Evidence level: B
孟德尔随机化想拿”体质”当暴露,前提是手里有一把够硬、够纯、够独立的工具变量——而”F>10″这个流传最广的门槛,恰恰是最容易骗人的那一关。
正文
上一篇文章讲了一个很诱人的前景:把中医体质当成”暴露”,用遗传位点做工具变量(IV),借孟德尔随机化(MR)推断”体质→疾病”的因果。韩国人已经用四象体质做到了——太阴型的遗传倾向显著推高心源性栓塞性卒中风险(β=4.190),少阴型则显著压低(β=−9.336)。但这里有个被许多人跳过的问题:那把”刀”(工具变量集)本身够不够硬?如果 IV 是钝的、脏的、彼此纠缠的,MR 给出来的因果结论就可能是幻觉。这就是本文要做的”强度审计”——九种体质要真正跑通 MR,必须依次闯过三道闸门。
第一道闸门:F 统计量,但别迷信”F>10″。 两阶段最小二乘(2SLS)的第一阶段 F 统计量,衡量的是遗传位点对暴露(体质)的解释力——F 越大,工具越硬。流传最广的说法是”F>10 就安全”,但这篇 2011 年的奠基性方法学综述(PMID 21414999)用模拟和真实数据把这条规则击穿了:弱工具偏倚的严重程度随 F 的期望值连续变化,F>10 只是经验法则而非安全线;更麻烦的是,事后根据实测 F 挑选工具或遗传模型会人为放大偏倚,而”F 越大、标准误越小”这一相关性还会给 MR 的 meta 分析引入第二重偏倚。同年另一项研究(PMID 20813862)给出了更精确的账:当 2SLS 的相对偏倚要控制在 10% 以内时,单工具情形 F 至少要达到 9-10;而真实偏倚的大小不仅取决于 F,还取决于暴露-结局混杂的强度——混杂越强,同样的 F 对应的偏倚越大。换句话说,F>10 不是免检章,而是”至少还要做敏感性分析”的起点。
第二道闸门:多效性检验,水平多效性是最大的污染源。 一个 SNP 同时影响体质和结局(不经过体质)就是水平多效性,它会直接违反 MR 的排他性假设。2018 年 Nature Genetics 的 MR-PRESSO 研究(PMID 29686387)对 82 个复杂性状做了 4250 对因果检验,发现超过 48% 的显著因果关联存在可检测的水平多效性,多效性对因果估计的扭曲平均在 −131% 到 +201% 之间,还能在高达 10% 的关系里制造假阳性。更关键的是它给体质 MR 提了个醒:MR-PRESSO 只在”多效性工具占比 <50%”时表现最好——也就是说,如果九种体质的位点里混入大量一因多效的 SNP,连纠错工具都会失灵。2021 年发布的 STROBE-MR 报告规范(PMID 34698778)把这一整套要求固化为标准动作:明确陈述三条 IV 假设、报告工具选择与 LD 修剪细节、报告 F 与 Q 统计量、执行 MR-Egger/加权中位数/留一法敏感性分析——这些条目逐条对下来,就是一份完整的”强度审计清单”。
第三道闸门:LD 修剪后的独立 IV 数量。 这是九种体质最现实的痛点。韩国的四象体质 GWAS 全基因组扫描了 34 万个位点,但显著关联停留在 10⁻⁶ 量级(PMID 22394158)——远够不到全基因组显著线(5×10⁻⁸),这意味着经过 LD 修剪后能站出来的独立 IV 很可能只有个位数甚至为零。再看中医体质这边:阳虚体质小样本 GWAS(30 vs 30 人,PMID 25533513)找到 42 个 p<10⁻⁴ 的 SNP,血瘀(瘀滞)体质在 947 名 2 型糖尿病患者里找到 22 个 SNP 中 PON2 最显著(PMID 26169365),痰湿体质则是基因表达谱层面的差异(PMID 23548117)加全基因组甲基化特征(PMID 29518789)。这些研究证明”体质有遗传基础”,但离”能当 MR 工具变量”还差着数量级:p 值不够全基因组显著、样本量千人级、效应量小——三样叠在一起,IV 强度审计的第一张表(F 统计量)大概率就过不了。
好消息是”可复制性”已经出现曙光。 韩国 2015 年的 FTO 研究(PMID 25888059)在 5490 人的发现+验证双队列里,确认了肥胖风险变异 rs7193144 与少阴型反向关联(OR=0.729,p=1.47×10⁻⁷),且独立于 BMI——这是四象体质第一个真正跨队列复制的位点,说明”体质有稳健遗传位点”并非空想,只是密度太低。痰湿体质的 LDL-R 基因多态性研究(PMID 17898945)也提示 PvuII/AvaII 位点可作为遗传标记。把这些拼起来,体质 MR 的强度审计结论很明确:九种体质要跑双样本 MR,先把痰湿、阳虚、阴虚三个证据最厚的体质做到万人级 GWAS,拿到一批全基因组显著、LD 修剪后互相独立的位点,再按 STROBE-MR 清单走完 F 统计量、MR-PRESSO、MR-Egger 全套体检——这跟韩国四象体质的路径一致,也是一切”体质→疾病”因果结论的入场券。
证据与局限
已验证: 弱工具偏倚的严重性随 F 连续变化、F>10 非安全线的结论有直接方法学证据(PMID 21414999, 20813862);水平多效性在 MR 中普遍存在且能制造假阳性(PMID 29686387);STROBE-MR 报告规范已确立(PMID 34698778);四象体质有跨队列复制的稳健位点(FTO rs7193144,PMID 25888059)。
部分验证: 九种体质各有候选位点但 p 值多为 10⁻⁴ 量级、样本千人级,未达全基因组显著线(PMID 25533513, 26169365, 22394158);痰湿体质有基因表达与甲基化特征但非 GWAS 位点强度(PMID 23548117, 29518789)。
待验证: ①九种体质万人级 GWAS 能否产生一批通过 F>10(及更严格准则)与 MR-PRESSO 多效性检验的独立 IV;②痰湿体质 LDL-R 多态(PMID 17898945)能否在独立队列复制;③四象体质的 FTO 等位点能否在中医九种体质(尤其痰湿质)上复现同一效应方向。
相关阅读
证据等级: B
参考文献: PMID 21414999, PMID 20813862, PMID 29686387, PMID 34698778, PMID 25888059, PMID 25533513, PMID 26169365, PMID 22394158, PMID 23548117, PMID 29518789, PMID 17898945
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。
Audit of Instrument Strength for Phenotypic Tools: Nine Phenotypes Must Pass Three Mendelian Randomization Gates
Evidence level: B
Mendelian randomization (MR) intends to use “constitution” as exposure only if it possesses a robust, pure, and independent instrumental variable (IV); yet the widely circulated threshold of “F > 10” is precisely the most deceptive hurdle.
Main Text
The previous article presented an appealing prospect: treating TCM constitution as “exposure,” using genetic loci as instrumental variables (IVs), and inferring causality (“constitution → disease”) via Mendelian randomization (MR). Koreans have already achieved this with the Four Temperaments constitution—Taiyin type showed a significant genetic tendency increasing risk of cardioembolic stroke ($\beta=4.190$), while Shaoyin type significantly reduced it ($\beta=-9.336$). However, a problem often overlooked is whether the “knife” (the IV set) itself is sharp enough. If the IVs are dull, contaminated, or entangled, the causal conclusions derived from MR may be illusions. This article performs an “intensity audit”: for nine constitutions to truly pass MR, they must sequentially clear three gates.
Gate 1: The F-statistic, but do not迷信 “F > 10”. The first-stage F-statistic in two-stage least squares (2SLS) measures the explanatory power of genetic loci on the exposure (constitution)—the larger the F, the stronger the instrument. The most prevalent rule states “F > 10 ensures safety,” but a foundational methodological review from 2011 (PMID 21414999) shattered this rule using simulations and real data: the severity of weak instrument bias varies continuously with the expected value of F; F > 10 is an empirical heuristic rather than a safety line. More troublingly, selecting instruments or genetic models post-hoc based on observed F values artificially inflates bias, and the correlation between “higher F and smaller standard errors” introduces a second layer of bias into MR meta-analyses. Another study from the same year (PMID 20813862) provided a more precise accounting: to keep relative bias in 2SLS within 10%, F must reach at least 9–10 for single-instrument scenarios; furthermore, the magnitude of actual bias depends not only on F but also on the strength of exposure-outcome confounding—stronger confounding yields larger bias for the same F. In other words, F > 10 is not a pass mark, but rather the starting point for “at least conducting sensitivity analyses.”
Gate 2: Pleiotropy Testing, Horizontal Pleiotropy is the Largest Contaminant. A SNP affecting both constitution and outcome directly (not via constitution) constitutes horizontal pleiotropy, violating MR’s exclusion restriction assumption. A 2018 Nature Genetics study on MR-PRESSO (PMID 29686387) performed 4,250 causal tests across 82 complex traits and found that over 48% of significant causal associations exhibited detectable horizontal pleiotropy. Pleiotropy distorted causal estimates by an average of −131% to +201% and generated false positives in up to 10% of relationships. Crucially, it alerted constitution MR: MR-PRESSO performs best only when the proportion of pleiotropic instruments is <50%. Thus, if the loci for nine constitutions contain a large number of one-causes-many-effects SNPs, even correction tools may fail. The 2021 STROBE-MR reporting guideline (PMID 34698778) codified these requirements into standard procedures: explicitly state the three IV assumptions, report details on instrument selection and LD pruning, report F and Q statistics, and execute sensitivity analyses using MR-Egger, weighted median, and leave-one-out methods. Taken together, these items form a complete “intensity audit checklist.”
Gate 3: Number of Independent IVs After LD Pruning. This is the most realistic pain point for the nine constitutions. The Korean Four Temperaments GWAS scanned 340,000 loci, but significant associations remained at the $10^{-6}$ level (PMID 22394158)—far below the genome-wide significance threshold ($5 \times 10^{-8}$). This implies that independent IVs surviving LD pruning may number only in the single digits or even zero. Looking at TCM constitution research: the Yangxu constitution small-sample GWAS (30 vs 30 individuals, PMID 25533513) identified 42 SNPs with $p < 10^{-4}$; the Blood Stasis (Yu Zhi) constitution found 22 SNPs in 947 type 2 diabetics with PON2 being most significant (PMID 26169365); and the Phlegm-Damp constitution showed differences at the gene expression profile level (PMID 23548117) plus whole-genome methylation features (PMID 29518789). These studies prove “constitution has a genetic basis,” but they are orders of magnitude away from being suitable as MR IVs: p-values fail genome-wide significance, sample sizes are in the thousands rather than tens of thousands, and effect sizes are small. Combined, these factors mean the first table of the IV strength audit (F-statistic) will likely fail for most constitutions.
Good news: signs of “reproducibility” are emerging. The Korean 2015 FTO study (PMID 25888059) confirmed in a discovery-validation dual cohort of 5,490 individuals that the obesity risk variant rs7193144 showed a reverse association with Shaoyin type (OR=0.729, $p=1.47 \times 10^{-7}$), independent of BMI. This was the first truly cross-cohort replicated locus for the Four Temperaments constitution, indicating that “robust genetic loci for constitution” are not mere fantasies, albeit with low density. The study on LDL-R gene polymorphism for Phlegm-Damp constitution (PMID 17898945) also suggested the PvuII/AvaII sites could serve as genetic markers. Putting these together, the conclusion of the constitution MR intensity audit is clear: for the nine constitutions to run two-sample MR, first achieve genome-wide significant, LD-pruned independent loci for the three constitutions with the strongest evidence (Phlegm-Damp, Yangxu, Yinxu) via million-person-scale GWAS. Then proceed through the full STROBE-MR checklist covering F-statistics, MR-PRESSO, and MR-Egger. This aligns with the path taken by the Korean Four Temperaments constitution and serves as the ticket for all “constitution → disease” causal conclusions.
Evidence and Limitations
Verified: Conclusions that weak instrument bias severity varies continuously with F and that F > 10 is not a safety line are supported by direct methodological evidence (PMID 21414999, 20813862); horizontal pleiotropy is prevalent in MR and can generate false positives (PMID 29686387); the STROBE-MR reporting guideline is established (PMID 34698778); and Four Temperaments constitution has robust loci with cross-cohort replication (FTO rs7193144, PMID 25888059).
Partially Verified: Nine constitutions have candidate loci but p-values are mostly at the $10^{-4}$ level with sample sizes in the thousands, failing to reach genome-wide significance (PMID 25533513, 26169365, 22394158); Phlegm-Damp constitution has gene expression and methylation features but lacks GWAS locus intensity (PMID 23548117, 29518789).
To Be Verified: ① Whether million-person-scale GWAS for nine constitutions can generate a batch of independent IVs passing F > 10 (and stricter criteria) and MR-PRESSO pleiotropy tests; ② Whether the Phlegm-Damp LDL-R polymorphism (PMID 17898945) can be replicated in an independent cohort; ③ Whether the FTO allele from Four Temperaments constitution can be reproduced in TCM Nine Constitutions (especially Phlegm-Damp) with the same effect direction.
Related Reading
- Constitution as Exposure: How Mendelian Randomization Turns “Constitution → Disease” from Correlation to Causality
- Multicenter GWAS Replication of TCM Nine Constitutions
- TCM Constitution = Genetic Basis
- Constitution Polygenic Risk Scores: From Innate Endowment to Quantifiable Causal Chains
- Biological Frequency Research Pool
Evidence Level: B
References: PMID 21414999, PMID 20813862, PMID 29686387, PMID 34698778, PMID 25888059, PMID 25533513, PMID 26169365, PMID 22394158, PMID 23548117, PMID 29518789, PMID 17898945
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.