SteeraMed Bench evaluates and optimizes human-scale biological maps for virtual patients—making representation selection a measurable driver of drug-repurposing decision quality.

tags: steeramed, ai4science, virtualpatients, drugdiscovery

Beyond Virtual Cells: SteeraMed Bench Gives AI the Map It Needs for Virtual Patients

AI can now predict protein structures, simulate cellular perturbations, and design molecules de novo. But ask it the question that actually determines whether a drug succeeds or fails—"How will this person's biology change under this intervention?"—and today's models go silent.

Molecules can be modeled. Cells can be simulated. But at the level of the person—precisely where drug-development value is won or lost—the infrastructure for systematic modeling remains absent.

This gap is what DeepoMe and The Future Laboratory at Tsinghua University are addressing with SteeraMed Bench, an evaluation and optimization engine for human-scale biological maps, released alongside a joint research paper on Preprints.org (ID 202608.0998, DOI: 10.20944/preprints202608.0998.v1).

The $900 Million Question: Why Phase II Is Where Value Lives

A recent Nature Reviews Drug Discovery Perspective lays out the economics: reducing failure by 20% at Phase II—the first test of whether a medicine delivers meaningful benefit in people—can save nearly $900 million per successful drug launch. The same improvement at early candidate selection has a comparatively small effect.

The article also finds that clinical programs using biomarkers to identify likely responders can cut the average cost per successful launch to roughly half that of unstratified programs.

In other words, the big money isn't lost in early "pick the molecule" decisions—it's lost at the human response stage. DeepoMe's positioning is clear: "We optimize the Phase II success-or-failure juncture—each successfully launched drug corresponds to nearly $900 million in value."

SteeraMed Bench: Context and Positioning

SteeraMed—a Steerable World Model framework for biomedicine—was first introduced by DeepoMe Chief Scientist Dr. Jianghui Xiong at the Beijing Zhiyuan Conference (BAAI). SteeraMed Bench is a new module within that framework: an evaluation-and-optimization engine that answers the question, "Which set of biological coordinates is most useful for the decision at hand?"

The Core Logic in Three Sentences

DeepoMe is building Human Response Intelligence, and its core logic can be stated in three steps: Use a steerable world model to compute virtual patients—predict whether a person will respond to an intervention. After the intervention, bring them back for retesting—whether the prediction was right is immediately verifiable. And this maps directly onto the most expensive, most valuable cut in drug development—Phase II clinical trials (each successfully launched drug corresponds to nearly $900 million; responder stratification can cut the cost per successful launch to approximately half).

Defining State: The Human-Scale Biological Map

Before you can predict how someone will respond, you need to define which biological dimensions matter for that person and that decision. DeepoMe organizes these into a human-scale biological map with two levels:

• Modules: Reusable, evidence-linked units of the map—aging processes, immune state, organ function, nutrition, exposure history, and more.
• Representations: Decision-ready combinations of modules selected for a particular disease, intervention, and response question.

A world model without an explicit map is like a navigator without a road network—no matter how good the rendering, it can't compute "from this state, under this intervention, where do I arrive?"

Reasoning over explicit, testable representations of human state, mechanism, and intervention is what makes a Biomedical World Model genuinely steerable.

SteeraMed Bench: Methodology and Key Results

Once you have a map, the methodological question follows: among all modules and their combinations, which map is most useful for the current decision?

SteeraMed Bench provides a measurable answer: it tests which modules and combinations improve intervention reasoning, then selects and integrates the strongest-performing combinations into optimized maps.

In the reported analyses:

Parameter	Value
Total modules	332 (aging hallmarks, organ systems, immunity, nutrition, food-derived interventions, TCM syndrome proxies, food-as-medicine targets)
Small molecules	1,916 DrugBank compounds
Disease tasks	5 chronic diseases, extrapolated to 23 disease categories
Core metric	recall\@20 (proportion of known drug-disease pairs captured in top-20 ranked candidates)

Three core results emerged:

1. No single biological map is universally optimal. The full 332-module atlas is strictly optimal in only 9 of 23 disease categories. Different diseases have different best-fitting coordinate combinations: extended aging hallmarks lead in type 2 diabetes and osteoporosis; food-as-medicine combinations are strongest for depression; nutrient modules dominate atherosclerosis/hyperlipidemia.

2. Representation selection is a measurable driver of decision quality. Just 117 nutrient and extended modules achieve an average recall\@20 of 0.494 across five diseases—already approaching the full 332-module atlas at 0.524. This fact was previously judged mostly by expert intuition and default single-atlas assumptions.

3. Negative results are reported honestly. LLM-proposed candidate modules showed no significant improvement after multiple-testing correction; leave-one-disease-out validation performed near random. This is methodologically important: it demonstrates the benchmark does not stack the deck for any particular approach.

Quantitative Stress-Testing of Knowledge Frameworks

The framework also includes an LLM-assisted workflow for proposing and evaluating candidate modules against explicit utility and redundancy criteria.

Consider the Hallmarks of Aging framework: it has expanded from 9 to 12 and now to 14 major hallmarks. Such scientific frameworks traditionally evolve through expert synthesis of growing literature. SteeraMed Bench adds a quantitative layer—it tests whether a proposed hallmark is a genuinely independent dimension that improves decision quality, or merely redundant with existing representations.

For researchers, this is a new instrument: a ruler for scoring knowledge frameworks.

# Conceptual: How SteeraMed Bench evaluates a candidate module
def evaluate_module(candidate_module, existing_modules, task):
    """
    Test whether adding candidate_module improves decision quality
    beyond what existing_modules already provide.
    """
    baseline_recall = compute_recall(existing_modules, task)
    augmented_recall = compute_recall(existing_modules + [candidate_module], task)
    delta = augmented_recall - baseline_recall
    is_significant = multiple_testing_correction(delta, alpha=0.05)
    return {
        "module": candidate_module,
        "delta_recall": delta,
        "is_significant": is_significant,
        "verdict": "ADD" if is_significant else "REDUNDANT"
    }

Expert Perspective

"AI for Science has become remarkably good at seeing molecules and cells. 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."
— Dr. Jianghui Xiong, Chief Scientist, DeepoMe

Why Longevity Science as the First Domain

For DeepoMe, longevity science and aging-related disease are the initial learning domain for this AI mission. These settings combine four conditions that make human-biology learning feasible:

1. Repeated measurement: Biological state can be sampled multiple times
2. Multidimensional state: Multiple biological axes can be characterized simultaneously
3. Modifiable interventions: Diet, supplements, drugs, lifestyle can be documented
4. Cross-disease relevance: Aging processes span multiple disease categories

DeepoMe's multidimensional aging-assessment work, including Capome, is an entry point for measuring human biological state longitudinally and interpretably—linking repeated measurements with structured intervention exposure and follow-up creates the conditions to learn how, and for whom, an intervention changes that state.

Toward N-of-1 Prediction

For research and biopharma partners, DeepoMe is developing an auditable translational-decision workflow for:

• Prioritizing drug-repurposing hypotheses
• Defining candidate responder populations
• Specifying the next validation experiment or longitudinal study

For aging and Longevity Medicine, the long-term significance is moving intervention reasoning from population-level associations toward individualized N-of-1 prediction and, ultimately, the generation of individualized intervention programs.

A "virtual patient" is not a rendered human image, nor a declared-complete digital twin—it is a testable, auditable inference substrate: identify the biological dimensions most relevant to a given individual, prioritize intervention hypotheses, pair them with longitudinal follow-up, and verify whether predictions hold.

The next phase of world models may not be about parameter scale—it may be about who first accumulates auditable, testable human-representation assets.

References

1. Xiong J, Xia Q. Toward a Self-Learning AI Agent for Drug Repurposing: Building Human-Scale Representations for Virtual Patients. Preprints, 2026. DOI: 10.20944/preprints202608.0998.v1. [This paper]
2. López-Otín C, Blasco MA, Partridge L, et al. The hallmarks of aging. Cell, 2013, 153(6): 1194-1217.
3. López-Otín C, Blasco MA, Partridge L, et al. Hallmarks of aging: An expanding universe. Cell, 2024, 187(2): 361-378.

---

Disclaimer: This article is a technology overview. The cited research has not yet completed peer review and does not constitute medical advice. For health concerns, consult a qualified healthcare professional.