SteeraMed-bench

Building human-scale representations for virtual patients — a module-panel evaluation benchmark that uses drug repurposing as its human-scale testbed (332 modules, 1,916 drugs, five knowledge families).

From virtual cells to virtual patients

Virtual patients need a useful coordinate system before they can model patient-specific dynamics. Virtual-cell programs are building the molecular and cellular scales; treatment decisions, however, are made across a whole person and across different forms of intervention. SteeraMed Bench is a pioneering step toward that missing human scale: it uses drug repurposing as a human-scale testbed for a central virtual-patient question — can a structured representation of biological directions organize intervention-relevant knowledge well enough to prioritize known drug–disease relationships?

Beyond virtual cells — modeling the human response to intervention. AI is already strong at modeling proteins, generating molecules and simulating cellular perturbations. The central opportunity is now shifting: not only molecules or cells, but how a whole person responds to intervention. Virtual patients are the goal — and a useful coordinate system of biological directions is the prerequisite. (AI-generated concept illustration)

Drug development is gated at the human scale. Drug-development value is won or lost at the clinical gate: a Nature Reviews Drug Discovery perspective estimates that cutting Phase II failure by 20% can save nearly $900 million per successful drug launch. Drug repurposing compresses this human-scale question into a computable benchmark — if a representation of biological directions cannot prioritize known drug–disease relationships, it is unlikely to support harder intervention decisions. That is the gatekeeper role repurposing plays here. (AI-generated concept illustration)

What is SteeraMed?

SteeraMed Bench is a framework for evaluating module panels built from a 332-module atlas, developed by DeepoMe. The atlas combines extended aging hallmarks, traditional Chinese medicine syndrome proxies, nutraceutical targets, and food-as-medicine targets. It scores 1,916 drugs against every module using the Guney network-proximity metric on the STRING v12 interactome, then asks one question: which module panels best recover the approved drugs for each disease?

Across five chronic disease tasks, the full atlas reaches Recall@20 = 0.524, and the nutraceutical panel including extensions (NUT+NUTX, 117 modules) reaches 0.494 — both far above the column-permutation null (0.017–0.032 in the paper's Table 1; 0.01–0.07 across panels in this package's reproduction). No single panel wins everywhere: the atlas works as a panel-selection system, exposing which biological organization is most useful for each disease — hallmarks for type 2 diabetes and osteoporosis, food-as-medicine for depression, nutraceutical modules for atherosclerosis/hyperlipidemia. As the paper reports, these standard-CV numbers are within-benchmark estimates: performance decreases under target-family-separated evaluation and approaches chance under leave-one-disease-out evaluation.

SteeraMed-bench (this package) is the reproducibility companion: a dependency-light Python package plus pre-computed data that reproduces the paper's Table 1 panel results (via the predefined panels and the custom-module entry point) and lets you evaluate your own module panels under the exact same protocol. You can also explore the atlas and try panel evaluation directly in the browser: https://steeramed.com/bench

The framework as a whole (paper, Figure 1): curated module sources are assembled into a 332-axis module library; an LLM-assisted workflow proposes and revises candidate gene sets, scored for incremental benchmark value and redundancy; the benchmark then evaluates predefined panels and candidate modules against drug-repurposing tasks. Five chronic diseases are the primary case studies; 23 disease categories provide the exploratory extension.

The 332-module atlas

The benchmark is organized around five module families — four knowledge paradigms in the paper (nutraceutical and its extensions count as one) — covering complementary angles of aging biology and interventions:

Family	Modules	What it captures
Hallmarks:	72	Aging-hallmark modules in five tiers (A1–A5)
TCM:	38	Traditional-Chinese-Medicine modules in three tiers (T1–T3)
NUT:	80	Nutrient / dietary-supplement modules — vitamins, minerals, amino acids, cofactors and botanicals (e.g. NUT:Thiamine, NUT:Ginseng)
NUTX:	37	Extended nutraceutical modules (e.g. NUTX:Betaine)
FAM	105	Food-as-medicine (FAM) herb modules — one per herb (e.g. clove, Chinese yam); module names carry a YFY: prefix in the released data files

Every module ships as a name + pre-computed z-score profile, so the full atlas can be redistributed and evaluated without disclosing any gene list (see the policy section below).

Module-definition policy (minimal definition)

The gene-set definitions of all 332 modules (aging hallmarks, TCM, nutrient and functional-aging modules) are proprietary and intentionally not distributed with this benchmark. The release contains only:

• the aggregated 1916 drugs × 332 modules network-proximity z-score matrix,
• positive drug labels for the five benchmark diseases,
• module-to-panel mapping and family/gene-count metadata.

This is sufficient to reproduce every panel result in Table 1 exactly, while keeping how each module is built (its gene list) confidential. The custom evaluation entry point therefore operates on module names, not gene symbols.

---

Installation

Install directly from GitHub (not yet on PyPI):

pip install git+https://github.com/DeepoMe/SteeraMed-bench.git

For development:

git clone https://github.com/DeepoMe/SteeraMed-bench.git
cd SteeraMed-bench
pip install -e ".[dev]"

Data files are required before evaluation. See Data below.

---

Quick Start

from steeramed_bench import Bench

bench = Bench()                              # load pre-computed matrices
res = bench.evaluate_panel("ALL")            # reproduce Table 1
print(res.recall_at_20)                      # 0.524

Evaluate a custom module panel against a disease:

bench.list_modules("NUT")[:5]                # pick modules by family
out = bench.evaluate_custom_modules(
    ["NUT:Thiamine", "Hallmarks:A1_telomere"],   # any names from list_modules()
    disease="T2D",
)
print(out.recall_at_20)

List what is available:

bench.list_panels()     # 16 panels: tiers A1-A5/T1-T3, families, combinations, ALL
bench.list_diseases()   # 5 benchmark diseases with positive drug labels
bench.list_modules()    # 332 module names (no gene lists)

---

Data

The package depends on four pre-computed artefacts that are distributed separately (they are not bundled with the source and are not committed to git):

File	Description	Approx. size
zscore_matrix.npz	1916 drugs × 332 modules network-proximity z-scores	~5 MB
disease_labels.csv	Positive (approved) drug labels per disease	< 1 MB
panel_mapping.csv	Module → panel assignment	< 1 MB
module_metadata.csv	Module family, panel, gene count (no gene lists)	< 1 MB

Download them from the latest GitHub Release and place the files in steeramed_bench/data/ (or pass data_dir= to Bench). Detailed instructions live in data/README_DATA.md.

What is not included:

• Module gene-set definitions — proprietary; only z-scores and metadata are redistributed (see the policy section above).
• DrugBank raw XML — obtain your own license from https://go.drugbank.com/.
• STRING PPI network (~400 MB) — download from https://string-db.org/.
• repoDB — https://apps.chiragjpgroup.org/repoDB/.

The pre-computed z-scores are derived from these sources but redistributed in aggregated form only.

---

Panels

Panel	# modules	Description
A1–A5	14/36/10/6/6	Aging hallmark tier sub-panels
T1–T3	7/13/18	TCM tier sub-panels
HALLMARKS	72	Aging hallmark modules (tiers A1–A5)
TCM	38	Traditional Chinese Medicine modules
NUT	80	Nutrient modules
NUTX	37	Extended nutraceutical modules
FAM	105	Food-as-medicine herb modules (Chinese food-therapy herbs, one per herb)
HALLMARKS_TCM	110	Hallmarks + TCM combination
TCM_NUT	118	TCM + nutrient combination
ALL	332	Full module atlas

Reference Recall@20 values (averaged over 5 diseases; paper Table 1 rows shown for comparison — note the paper's nutraceutical row is the combined NUT+NUTX (117) panel, evaluated here via the custom-module entry point):

Panel            Recall@20    Paper Table 1
HALLMARKS          0.433         0.433
TCM                0.324         0.324
NUT (80)           0.494           —
NUT+NUTX (117)       —           0.494
FAM                0.403           —
ALL                0.524         0.524

Across the paper's 23 exploratory disease categories, no single panel dominates — the atlas works as a panel-selection system (paper, Figure 2a):

Recall@20 for 23 disease categories × 5 panels (paper, Figure 2a). Hallmarks lead in cancer and osteoporosis, nutraceuticals in analgesic / antipsychotic / cardiovascular categories, food-as-medicine in depression, TCM in sedative and antiepileptic categories. Per-category winners are exploratory; the primary benchmark remains the five chronic-disease case studies.

The panel-evaluation protocol follows the paper (Table 1): stratified 5-fold cross-validated logistic regression (C=0.1, seeds 42/123/456) produces out-of-fold drug scores, and Recall@20 uses the capped denominator min(20, n_positives). The permutation-null runs use this package's own column-permutation implementation (see the null note in the intro), so null values may differ slightly from the paper's Table 1.

---

Examples

Script	What it does
examples/01_reproduce_table1.py	Reproduce Table 1 panel Recall@20
examples/02_custom_module.py	Evaluate a custom module panel end-to-end

python examples/01_reproduce_table1.py

---

Relationship to the paper

This package is the reproducibility companion to:

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

_{Concept illustration: the closed learning loop the paper works toward (AI-generated, not a paper figure)}

In the paper, SteeraMed Bench is the evaluation framework that tests whether module panels — human-scale biological directions drawn from four knowledge paradigms — can prioritize known drug–disease relationships, laying the coordinate foundation for virtual patients and future self-learning agents.

This package implements the panel evaluation and custom module-panel workflows of that framework. The LLM-assisted agent loop for proposing new modules, the module gene-set definitions, the 23-disease-category extension, and the full STRING-proximity re-computation are out of scope for this release; see the paper for details.

---

Links

Resource	URL
Website	https://steeramed.com
Live benchmark demo	https://steeramed.com/bench
Paper	Read the paper (DOI)
Data downloads	GitHub Releases
DeepoMe	https://deepome.com

---

License

MIT © 2026 DeepoMe

---

中文文档 / Chinese README

阅读简体中文版 README · English README