填 43 道题,还是花几千块测菌群?——中医体质问卷 vs 菌群测序的”预测力对决

Written by

in

Language

Evidence level: B

同样一个问题——”我该吃哪种益生菌、哪个调理方案对我最管用”——现在有两条分型路径:中医九种体质问卷填 43 道题(几分钟、几乎零成本),或肠道菌群测序(几百到几千元、数周出报告)。这两套分型到底谁更能预测干预效果?它们差多少?这决定了一个被忽略的关键问题:贵的测序,值不值那个钱?

正文

先看体质这一侧凭什么说话。近十年研究反复确认:中医九种体质不是玄学标签,而是在菌群上留得下指纹的分类。2018 年一项研究发现,气虚体质者的肠道菌群 α 多样性显著低于平和质,且菌群组成与疾病易感性升高直接相关(PMID 30425748);2025 年一项三体质整合分析显示,阴虚、平和、阳虚三质的菌群与代谢物谱显著分离(PMID 40258500);2023 年一项九体质研究更进一步——不同体质供体的粪便菌群不仅组成不同,移植到高脂饮食小鼠后抗肥胖效果也不同,平和质供体的合格率更高(PMID 41853345)。换句话说,体质问卷捕捉到的差异,在菌群层面能”验到货”。

体质侧的实战成绩单也不差。2024 年一项失眠干预研究给出最漂亮的证据:63 名患者按两种中医证型(脾气虚、脾气虚兼胃热)分别用两首不同方剂,基线菌群即按证型分开,机器学习仅凭菌群就能区分证型,且两方各按自己的菌群-炎症通路起效(PMID 38817449)。2024 年另一项湿热质 RCT 中,二妙丸+生活方式干预 1 个月显著降湿热评分,菌群向平和质靠拢(PMID 38851100)。这些都指向:辨证选方,在菌群层面有真实的生理基础。

再看菌群测序这一侧。它最大的卖点是”预测应答”——在干预之前告诉你谁会有效。2024 年一项非便秘型肠易激综合征试验中,副干酪乳杆菌 DG 干预应答者基线富集 Collinsella aerofaciens 等 18 个类群(PMID 38178601);2018 年 Cell 经典研究证明 2/3 的人对益生菌存在”定植抗性”,且这种抗性可被基线宿主与菌群特征预测(PMID 30193112);2015 年老年益生菌研究显示应答者基线 α 多样性显著更高(PMID 26441879)。测序+机器学习甚至能预测更细的结局:2015 年 Zeevi 团队用菌群等多组学数据预测个体餐后血糖反应,并在小 RCT 中验证了个性化饮食有效(PMID 26590418);2020 年 1102 人的 PREDICT 1 研究把话挑明——对餐后甘油三酯反应,个体菌群解释的方差(7.1%)比餐食宏量营养素(3.6%)还大(PMID 32528151)。

但高成本是否换来了成比例的预测力?这才是 #370 要回答的问题。目前两套证据链各自独立:体质问卷有菌群关联和辨证干预 RCT,菌群测序有应答预测模型和个性化膳食 RCT,却没有任何一项研究把两者放进同一批受试者、用同一把尺子(干预应答 AUC)头对头比较过。这正是缺口所在。

成本差值得先说清楚。中医体质问卷(CCMQ)由王琦团队开发,43 个条目、9 个分量表,已在加拿大、日本、韩国、美国等跨文化人群中使用,几分钟填完、几乎零边际成本(PMID 30425748);而 16S 测序每例数百元、宏基因组数千元,还需采样、测序、生信分析、报告解读数周周期。如果两者预测力相当,一线筛选用问卷即可;只有当测序的增量预测力超过其增量成本,个性化医学的”测序优先”才成立——这是一个纯粹的成本-效益实证问题,不该由商业叙事替我们回答。

验证设计其实不复杂:同一队列,先做九种体质问卷,再取粪便测序,然后给同一菌群干预,最终用三套模型预测应答——只靠问卷、只靠菌群、两者联合,比较三者的 ΔAUC 与净重分类改善指数(NRI)。结果有三种可能:若菌群模型的 ΔAUC 显著大 → 测序确实值钱,问卷只是替代品;若两者 AUC 相当 → 低成本问卷可以一线筛选,测序只留给”问卷判不定”的分歧区;若联合模型明显更强 → 两套分型携带互补信息,各管一块。

这个”共识区 vs 分歧区”的剖析尤其有意义:当问卷和菌群指向同一干预时,可以自信处方;当两者冲突时,恰恰是需要更细数据(代谢组、连续血糖、FMT 验证)裁决的高价值人群——这对应中医”望闻问切先于高科技,但不能被高科技取代”的方法论分层。分型不是二选一,而是成本-效益的最优叠加。

证据与局限

已验证: 九种体质与菌群组成存在稳定关联,平和/偏颇质的菌群差异可被测序与代谢组复现(PMID 30425748、40258500、41853345);辨证用药的效应有菌群机制支撑,证型可被菌群+机器学习区分(PMID 38817449、38851100);菌群测序可预测益生菌定植与应答(PMID 38178601、30193112、26441879);菌群+机器学习可预测餐后代谢反应且个性化膳食 RCT 有效(PMID 26590418、32528151)。

部分验证: “体质问卷 vs 菌群测序”在同一数据集上对干预应答的预测力增量对比尚无直接研究;现有比较集中在”菌群个性化 vs 一刀切”,未纳入体质问卷作为低成本对照。

待验证: 同一队列双套分型(体质问卷+16S/宏基因组)+ 同菌群干预 RCT + 两套分型分别/联合建模应答预测 AUC + 增量检验;分型一致率与冲突区(问卷说 A、测序说 B)的特征分析——冲突人群是否需要更高阶数据介入。

相关阅读

证据等级: B(体质-菌群关联、菌群应答预测、个性化膳食 RCT 各有 PubMed 直接支撑;但”两套分型在同一数据集头对头预测干预应答”的直接研究尚不存在,属强科学类比/可检验假说)
参考文献: PMID 38178601, PMID 30193112, PMID 26441879, PMID 32528151, PMID 26590418, PMID 38817449, PMID 38851100, PMID 41853345, PMID 30425748, PMID 40258500


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

Fill 43 Questions or Spend Thousands for Sequencing? A Duel on Predictive Power Between Traditional Chinese Medicine Constitution Questionnaires and Gut Microbiome Sequencing

Evidence level: B

The same question—”Which probiotic and regimen work best for me?”—now has two stratification pathways: completing a TCM Nine-Constellation questionnaire (43 items, minutes, near-zero cost) or gut microbiome sequencing (hundreds to thousands of dollars, weeks for results). Which system better predicts intervention outcomes? What is the gap between them? This determines an overlooked critical issue: Is the expensive sequencing worth the cost?

Main Text

First, let’s examine the basis for the constitution side. Over the past decade, research has repeatedly confirmed that the TCM Nine-Constellations are not esoteric labels but classifications that leave a distinct fingerprint on the microbiome. A 2018 study found that individuals with Qi-deficiency constitution showed significantly lower gut microbial alpha diversity compared to those with Balanced constitution, and microbial composition was directly correlated with increased disease susceptibility (PMID 30425748); a 2025 integrated analysis of three constitutions showed that the microbial profiles and metabolomic spectra of Yin-deficiency, Balanced, and Yang-deficiency constitutions were significantly distinct (PMID 40258500); a 2023 study on nine constitutions went further—fecal microbiomes from donors of different constitutions not only differed in composition but also showed varying anti-obesity effects when transplanted into high-fat diet mice, with donors from the Balanced constitution showing higher success rates (PMID 41853345). In short, the differences captured by the constitution questionnaire can be “verified” at the microbial level.

The practical track record on the constitution side is also strong. A 2024 insomnia intervention study provided compelling evidence: 63 patients were divided into two TCM syndromes (Spleen Qi deficiency, Spleen Qi deficiency with Stomach Heat) and treated with two different formulas; baseline microbiomes were already stratified by syndrome, and machine learning could distinguish syndromes based solely on microbiome data, with each formula activating via its own microbiome-inflammation pathway (PMID 38817449). In another 2024 Randomized Controlled Trial (RCT) on Damp-Heat constitution, Er-Miao Pill plus lifestyle intervention significantly reduced damp-heat scores after one month, with microbiomes shifting toward the Balanced constitution (PMID 38851100). These findings point to: Syndrome differentiation and formula selection have a genuine physiological basis at the microbial level.

Now, let’s look at the microbiome sequencing side. Its primary selling point is “predicting response”—telling you beforehand who will respond to an intervention. A 2024 trial in non-constipated Irritable Bowel Syndrome found that responders to Lactobacillus paracasei DG baseline enrichment included 18 taxa such as Collinsella aerofaciens (PMID 38178601); a classic 2018 Cell study demonstrated that 2/3 of individuals exhibit “colonization resistance” to probiotics, which can be predicted by baseline host and microbiome characteristics (PMID 30193112); a 2015 study on elderly probiotics showed that responders had significantly higher baseline alpha diversity (PMID 26441879). Sequencing combined with machine learning can even predict finer outcomes: Zeevi’s team in 2015 used microbiome and multi-omics data to predict individual postprandial glucose responses, validating personalized diets in a small RCT (PMID 26590418); the 2020 PREDICT 1 study involving 1,102 individuals made it explicit—for postprandial triglyceride response, the variance explained by individual microbiomes (7.1%) was larger than that by macronutrients in the meal (3.6%) (PMID 32528151).

But does high cost yield proportional predictive power? This is the question #370 seeks to answer. Currently, the two evidence chains operate independently: the constitution questionnaire has microbiome associations and RCTs for syndrome-based interventions, while microbiome sequencing has response prediction models and personalized dietary RCTs. However, no study has yet placed both systems within the same cohort, using the same metric (intervention response AUC) for a head-to-head comparison. This is the gap.

The cost disparity warrants clarification first. The TCM Constitution Questionnaire (CCMQ), developed by Wang Qi’s team, comprises 43 items across nine subscales and has been used in cross-cultural populations including Canada, Japan, South Korea, and the US; it takes minutes to complete with near-zero marginal cost (PMID 30425748). In contrast, 16S sequencing costs hundreds per sample, while metagenomics costs thousands, requiring sampling, sequencing, bioinformatics analysis, and report interpretation over several weeks. If predictive power is equivalent, the questionnaire suffices for frontline screening; only if the incremental predictive gain of sequencing exceeds its incremental cost does the “sequencing-first” approach for personalized medicine hold merit—this is a pure cost-benefit empirical question that should not be answered by commercial narratives.

The validation design is not complex: use the same cohort, administer the Nine-Constellation questionnaire first, then collect stool for sequencing, apply the same microbial intervention to all, and finally predict response using three models—questionnaire only, microbiome only, or combined. Compare the ΔAUC and Net Reclassification Improvement (NRI) across the three. There are three possible outcomes: If the microbiome model’s ΔAUC is significantly larger → sequencing is indeed valuable, and the questionnaire is merely a substitute; If both AUCs are comparable → low-cost questionnaires can serve as frontline screening, reserving sequencing for cases where the questionnaire yields inconclusive results; If the combined model is clearly superior → both stratification systems carry complementary information, each managing a specific domain.

This analysis of “consensus zones vs. divergence zones” is particularly meaningful: when the questionnaire and microbiome point to the same intervention, prescribing can be done with confidence; when they conflict, it is precisely this high-value population requiring finer data (metabolomics, continuous glucose monitoring, FMT validation) for adjudication—corresponding to the TCM methodological stratification of “observation and inquiry preceding high-tech, but not being replaced by it.” Stratification is not an either/or choice, but an optimal superposition based on cost-benefit.

Evidence and Limitations

Verified: Stable associations exist between Nine-Constellations and microbiome composition; differences in Balanced/Biased constitutions can be replicated via sequencing and metabolomics (PMID 30425748, 40258500, 41853345); efficacy of syndrome-based medication is supported by microbial mechanisms, and syndromes can be distinguished via microbiome + machine learning (PMID 38817449, 38851100); microbiome sequencing can predict probiotic colonization and response (PMID 38178601, 30193112, 26441879); microbiome + machine learning can predict postprandial metabolic responses, with validated personalized dietary RCTs (PMID 26590418, 32528151).

Partially Verified: No direct study exists comparing the incremental predictive power of “TCM questionnaire vs. microbiome sequencing” on intervention response within the same dataset; existing comparisons focus on “microbiome personalization vs. one-size-fits-all,” without incorporating the constitution questionnaire as a low-cost control.

To Be Verified: Same cohort with dual stratification (TCM questionnaire + 16S/metagenomics) + same-microbial-intervention RCT + separate and combined modeling of response prediction AUC for both stratifications + incremental testing; analysis of concordance rates and divergence zones (questionnaire says A, sequencing says B)—whether conflicting populations require higher-order data intervention.

Related Reading

Evidence Level: B (Associations between constitution and microbiome, microbiome response prediction, and personalized dietary RCTs each have direct PubMed support; however, a direct study comparing both stratification systems in the same dataset for predicting intervention response does not yet exist, representing strong scientific analogy/testable hypothesis)
References: PMID 38178601, PMID 30193112, PMID 26441879, PMID 32528151, PMID 26590418, PMID 38817449, PMID 38851100, PMID 41853345, PMID 30425748, PMID 40258500


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *