Nucleus of the soy cluster — the canonical page for what soy does to patient-important
outcomes, whether the feared effects are real, and how the forms differ. Facets (thyroid,
natto/K2-bone, phytate handling) attach here.
The reframe — not “is soy good or bad?” but three fears under symmetric standards
The public question is polarized between soy feminizes men / causes breast cancer and soy is a superfood. Neither pole is the decision. The wiki holds a health axis only and applies the same appraisal bar to the fear and to the boast — a favourable result buys no exemption, an alarming one no free pass. Three fears are separable and each resolves to a direction + magnitude + certainty, not to it’s complicated:
- Hormonal (isoflavones as phytoestrogens -> feminization in men) — refuted, a well-powered null (Reed et al., 2021).
- Breast cancer — neutral-to-protective, and conditional on menopausal status, population and study design (Chen et al., 2014).
- LDL cholesterol — modest lowering via the isolated soy-protein form; the best current intrinsic estimate is ~3% (−4.76 mg/dL) at ~25 g/d (Blanco Mejia et al., 2019), a downward revision of the dated ~13% figure (Anderson et al., 1995). A surrogate — the events benefit is inferred through LDL ApoB and Cumulative Exposure, not measured.
Thyroid (goitrogen) and antinutrient (phytate) fears are named gaps below — not searched here, so neither asserted nor dismissed.
Why the hormonal null is mechanistically expected — isoflavones are SERMs, not estrogen
Isoflavones (genistein, daidzein) have a structure similar to estrogen but preferentially bind and activate ERβ, whereas estrogen has roughly equal affinity for ERα and ERβ; because the two receptors have different tissue distributions and can exert opposite effects, isoflavones act tissue-selectively and are classified as selective estrogen-receptor modulators (SERMs) (Reed et al., 2021). So a naive phytoestrogen = estrogen inference is the wrong model, and the human null below is what the mechanism predicts — mechanism carries directional weight here, but the outcome evidence, not the mechanism, is what settles it (net-effect-not-intended).
Fear 1 — HORMONAL: a genuine null, not “insufficient evidence”
Reed 2021 meta-analyzed the clinical data on whether soy or isoflavone intake moves male reproductive hormones — total testosterone (TT), free testosterone (FT), estradiol (E2), estrone (E1), sex hormone-binding globulin (SHBG): 41 studies; TT and FT measured in 1753 and 752 men, E2 and E1 in 1000 and 239, SHBG in 967 (Reed et al., 2021).
- effect_measure: standardized mean difference (SMD), random-effects. Total testosterone SMD −0.06 (95% CI −0.29 to 0.17; P=0.59; 20 groups, 1241 men, treatment-vs-control); change-over-time 0.09 (−0.02 to 0.20; P=0.12). Free testosterone 0.01 (−0.33 to 0.32; P=0.98). SHBG −0.03 (−0.45 to 0.38; P=0.88). E2 and E1: no significant effect on any model. (Reed et al., 2021)
- population_and_comparator: adult men consuming soyfoods / soy protein (SPI, SPC) / isoflavone extracts, vs control (usually animal protein or usual diet).
- outcome: circulating reproductive hormones — a surrogate, not a patient-important outcome (the feared endpoints are fertility/feminization; hormone levels stand in for them) -> Surrogate Outcomes.
- dose_response_shape: no dose or duration effect — sub-analysis by isoflavone dose (<75 vs ≥75 mg/d) and by study duration (≤12 vs >12 weeks) showed no effect; none of the four ≥12-month studies found an effect on testosterone. (Reed et al., 2021)
- certainty: high for the TT null specifically — n=1753, SMD −0.06 with a tight CI (−0.29 to 0.17) that excludes a meaningful effect, low heterogeneity (I2 30% on the change-over-time model), no funnel-plot publication bias, superseding and confirming a 2010 MA. FT is only moderate: its CI (−0.33 to 0.32) does not exclude a small effect in either direction (n=752 — an imprecise estimate, not a tight null). E1/E2 rest on smaller samples (239/1000). The author’s sweeping summary «neither soy protein nor isoflavone exposure affects TT, FT, E2 or E1 levels in men» is best read as carried by the TT endpoint (Reed et al., 2021).
This is a null, not a gap — most firmly for testosterone. Apply the expectancy test: if soy lowered testosterone, a 41-study MA measuring TT in 1753 men, powered to detect it, would have. It did not, across every model, dose, and duration — that is evidence of no meaningful effect on the endpoint the feminization fear is actually about, one of the four evidence states, not unproven. The estradiol/estrone arms are consistent with no effect but less tightly bounded. The feared cases (gynecomastia, hypogonadism) involved intakes near 360 mg/d isoflavones, roughly nine times a typical native-Japanese intake (~30-50 mg/d) — outside the range achievable from foods (Reed et al., 2021).
Provenance caveat (symmetric standards): the corresponding author (Messina) and one co-author work in soy-industry-adjacent nutrition consultancy. The null is nonetheless robust — low heterogeneity, no funnel-plot asymmetry, a pre-existing 2010 MA reaching the same result, and it runs toward the direction a SERM mechanism predicts — so the finding is not carried by the framing. But the source is flagged, not exempt.
Fear 2 — BREAST CANCER: neutral-to-protective, conditional, and study-design-confounded
Chen 2014 pooled 35 epidemiological studies (odds ratios, high-vs-low isoflavone intake), stratified by menopausal status and study region/design (Chen et al., 2014).
| Stratum | Premenopausal OR (95% CI) | Postmenopausal OR (95% CI) |
|---|---|---|
| Summary | 0.74 (0.64–0.85) | 0.75 (0.63–0.86) |
| Asian countries | 0.59 (0.48–0.69) | 0.59 (0.44–0.74) |
| Western countries | 0.90 (0.77–1.04), NS | 0.92 (0.83–1.00), marginal |
- The pooled protection is carried by the weaker designs (read this before the headline number). The summary OR ~0.75 is confounded by study design: prospective cohort / nested case-control studies — the stronger designs — showed no significant protection (premenopausal cohorts OR 0.94, 0.74–1.14; postmenopausal cohorts 0.86, 0.73–1.00 — borderline, upper bound touching the null), while the inverse association came from retrospective case-control studies, whose recall and selection biases systematically inflate a diet-disease association. So the ~25% reduction is a case-control-driven figure, not a cohort-confirmed one. This is a weaker-design inflates the favourable arm pattern (a cousin of the artifact-arm problem on The U-Shaped Association Artifact, though the mechanism here is recall/selection bias, not unequal reporting precision). (Chen et al., 2014)
- Direction: inverse (protective) or neutral — never a raised risk in any stratum.
- Population-conditional: in Western women (pre- or post-menopausal) the marginal Western signal vanished once stratified by design — Chen concludes there is no evidence of an association in Western women, attributing the Asian protection to early-life and higher lifelong intake rather than to adult supplementation. Transportability caveat: the protective estimate is an Asian / lifelong-exposure finding, not a promise for a Western adult adding soy.
- certainty: low-to-moderate — observational, heterogeneous (I2 up to 84%), publication bias detected (Egger), and the strongest designs are null. The defensible claim is “not a risk, and plausibly protective under lifelong high intake,” not “soy prevents breast cancer.”
Fear 3 — LDL CHOLESTEROL: modest lowering, baseline-dependent, via isolated soy protein
Anderson 1995 — the landmark MA of 38 controlled trials substituting soy protein (mean 47 g/d) for animal protein (Anderson et al., 1995):
- effect_measure (absolute, net vs control diet): total cholesterol −23.2 mg/dL (−9.3%); LDL −21.7 mg/dL (−12.9%); triglycerides −13.3 mg/dL (−10.5%); HDL non-significant (+2.4%). (Anderson et al., 1995)
- dose_response_shape: the soy-diet-alone model estimated 25 / 50 / 75 g/d -> cholesterol decreases of 8.9 / 17.4 / 26.3 mg/dL — roughly monotone over the studied range, no knee located. (Anderson et al., 1995)
- effect_modifiers: baseline risk (route (a)) dominates — initial cholesterol accounted for ~77% of between-study variance. Normal (<200 mg/dL): non-significant −3.3%; severe hypercholesterolemia (>335 mg/dL): −19.6%. The lever is real mainly for people who already have high cholesterol. (Anderson et al., 1995)
- population_and_comparator: isolated or textured soy protein replacing animal protein — whole soybeans were excluded from the analysis. This is the specify-the-exposure catch: the LDL evidence is for the isolated-protein form, not for tofu/edamame, and not for soy sauce -> Is the Food Category Doing Any Work.
- outcome: LDL/total cholesterol — a surrogate for ASCVD; its causal transmission to events is the LDL ApoB and Cumulative Exposure claim, not assumed here.
The magnitude is revised DOWN — Blanco Mejia 2019 (type-F refinement of the Anderson leg)
Blanco Mejia 2019 re-pooled the 46 soy-protein RCTs the FDA compiled to decide whether to revoke the soy heart-health claim (43 with data; 41 for LDL / 50 comparisons, 43 for TC / 52 comparisons; 2607 participants). At a median dose of 25 g/d over a median 6 wk, soy protein (79% isolated, vs mostly dairy-protein controls) lowered:
- LDL by −4.76 mg/dL (−0.12 mmol/L); 95% CI −6.71, −2.80; = −3.2% (95% CI −4.5, −1.9%); I2=55%, no publication bias, no subgroup effect modification, and no dose-response or threshold located (the studied dose range was narrow) (Blanco Mejia et al., 2019).
- TC by −6.41 mg/dL (−0.17 mmol/L); 95% CI −9.30, −3.52; = −2.8% (95% CI −4.1, −1.5%); I2=74% (Blanco Mejia et al., 2019).
- The abstract states the effect as «approximately 3-4%» LDL lowering (Blanco Mejia et al., 2019).
This is F (claim-refinement), not E (independent backing): both are meta-analyses pooling soy-protein RCTs on the same surrogate, over an overlapping trial pool (Blanco cites Anderson as ref 70 and compares directly), so they are the same route re-estimated, not two independent routes converging — no confidence bump is bought by their agreement.
Parameter table — why the two magnitudes differ (BLOCKING before the comparison prose):
| Parameter | Blanco Mejia 2019 (chunk 01) | Anderson 1995 (chunk 01) | Same quantity? |
|---|---|---|---|
| LDL net change vs control | −4.76 mg/dL (−0.12 mmol/L); −3.2% | −21.7 mg/dL (0.56 mmol/L); −12.9% | Same measure (net between-group MD on LDL, matched-protein control), but NO on the overall figure — confounded by dose + baseline (below) |
| TC net change vs control | −6.41 mg/dL (−0.17 mmol/L); −2.8% | −23.2 mg/dL (0.60 mmol/L); −9.3% | same measure; differs by dose + baseline |
| Soy dose | median 25 g/d (IQR 23.8, 38.1) | mean 47 g/d (range 17-124) | NO — Anderson ~2x the dose |
| Baseline population | LDL 110-201 mg/dL, mean 147.6 (borderline) | wide, incl. severe hypercholesterolemia (>335); baseline drove ~77% of variance | NO — Anderson spans much higher (more responsive) baselines |
| Comparator | non-soy protein (72% dairy) | animal protein, fat/SFA/cholesterol-matched in most | ~yes (both matched-protein = “intrinsic”) |
| Exposure form | 79% ISP + soy foods/milk | isolated/textured SP; whole soybeans excluded | ~yes (predominantly isolated protein) |
Reconciliation. The overall Anderson −12.9% and Blanco −3.2% are not the same quantity: Anderson pooled both higher doses and higher-baseline populations, and both inflate the effect (dose-response + the route-(a) baseline dependence Anderson itself found). Matched at Blanco’s population, the gap largely closes — Anderson’s own normal-baseline (<200 mg/dL) quartile was NS −3.3% cholesterol (LDL Q1 −7.7%), and its 25 g/d dose-model point was −8.9 mg/dL TC (vs Blanco’s −6.41). So the downward revision is partly a genuine update (a cleaner FDA-identified set, no detectable dose-response) and partly explained by lower dose + lower baseline, not a contradiction (Anderson et al., 1995; inferred from Blanco Mejia et al., 2019). Both MAs agree on direction and on baseline-amplification.
Intrinsic vs extrinsic (a decision-relevant distinction, type-B). Blanco’s pooled effect is the
intrinsic effect (soy vs a matched-protein control). The real-world effect adds an extrinsic /
displacement component when soy replaces saturated-fat-rich animal protein — «The overall effect
in real life could be potentially higher than that seen in these trials» — an NHANES-III estimate put
the displacement advantage of 25 g/d soy for animal protein at ~4.3% additional LDL lowering
(Blanco Mejia et al., 2019). So the trial figure is a floor
for a substitution framing (layer-3), not the ceiling.
Surrogate bounding (the sharp point). LDL/TC are surrogates — Blanco reports no hard cardiovascular outcomes, calling them «surrogates of CVD». The events benefit of a −0.12 mmol/L LDL reduction is inferred through the cumulative-exposure causality on LDL ApoB and Cumulative Exposure, never measured here — and in absolute terms 0.12 mmol/L is a small LDL move beside a standard drug lever (a statin lowers LDL ~1.5-2 mmol/L), so soy protein sizes as a small, non-substitutable-for-a-drug lever on this outcome (layer-1). The intrinsic effect is real, directionally certain, and modest.
Mechanism (proposed, not established). Blanco attributes the intrinsic effect to the 7S globulin
fraction of soy protein (shared with other legumes), which «appears to inhibit hepatic Apo B
synthesis»; isoflavones are judged a «less likely» contributor to the lipid effect
(Blanco Mejia et al., 2019) [EXTRACTED — asserted]. This
is a mechanistic proposal with directional weight, not an outcome finding.
Two live caveats (symmetric standards):
- The 1995 magnitude is superseded. Anderson’s ~13% LDL figure was DATED and industry-funded (Protein Technologies International, lead author on its advisory group); treat Blanco Mejia’s ~3% intrinsic effect at ~25 g/d as the current estimate, holding Anderson for the baseline-dependence and dose structure it still supplies. Blanco is itself soy/food-industry-entangled (co-author Messina directs the Soy Nutrition Institute) — but it runs against sponsor interest by shrinking the claim, on a gold FDA-identified set with no funnel-plot asymmetry, so the direction is not carried by the framing. Both favourable-industry results get more scrutiny, not less (Blanco Mejia et al., 2019).
- The mechanism was not established in either MA; Anderson notes primate data attributing 60–70% of the effect to soy estrogens (isoflavones), but this is animal-derived and speculative, and Blanco judges the isoflavone route the less likely one (Anderson et al., 1995).
The mortality cell — soy / isoflavones / soy protein x all-cause, cancer, CVD, breast cancer (Nachvak 2019)
The three fears above are surrogate-or-incidence outcomes; this is the direct patient-important outcome the page previously left blank. Nachvak 2019 is a gold-tier SR + dose-response MA of 23 prospective cohorts (330,826 participants; 26,900 all-cause, 11,005 cancer, 13,418 CVD deaths; published 2002-2018) that keeps the three exposures distinct — whole soy/soy products, soy isoflavones, soy protein — against four endpoints. All estimates are highest-vs-lowest pooled RR, random-effects, observational (Nachvak et al., 2019).
Do not collapse the three forms into “soy” — the matrix (type-B):
| Exposure | All-cause | Cancer (total) | CVD | Breast cancer |
|---|---|---|---|---|
| Whole soy / soy products | 0.90 (0.77-1.04) null overall; inverse only in higher-quality subgroups | 0.88 (0.79-0.99) ✓ (soy/soy-products); total-soy 0.90 (0.81-1.00) marginal; soy-specific foods 0.99 null | 0.85 (0.72-0.99) ✓ (soy/soy-products); total-soy 0.91 (0.81-1.03) null; soy-specific 1.03 null | not pooled for whole soy |
| Soy isoflavones | 0.90 (0.82-0.98) ✓ categorical; linear +10 mg/d 0.98 (0.94-1.01) null | 0.80 (0.67-0.94) ✓; linear +10 mg/d 0.93 (0.89-0.98) ✓ (−7%) | 0.98 (0.90-1.06) null; linear null | 0.83 (0.69-0.99) ✓; linear +10 mg/d 0.91 (0.84-0.99) ✓ (−9%); ER− 0.77 ✓, ER+ 0.82 null |
| Soy protein | 1 study only — not poolable | (all cancer effect sizes were breast cancer) → | 0.99 (0.84-1.17) null; dose-response null | 0.73 (0.55-0.96) ✓; linear +5 g/d 0.88 (0.83-0.93) ✓ (−12%, I²=7.4%); ER+ 0.75 ✓, ER− 0.77 ✓ |
Whole soy, cancer-site-specific (total soy, highest-vs-lowest): gastric 0.49 (0.35-0.68), lung 0.79 (0.71-0.87), colorectal 0.59 (0.41-0.84) all inverse; hepatic 0.89 (0.71-1.12) null. Ischemic CVD 0.79 (0.63-0.99) inverse; stroke 0.87 (0.73-1.04) null (Nachvak et al., 2019).
Four caveats that bound every cell — read before the headline “soy for longevity”
Nachvak’s own conclusion is that soy «may support the current recommendations to increase intake of soy for greater longevity», but it pairs that with «However, soy protein was not significantly associated with all-cause and CVDs mortal- ity» (Nachvak et al., 2019). Four structural caveats matter more than the boast:
- The all-cause signal for whole soy is NULL overall, significant only in the “high-quality” subgroups. «this significant association for all-cause mortality was observed only in the high-quality studies (studies with sample size 10,000 participants and those that presented energy-adjusted effect sizes)» (Nachvak et al., 2019). A subgroup that becomes significant after the full-set null is a hypothesis, not a confirmed effect — the headline all-cause benefit for soy foods is not carried by the primary pooled estimate (0.90, P=0.16).
- Categorical-significant, but the LINEAR per-mg dose-response is null (isoflavone → all-cause). The highest-vs-lowest contrast is 0.90 (✓, P=0.02) and a nonlinear inverse relationship was significant (P for nonlinearity=0.04), yet the linear +10 mg/d dose-response is null: 0.98 (0.94-1.01, P=0.17). So there is no clean per-mg gradient — the effect is not linear-dosable and a per-mg isoflavone target is not derivable. Nachvak explicitly cautions «Overall, caution in the interpretation of findings from the doseeresponse analysis is required» (Nachvak et al., 2019), the dose-response models being «only empowered to detect massive associations» (Nachvak et al., 2019). Where a cell IS both categorical- and linear-significant (isoflavone → cancer −7%/10 mg, 0.93 (0.89-0.98); isoflavone → breast-cancer −9%/10 mg, 0.91 (0.84-0.99); soy-protein → breast-cancer −12%/5 g, 0.88 (0.83-0.93)) the claim is firmer, but still over the studied range only.
- The breast-cancer-mortality cell is a SURVIVOR endpoint, not primary prevention (
type-A, the cross-cell observation the flat abstract hides). Every study feeding the isoflavone → breast-cancer and soy-protein → breast-cancer cells is a breast-cancer patient/survivor cohort measuring post-diagnosis intake against survival (Conroy, Nechuta, Shu, Boyapati, Kang, Zhang) — Nachvak even folds Shu’s relapse deaths into the endpoint: «we considered risk estimates on relapse/breast cancer-specific mortality as breast cancer mortality» (Nachvak et al., 2019). So the ~12%/5 g soy-protein and ~9%/10 mg isoflavone breast-cancer figures answer “does soy intake improve survival in women already diagnosed?”, NOT “does soy prevent breast-cancer death in the general population?” — and reverse causation (sicker patients eat/report differently) is un-excluded in a survivor cohort. This directly cashes the page’s ER-positive-survivor / tamoxifen-interaction gap (below): the survivor evidence is protective-or-neutral in BOTH ER strata for soy protein and ER-negative-only for isoflavones — no raised risk in any stratum, contra the stimulate-the-tumour fear. (inferred from Nachvak et al., 2019) - Transportability + a publication-year split the authors discount. 19 of 23 cohorts are Asian (Nachvak et al., 2019), where habitual soy intake is far above Western levels — the estimate is an Asian/high-lifelong-intake finding, the same transportability caveat the Chen breast-cancer leg carries. Year of publication was one of two subgroups that explained the whole-soy all-cause heterogeneity: the protective association held in studies published up to 2010 (0.77, 0.69-0.86) but not 2010-and-later (1.06, 0.93-1.21) (Nachvak et al., 2019), and Nachvak invokes Kicinski’s mechanism — «the amount of overrepresentation of findings that are favorable to treatment is larger in older studies … found protective effects for soy intake in studies published up to 2010, but not for those studies that were published during the year 2010 and later» (Nachvak et al., 2019). But the authors’ own bottom line discounts it: «Overall, it seems that publication year has a low effect on the overall estimates in the current meta-analysis» (Nachvak et al., 2019). Isoflavones ran the opposite way (protective effects mostly in 2010-and-later studies). Reading the pre-2010 confinement as a decline-effect artifact is a possible interpretation the authors do not endorse . All cells are observational: «we cannot rule out the possible effect of residual confounding» (Nachvak et al., 2019).
Layer-1 read. On the direct mortality outcome soy is a small, form-dependent, observational lever: the firmest cells are whole-soy/soy-products → cancer (0.88) and → CVD (0.85) mortality and the dosed isoflavone → cancer (−7%/10 mg) arm; the all-cause benefit is soft (null overall for foods, dose-null for isoflavones); soy protein moves only a survivor breast-cancer endpoint and is null on all-cause/CVD. No cell is a big rock, and none rests on a causal design — a reason to eat soy without fearing it, not a longevity lever to chase.
The forms are NOT interchangeable — specify the exposure
“Soy” names several different exposures with different active fractions and different evidence
(a terminological disambiguation with decision consequences, type-B):
- Isolated / textured soy protein — the LDL-lowering form (Anderson’s exposure); high-DIAAS complete plant protein -> Protein Quality and the DIAAS Score.
- Whole / minimally processed (tofu, edamame, soy milk) — the isoflavone + complete-protein workhorse; carries the hormone/cancer evidence (foods, not isolates).
- Fermented (miso, tempeh, natto) — lower phytate (fermentation defuses the antinutrient concern -> Antinutrients in Plant Foods); natto is a notably rich dietary source of vitamin K2 (MK-7) (a widely-reported compositional fact; whether the MK-7 moves a bone or vascular outcome is the GAP below). Cross-links Fermented Foods for the live-culture question, but the soy matrix lives here.
- Soy sauce / tamari — a condiment and salt vehicle with negligible isoflavone dose; do not credit it with soy’s effects. A specify-the-exposure catch.
Synthesis — soy’s evidence is unusually FORM-dependent
Placing the three MAs side by side yields a move present in no single source (type-A): each fear was
tested with a different soy exposure — Reed pooled soyfoods + isolates + isoflavone extracts on
hormones, Chen pooled dietary-isoflavone epidemiology on cancer, Anderson isolated only the isolated
soy-protein form on lipids and excluded whole soybeans. So the composite answer is not “soy does X”
but a form-indexed one: the isolated protein moves LDL, the isoflavone-bearing foods leave male
hormones unmoved and sit neutral-to-protective on breast cancer, and the condiment does nothing. This
is Is the Food Category Doing Any Work instantiated — the category “soy” is too coarse to carry a
single verdict, and the better-designed evidence tracks the better-specified exposure.
The three sources answer three different outcomes (hormones · cancer · lipids), so they are
facets of this nucleus, not a tension — no joined issue, no forced independent-backing claim. The
lipids outcome now rests on two MAs, but F-related (Blanco Mejia refines Anderson’s magnitude on
an overlapping trial pool — not independent E), so confidence stays medium overall (four gold-tier
MAs, but each outcome effectively single-route, and the LDL/TC endpoints are surrogates).
One axis only
This page holds a health axis. Soy also carries environmental (low-footprint protein) and economic loads and is entangled with the animal-feed / deforestation debate; the wiki records only that those trade-offs exist and does not price them against the health finding.
Facets held elsewhere
- Cognition (a fourth outcome) — HELD. Supplemental soy isoflavones produce a small, memory-carried benefit on neuropsychological test scores; the earlier tofu-harm scare does not survive confounding scrutiny; and the observational isoflavone-cognition signal is null. The full appraisal — the Cui 2020 RCT MA effect estimates, the tofu-harm rebuttal, and why the RCT/observational split is a distinction rather than a tension — lives on Soy Isoflavones and Cognitive Function. The ERβ-SERM mechanism above is the same one invoked there for the memory-domain specificity.
Gaps and held threads
- LDL magnitude — CASHED (Blanco Mejia 2019 held). The intrinsic effect is ~3% (−4.76 mg/dL) at
~25 g/d, revising Anderson’s ~13% downward; see Fear 3. Remaining LDL gap: no hard CVD-outcome
trial of soy protein — the events benefit stays inferred through LDL ApoB and Cumulative Exposure.
G (needs a soy -> ASCVD-events trial, which does not exist). - Thyroid (goitrogen) — GAP, not searched. Prior (to test, not asserted): mild goitrogen only
under iodine deficiency; in euthyroid, iodine-replete people likely no clinically important harm
(may modestly raise levothyroxine dose needs — an absorption/timing issue).
AWAITS a soy-thyroid SR/MA conditioned on iodine status. - Antinutrient / phytate — the mineral-binding concern is handled at Antinutrients in Plant Foods (reduced by fermentation/soaking; a marginal-mineral-status issue, not a general harm); no soy-specific outcome evidence extracted here.
- Natto / vitamin K2 (MK-7) -> bone and vascular calcification — GAP.
AWAITS a natto/MK-7 bone or CVD outcome SR/MA(cross-links Fermented Foods). - Prostate cancer, menopausal symptoms, bone density — named in the deliverable scope, no source held. GAP.
- Soy in ER-positive breast-cancer survivors (the genistein / tamoxifen-interaction fear) —
PARTLY CASHED (Nachvak 2019). The survivor-cohort breast-cancer-mortality cells run protective or
neutral in BOTH ER strata for soy protein (ER+ 0.75, ER− 0.77, both ✓) and ER-negative-only for
isoflavones (ER− 0.77 ✓, ER+ 0.82 null) — no raised mortality in any stratum, contra the
stimulate-the-tumour worry (Nachvak et al., 2019).
RESIDUAL: this is observational post-diagnosis survival with un-excluded reverse causation, and it
does NOT directly test the tamoxifen-interaction mechanism (concurrent-endocrine-therapy
subgroups not isolated).
AWAITS a soy/isoflavone x tamoxifen interaction SRfor the mechanism leg. - Absolute breast-cancer risk reduction — Chen reports relative ORs only; the absolute benefit
depends on baseline incidence in the stratum, which the MA does not supply.
G (needs aggregation).