Direct Intervention Ranking from a Health State-Representation Map for the Multimorbidity Challenge 从健康状态表征地图直接排序候选干预·面向多病共存挑战
MorbiMap is a candidate-ranking component of SteeraMed. It connects a health state-representation map — the module-weighted representation of a molecular profile — with frozen drug–module relevance tables and uses a fixed scoring program to rank candidate drugs — without requiring a disease label as input to the scoring function.
Four DNA-methylation cohorts and two disease-associated gene signatures → module weights → frozen drug–module tables → fixed scoring → three-level evaluation.
Synthetic paired-profile analyses test whether a single top-20 list can recover drugs from two disease-label sets. Depression–Parkinson's mixtures retained both baselines on average; the depression-only endpoint already attained dual coverage.
- A hypertension gene signature recovered 6–7 antihypertensives in the top 20; removing 30 direct-target genes retained 6
- Depression and Parkinson's high recovery was shared with healthy controls — detected by a healthy-reference differential, not by drug-only or permutation checks
- 32% of depression–Parkinson's pairings met both baselines; the depression-only endpoint also met both
- A degree- and weight-preserving graph null reproduced the overlap–coverage association (ρ_null = 0.85 vs observed 0.90)
Version 1 (preprint). This repository provides the result-to-artifact index, representative figures, and version declaration. The analysis code, frozen candidate pools, positive-set files, drug–module tables, audit trail, and reproduction scripts will be released in version 2 (v2).
| Headline result | Frozen artifact (v2) |
|---|---|
| Candidate pool and sealed positive sets | results/pool200_clinical_v2.json |
| Per-participant 72-module |ES| profiles | results/_repr_gsea_signed_v2.npy + item_map_v2.json |
| Grounded drug–module tables (9 batches) | results/geneanchor72_claude_batch*.json |
| Depression and PD healthy-control differential gates | results/e2_triangle_full.json |
| Cross-fitted calibration | results/e15_calibration.json + e15_calibration_seed7.json |
| Per-drug differential audit | results/e15_diff_audit.json |
| HTN direct-target deletion, class deletion, null | results/e16_htn_independence.json + e16b_classholdout.json |
| Exhaustive pairing coverage (Table 3) | results/comorbidity_exhaustive_means.json |
| Self-mixing, healthy-partner controls, λ sweep | results/e14_e17_selfmix.json |
| Degree- and weight-preserving graph null | results/e19_graph_null.json |
| Blinded edge-level review package | results/e5_blind_edge_package.xlsx / .json |
All analyses use publicly available GEO datasets: GSE42861, GSE179325, GSE111223, GSE113725, GSE128235, GSE44132, GSE198904, GSE201287.
@preprint{xiong2026morbidmap,
title={Prescribing Without a Disease Diagnosis? Direct Intervention Ranking
from a Health State-Representation Map for the Multimorbidity Challenge},
author={Xiong, Jianghui},
year={2026},
note={Preprint. Code and frozen artifacts:
https://github.com/DeepoMe/SteeraMed-MorbiMap}
}- Paper (ResearchGate version) — preprint version; DOI to follow on preprints.org
- SteeraMed — the broader framework
- DeepoMe — the organization behind this work
- SteeraMed-RootMap — companion aging-dependency repository (with paper audio & alphaXiv discussion)
- SteeraMed-bench — companion benchmark repository
MIT (code) / CC BY 4.0 (data and documentation)
Jianghui Xiong — jianghui@deepome.com
MorbiMap 是 SteeraMed 框架的候选排序组件。 它将健康状态表征地图（分子谱的模块加权表征）与冻结的药物-模块关联表连接，用固定评分程序排序候选药物——评分函数不需要疾病标签输入。
四个 DNA 甲基化队列 + 两个疾病基因签名 → 模块权重 → 冻结药物-模块表 → 固定评分 → 三层评测。
合成配对谱分析测试单个 top-20 列表能否覆盖两组疾病标签。抑郁-帕金森混合平均保留双基线；纯抑郁端点已达双覆盖。
- 高血压基因签名在 top 20 中恢复 6–7 个降压药；删除 30 个直接靶基因后保留 6 个
- 抑郁与帕金森的高恢复与健康对照共享——由健康参照差分检出，药物排序/置换检查无法单独发现
- 抑郁-帕金森配对中 32% 同时达到双基线；纯抑郁端点也达到双覆盖
- 度数-权重保持的图空模型复现了重叠-覆盖关联（ρ_null = 0.85 vs 观测 0.90）
版本 1（预印本）。 当前提供结果-文件索引、代表性图表与版本声明。 分析代码、冻结候选池、阳性集文件、药物-模块表、审计链与复现脚本将在**版本 2（v2）**中发布。
所有分析使用公开 GEO 数据集： GSE42861, GSE179325, GSE111223, GSE113725, GSE128235, GSE44132, GSE198904, GSE201287.
MIT（代码）/ CC BY 4.0（数据与文档）
熊江辉 — jianghui@deepome.com