Effect of Low-Carbohydrate and Low-Fat Diets on Metabolomic Indices and Coronary Heart Disease in U.S. Individuals (Paper Feb 26, Diet Quality Matters)

#1

https://www.jacc.org/doi/10.1016/j.jacc.2025.12.038

chatGPT:

Paper in one paragraph (what they did + what they found)

This JACC study followed three large U.S. cohorts (HPFS men 1986–2016; NHS women 1986–2018; NHSII women 1991–2019) totaling ~198k participants, and related multiple “low-carb” (LCD) and “low-fat” (LFD) diet indices—including healthy vs unhealthy versions—to coronary heart disease (CHD) risk. Over ~5.25 million person-years, they documented 20,033 CHD cases. The headline result is that diet quality dominates macronutrient “low carb vs low fat” framing: healthy LCD and healthy LFD were associated with lower CHD risk, while unhealthy versions of both were associated with higher CHD risk. They also built metabolomics-based “multimetabolite scores” (elastic net) to act as objective correlates of these diet patterns and found those scores tracked CHD risk in a way consistent with the questionnaire-based indices. (PubMed)

Key quantitative results (from the abstract)

Comparing highest vs lowest scores (pooled multivariable HRs):

  • LCD
    • Overall LCD: HR 1.05 (1.01–1.10)
    • Animal LCD: HR 1.07 (1.02–1.12)
    • Vegetable LCD: HR 0.94 (0.90–0.99)
    • Unhealthy LCD: HR 1.14 (1.09–1.20)
    • Healthy LCD: HR 0.85 (0.82–0.89)
  • LFD
    • Overall LFD: HR 0.93 (0.89–0.98)
    • Animal LFD: HR 0.94 (0.90–0.98)
    • Vegetable LFD: HR 0.87 (0.83–0.91)
    • Unhealthy LFD: HR 1.12 (1.07–1.17)
    • Healthy LFD: HR 0.87 (0.83–0.91)

Healthy patterns also correlated with lower triglycerides, higher HDL-C, lower hs-CRP, and a “favorable metabolomic profile,” highlighted in the abstract as higher 3-indolepropionic acid and lower valine (example metabolites, not the whole signature). (PubMed)

What’s genuinely novel here

  1. “Low carb vs low fat” reframed into quality-stratified versions
    This isn’t just macronutrient quantity; it explicitly distinguishes healthy vs unhealthy implementations within both LCD and LFD categories (e.g., whole grains/plant fats vs refined carbs/animal-heavy patterns). That design is what allows the “quality > macro ratio” conclusion to be tested directly. (PubMed)
  2. Metabolomics used as a bridge between reported diet patterns and CHD
    They derived multimetabolite scores (elastic net) for each diet index in a validation subset and then tested whether those metabolite scores associated with CHD similarly to FFQ-based indices. This is a step toward objective diet pattern biomarkers, not just single metabolites. (PubMed)
  3. Scale + long follow-up + many CHD events
    The very large event count (20k CHD cases) gives high statistical power to compare multiple diet pattern variants and test consistency across cohorts. (PubMed)

Critique (what to be cautious about)

  1. Observational design → residual confounding is unavoidable
    People choosing “healthy” versions of diets usually differ in many ways (health-seeking behavior, healthcare use, smoking, exercise, socioeconomic factors). Multivariable adjustment helps, but causality still isn’t established.
  2. FFQs are noisy—metabolomics helps, but doesn’t fully solve it
    Diet indices are FFQ-derived, which introduces measurement error. The metabolomic scores are an improvement, but their replication correlations were modest (reported r ≈ 0.21–0.38 in the cohorts), meaning the “objective” signature still only partially captures the intended dietary pattern. (PubMed)
  3. Elastic net metabolomic signatures risk overfitting / instability
    Elastic net is reasonable for high-dimensional metabolomics, but signatures can be sensitive to: preprocessing choices, batch effects, feature selection, and the specific metabolomics platform. Without seeing robustness checks (e.g., repeated resampling stability, external cohorts, platform portability), treat the “objective index” as promising but not definitive.
  4. Multiple testing / selective emphasis in metabolomics
    Metabolomics involves many correlated features. The abstract highlights a couple of metabolites (3-indolepropionic acid, valine), but the full signature likely includes many. The risk is readers over-interpret a small set of “named” metabolites as mechanistic proof rather than markers.
  5. Generalizability
    HPFS/NHS/NHSII are U.S. health professionals—often more health-aware and not fully representative by ethnicity, occupation, or broader dietary contexts. Effects may differ in other populations and food environments. (PubMed)
  6. Effect sizes for “overall LCD” are tiny and directionally mixed
    The overall LCD HR is 1.05 (barely above 1), while the quality-stratified results are much more separated (healthy vs unhealthy). That supports the authors’ thesis—but also means “low-carb” as a generic label is not a useful exposure unless quality is specified.

My bottom-line interpretation

This paper’s main contribution is methodological and conceptual: it shows that “low carb” and “low fat” are too coarse—you need food-source quality to predict CHD directionally, and metabolomics-derived pattern scores can partially corroborate (and potentially improve) FFQ-based pattern measurement. The conclusions are compelling for diet guidance framing, but mechanistic claims should remain cautious until the metabolomic signatures are validated externally and (ideally) supported by intervention data. (PubMed)

If you want, I can also extract the diet index definitions (what exactly counted as “healthy” vs “unhealthy” for LCD/LFD) and turn them into a clean checklist you could apply to real meal patterns—but that will require access to the full text or supplementary methods (the abstract doesn’t list the full scoring rules).