optimization engine for this map. It tests which modules—and which combinations of them—improve intervention reasoning, then selects and integrates the combinations that add the strongest distinct value into optimized maps for the decision at hand.
In the reported analyses, SteeraMed Bench integrates 332 modules spanning aging, organ systems, immunity, nutrition and food-derived interventions. It evaluates their utility across drug-repurposing tasks involving 1,916 small molecules and multiple disease contexts. The results indicate that no single biological map is universally optimal: different diseases require different combinations of biological coordinates. In other words, representation selection is itself a measurable driver of decision quality.
The framework also includes an LLM-assisted workflow for proposing and evaluating candidate modules against explicit utility and redundancy criteria. For example, the Hallmarks of Aging framework has expanded from nine hallmarks to twelve and now fourteen major hallmarks. Such scientific frameworks have traditionally evolved through expert synthesis of a growing literature. SteeraMed Bench adds a quantitative layer: it tests proposed hallmarks against explicit utility, coverage and redundancy criteria, helping scientists build stronger knowledge frameworks by distinguishing dimensions that improve decision quality from those that add little beyond existing representations.
This infrastructure provides the coordinate system for steerable biological world models. It enables models to reason over explicit, testable representations of human state, mechanism and intervention; combine general-purpose maps with task-specific maps; and improve those representations as new evidence accumulates.
“AI for Science has become remarkably good at seeing molecules and cells,” said Dr. Jianghui Xiong, Chief Scientist of DeepoMe. “But before we can ask how an intervention changes a person, we must define the biological state that matters for that person and that decision. Human Response Intelligence has two tasks: first, identify the dimensions that best represent human state; second, learn how that state may change under intervention. SteeraMed Bench makes the first task measurable by evaluating and optimizing modules into the maps that provide the strongest basis for intervention reasoning.”
For DeepoMe, longevity science and aging-related disease are the initial learning domain for this AI mission. These settings make it possible to observe human biology repeatedly, characterize state across multiple dimensions, document modifiable interventions and study response over time.
DeepoMe’s multidimensional aging-assessment work, including Capome, is an entry point for measuring human biological state longitudinally and interpretably. By linking repeated measurements with structured intervention exposure and follow-up, it creates the conditions to learn how—and for whom—an intervention changes that state.
For research and biopharma partners, DeepoMe is developing an auditable translational-decision workflow for prioritizing repurposing hypotheses, defining candidate responder populations and specifying the next validation experiment or longitudinal study.
For aging and longevity science, this infrastructure can move intervention reasoning from population-level associations toward individualized N-of-1 prediction and, ultimately, the generation of individualized intervention programs. It identifies the biological dimensions most relevant to a given person, prioritizes intervention hypotheses and pairs them with longitudinal follow-up to learn whether that person’s state changes as predicted.
The preprint is available at https://www.preprints.org/manuscript/202608.0998 (DOI: 10.20944/preprints202608.0998.v1). SteeraMed Bench can be explored at https://steeramed.com/bench.
About SteeraMed Bench
SteeraMed Bench is a research framework and evaluation-and-optimization engine for human-scale biological maps. It tests reusable, evidence-linked modules and their combinations to identify the representations that add the strongest distinct value for a particular intervention decision. By measuring utility, coverage, redundancy and incremental value in drug-repurposing and intervention-reasoning tasks, it produces optimized maps for hypothesis generation, validation planning and future Human Response Intelligence systems.
About DeepoMe
DeepoMe is a Beijing-based research and technology company building Human Response Intelligence: infrastructure for connecting measured human biological state, structured interventions and longitudinally observed response. The company develops these capabilities first in longevity science and aging-related disease, through work spanning multidimensional state measurement, interpretable biomedical representations, intervention reasoning and longitudinal learning. Its aim is to improve the quality, auditability and testability of biomedical and translational hypotheses.
Media Contact
DeepoMe Media Relations
Email: info@deepome.com
Website: https://deepome.com
