方向性

步骤 3 · ★否决点 · 图 3 张 · 发表版表格 2 个　（流程见 v5 定稿）

变异层（MR-Steiger）与性状层（反向 MR）两个层面合起来判方向。全流程两个否决点之一。

图

Figure F3a. Trait-level directionality: evidence for a reverse effect of disease liability on plasma protein level, for every association taken forward from the screen

完整图注（英文，与投稿版一致）

Figure F3a. Trait-level directionality: evidence for a reverse effect of disease liability on plasma protein level, for every association taken forward from the screen.

Directionality is tested at two levels and neither replaces the other. Reverse Mendelian randomization uses the genome-wide instruments of the disease and asks whether liability to the disease changes the plasma protein level. MR-Steiger uses the protein's own cis sentinel and asks whether that variant explains more of the exposure than of the outcome; it is blind to a protein that is downstream of disease, because a cis variant explains its own protein strongly regardless. Figures F3a and F3b report the trait level, Figure F3c the variant level.

Every association that passed FDR < 0.05 in the forward direction (n = 142) is plotted with its forward effect on the x axis and the evidence for a reverse effect on the y axis. The y axis is -log10 of the FDR-adjusted P value for the reverse effect, so the single dashed line at FDR = 0.05 is the veto threshold for every panel: points above it are vetoed. The x axis is on a free scale per outcome because the outcomes differ greatly in effect magnitude. The 4 vetoed complication associations are named on the figure.

Reverse causation is concentrated in the diabetes traits, not in the complications: 4 of 57 complication associations were vetoed, against 41 of 85 in the two diabetes traits. This is the expected pattern if diabetes itself alters circulating protein levels, and it is one reason the complication estimates are interpreted separately.

WHAT IS EXCLUDED HERE. The veto is applied per association, not per protein. The 4 vetoed complication associations are APOE (Retinopathy), BTN3A2 (Retinopathy), ITGB7 (Retinopathy), TIGIT (Retinopathy); they are labelled on the figure. Of those proteins, 2 had no other complication association and therefore leave the complication analysis entirely (ITGB7 and TIGIT), while the rest are retained through their remaining complications. The complication side therefore goes from 57 associations involving 31 proteins to 53 associations involving 29 proteins. On the diabetes side 41 of 85 associations are vetoed, involving 40 distinct proteins.

Threshold definition and its sensitivity. The veto uses the FDR computed WITHIN each outcome. An alternative definition, FDR computed across all 142 associations at once, disagrees for 6 of them; under that alternative the number of vetoed complication associations would be 7 rather than 4. The within-outcome definition was chosen for consistency with the forward analysis, which also applies FDR within each outcome. Both values are reported in the supplementary table so that either convention can be checked.

Grey crosses mark associations that could not be assessed; they are not counted as passing. Their composition is given in Figure F3b.

Method: reverse MR used genome-wide significant variants for the outcome as instruments, with variants inside the cis region of the tested protein's own gene removed; effects were combined by fixed-effect inverse-variance weighting. The number of instruments per test ranged from 2 to 98.

Role in the analysis pipeline: this is a VETO point. Associations vetoed here are excluded from all downstream steps; no later evidence can reinstate them.

Data sources and sample sizes. Exposure: UKB-PPP plasma proteome (Olink Explore 3072), 34,557 European participants, 1,954 proteins with a cis-pQTL. Outcomes: FinnGen R9 -- Diabetic retinopathy 10,413 cases / 308,633 controls; Diabetic maculopathy 3,572 cases / 308,547 controls; Diabetic nephropathy 4,111 cases / 308,539 controls; Diabetic neuropathy 2,843 cases / 271,817 controls. Diabetes control arms: Type 1 diabetes 20,355 cases / 797,363 controls; Type 2 diabetes 55,005 cases / 400,308 controls.

Figure F3b. Composition of the trait-level directionality check, by outcome

完整图注（英文，与投稿版一致）

Figure F3b. Composition of the trait-level directionality check, by outcome.

Of 142 associations, 93 were supported, 45 were vetoed, and 4 could not be assessed. The three states are shown separately on purpose: NOT ASSESSABLE is not a result and is never counted as passing.

All 4 not-assessable associations are for diabetic neuropathy -- AGER (Neuropathy, 2 instruments), HCG22 (Neuropathy, 2 instruments), BTN2A1 (Neuropathy, 2 instruments), TNXB (Neuropathy, 2 instruments) -- where only 2 instruments remained after removing variants in the cis region of the protein's own gene; that exclusion is required to avoid a circular test, but it leaves too few instruments for a meaningful reverse estimate. These associations are CARRIED FORWARD, not excluded; the directionality check simply makes no statement about them.

Excluded here, by name: APOE (Retinopathy), BTN3A2 (Retinopathy), ITGB7 (Retinopathy), TIGIT (Retinopathy). ITGB7 and TIGIT leave the complication analysis entirely, having had no other complication association. After this step the complication side carries 53 associations involving 29 proteins, down from 57 and 31.

Reverse causation is concentrated in the diabetes traits: 4 of 57 complication associations were vetoed, against 41 of 85 in the two diabetes traits. The individual vetoed associations are named in Figure F3a.

Threshold: an association is vetoed when the reverse effect reaches FDR < 0.05, with the FDR computed within each outcome. Figure F3a reports how the counts change under a global FDR instead.

Role in the analysis pipeline: this is a VETO point. Associations vetoed here are excluded from all downstream steps; no later evidence can reinstate them.

Figure F3c. Variant-level directionality: the smallest MR-Steiger margin per outcome

完整图注（英文，与投稿版一致）

Figure F3c. Variant-level directionality: the smallest MR-Steiger margin per outcome.

MR-Steiger is computed AFTER the association test, on the 142 associations that passed FDR < 0.05 -- the same set that reverse Mendelian randomization is applied to (5 to 57 associations per outcome). It is not used to filter instruments before the association test. For each outcome the point is the SMALLEST ratio of exposure-explained to outcome-explained variance among that outcome's significant associations; the dashed line at 1 is where the direction would be called wrong. No association failed. The margin is not uniformly large: it falls to 1.33 for AIF1 (Type 1 diabetes), so this is a test that was passed rather than a formality. The protein named beside each point is the one with that smallest margin.

5 of the 6 tightest-margin proteins are encoded in the MHC region, where the disease signal is strong enough to raise the outcome-explained variance; that is a direct consequence of not excluding the MHC during instrument selection.

Method: MR-Steiger compares the variance in exposure and in outcome explained by the same variant, with the outcome-side R2 obtained on the liability scale from the log odds ratio, case and control counts and the population prevalence. In this workflow it is a directionality test applied to associations, not a filter applied to instruments: no variant is removed from the association test on the basis of it.

Threshold: the ratio must be >= 1 for the variant to explain more of the exposure than of the outcome; instruments below that threshold would be removed. No FDR or other multiple-testing correction applies to this panel because it reports one summary value per outcome rather than a set of tests.

Role in the analysis pipeline: this is a VETO point. Associations vetoed here are excluded from all downstream steps; no later evidence can reinstate them.

发表版表格

文件	下载
`directionality_mr_steiger.csv`	下载
`reverse_mr_verdict.csv`	下载

本页图为网页版 PNG（最宽 1600 px）。投稿用的矢量 PDF 与 600 dpi TIFF 体积较大，留在仓库 results/reverse_mr/figures/，不随文档站分发。

方向性 ​

图 ​

Figure F3a. Trait-level directionality: evidence for a reverse effect of disease liability on plasma protein level, for every association taken forward from the screen ​

Figure F3b. Composition of the trait-level directionality check, by outcome ​

Figure F3c. Variant-level directionality: the smallest MR-Steiger margin per outcome ​

发表版表格 ​

方向性

图

Figure F3a. Trait-level directionality: evidence for a reverse effect of disease liability on plasma protein level, for every association taken forward from the screen

Figure F3b. Composition of the trait-level directionality check, by outcome

Figure F3c. Variant-level directionality: the smallest MR-Steiger margin per outcome

发表版表格