The PREDICT Studies — Personalized Nutrition Data
The PREDICT studies — Personalized Responses to Dietary Composition Trial — represent the most comprehensive investigation of individualized metabolic responses to food ever conducted, recruiting over 2,000 participants across the UK and US, collecting continuous glucose monitoring data, blood lipid kinetics, gut microbiome sequencing, and detailed dietary diaries simultaneously to build predictive models of individual postprandial metabolic responses that outperform any single-biomarker or food-composition-based prediction. PREDICT 1 (Berry et al. 2020, Nature Medicine, n = 1,002 adult twins and unrelated individuals in the UK) established the foundational dataset; PREDICT 2 (n = 1,000 US participants, ZOE Ltd) extended the cohort and validated the predictive algorithm on an independent population.1 Together they constitute the scientific backbone of the ZOE personalized nutrition programme and have generated over 40 peer-reviewed publications rewriting conventional understanding of why identical diets produce different metabolic outcomes in different people.
Study Design: What Was Actually Measured
PREDICT 1 was exceptional in its measurement density. Participants ate standardized “muffin” test meals with precisely controlled macronutrient ratios on study days — each muffin contained fixed amounts of fat, carbohydrate, protein, and fibre, allowing the researchers to isolate the metabolic response to a known nutritional input rather than the chaotic variance of free-living meals. Participants wore Abbott Libre CGMs continuously for two weeks, providing 15-minute-interval interstitial glucose readings throughout the study period.1
Simultaneously, participants gave serial blood draws at 0, 1, 2, 4, and 6 hours after each muffin meal to measure postprandial triglyceride kinetics — an aspect of metabolic response that CGM alone cannot capture. Fasting blood insulin and glucose were measured to characterize insulin resistance at baseline. Stool samples were collected for gut microbiome profiling using 16S rRNA sequencing, yielding taxonomic composition data that could later be correlated with metabolic responses. Participants completed validated 7-day diet diaries via the ZOE smartphone app, wore Actigraph accelerometers for physical activity, and answered validated psychosocial questionnaires on sleep, stress, and appetite.1
The result was five simultaneous measurement streams — glucose dynamics, triglyceride kinetics, insulin, gut microbiome composition, and detailed free-living diet — collected on the same participants over the same period. This multivariate dataset is what distinguishes PREDICT from earlier nutrition intervention studies, which typically measured one or two biomarkers in response to a dietary intervention rather than decomposing the sources of individual metabolic variability.
PREDICT 2 replicated the protocol in 1,000 US-based participants, testing whether algorithms derived from the UK cohort generalised across geographic and dietary cultural differences. The validation was successful: the predictive models derived from PREDICT 1 explained similar proportions of variance in postprandial glucose and triglyceride response in the US cohort, supporting the generalisability of the biological mechanisms identified.2
Key Finding 1: Genetics Explains Less Than Expected
The most conceptually significant finding from PREDICT 1 emerged from its twin subsample. Identical twins share approximately 100% of their genomic sequence, while fraternal twins share approximately 50% — the same as non-twin siblings. By comparing postprandial responses between identical and fraternal twin pairs, the researchers could estimate the heritable fraction of metabolic variation using classical twin heritability analysis.
Host genetics explained only 34% of variance in postprandial triglyceride response and 49% of variance in postprandial glucose response. The remaining variance — 51–66% — was attributable to gut microbiome composition, habitual diet, lifestyle factors, and environmental influences.1 Crucially, identical twins showed substantially different postprandial glucose and triglyceride responses to the same standardized meal, confirming that the non-genetic factors driving these differences are real and not simply measurement noise. The broader science of postprandial glucose variability contextualises these twin findings within the wider inter-individual variability literature.
This finding directly challenges the “genetic determinism” framing of metabolic disease risk that dominated nutritional science for decades. If two people with identical DNA can have substantially different metabolic responses to the same meal, genetic testing alone cannot predict who will benefit from which dietary approach. The gut microbiome — which is highly modifiable through diet — emerges as a larger lever than genomic predisposition for a majority of individuals.
The practical implication extends to how we think about population-level dietary guidelines. Guidelines based on average responses to dietary changes (reduce saturated fat, increase fibre) are derived from cohort data that averages over precisely this individual variability. If 40–50% of individuals respond differently from the mean to a given dietary change, population-level advice is accurate guidance for the statistical center but systematically wrong for a large minority of people who receive it.
Key Finding 2: Postprandial Triglycerides Are the Overlooked Biomarker
Conventional nutrition science has focused on postprandial glucose as the primary metabolic outcome after a meal — driven largely by the development of CGM technology, which makes glucose visible in real time, and by the dominance of diabetes as the framing disease for dietary research. PREDICT’s simultaneous measurement of postprandial triglycerides exposed a significant gap in that narrative.
Postprandial triglyceride responses (measured by serial blood draws over six hours after standardized meals) were at least as variable and as predictive of cardiometabolic risk as glucose responses in the PREDICT dataset.1 The correlation between triglyceride response and cardiometabolic risk markers (visceral fat area, HOMA-IR, HDL cholesterol) was comparable in strength to the equivalent glucose correlations. Critically, triglyceride and glucose responses were only moderately correlated with each other — meaning a person with a high glucose response to a meal did not necessarily show a high triglyceride response, and vice versa.
This decoupling has practical consequences. High-fat meals (high olive oil, cheese, or nut content) that produce modest glucose spikes — which a CGM-based nutrition approach might score favorably — may produce large triglyceride excursions in individuals with underlying insulin resistance or elevated baseline triglycerides. These triglyceride excursions are invisible to a CGM. Nutrition apps that track only postprandial glucose as the metabolic outcome miss approximately half the metabolic story.3
The finding also reframes the discussion around dietary fat. High-saturated-fat meals that produce large triglyceride peaks represent cardiovascular risk even in individuals whose fasting triglycerides appear normal — because postprandial lipemia (elevated triglycerides after meals) is increasingly recognized as an independent cardiovascular risk factor, distinct from fasting lipid measurements taken during the standard lipid panel.
Key Finding 3: Meal Timing and Sleep Modulate Responses
One of PREDICT 1’s most actionable findings was the quantification of circadian and sleep effects on metabolic response — effects that occur independently of food composition and gut microbiome.
Eating the same standardized meal later in the day (after 5 PM vs before noon) increased postprandial glucose AUC by approximately 18% and triglyceride AUC by approximately 25% in the same participants.1 This circadian variation in metabolic response is driven by the daily rhythm of insulin sensitivity — peripheral tissues (muscle, liver, adipose) are most insulin-sensitive in the morning and least sensitive in the evening, a pattern governed by the circadian clock genes in each tissue. The same grams of carbohydrate consumed at 8 AM produce a smaller glucose excursion than at 8 PM, not because the food is different, but because the metabolic machinery processing it is less responsive by evening.
Poor sleep quality the preceding night — assessed via wrist actigraphy, which distinguishes sleep duration from sleep fragmentation — was associated with approximately 20% higher postprandial glucose responses to standardized breakfast meals.1 The mechanism involves elevated cortisol and growth hormone secretion during poor sleep, which promotes hepatic glucose production and reduces peripheral insulin sensitivity the following morning. Crucially, this sleep effect was additive to the food composition effect: a participant who slept poorly and ate a high-glycaemic breakfast showed substantially higher glucose excursions than either the food composition or sleep deprivation alone would predict.
These timing and sleep effects are not captured by any static food composition database. A meal’s macronutrient profile is fixed, but its metabolic impact depends substantially on when it is consumed and what preceded it the night before. Nutrition tracking systems that ignore circadian context and sleep data are modelling an important but incomplete slice of the actual metabolic picture.
The ZOE Algorithm: From PREDICT to Consumer Product
ZOE Ltd — the company that co-funded and conducted the PREDICT studies — translated the research findings into a consumer metabolic testing product. The ZOE kit involves a two-week home CGM wear period, a stool sample collection for gut microbiome sequencing, and a fasting finger-prick blood test for lipid baseline. Responses from all three measurement streams are fed into a gradient-boosted machine learning model trained on the full PREDICT dataset to generate personalized food scores for over 5,000 foods.2
Independent evaluation of the ZOE programme was published by Bermingham et al. (2022, Nature Medicine) — an 18-week randomized controlled trial comparing ZOE-guided dietary advice against standard population-level dietary advice in 347 adults. Participants following ZOE-personalized recommendations showed postprandial glucose improvements of 17% and postprandial triglyceride improvements of 22% compared with controls receiving standard dietary guidance.2 These are clinically meaningful differences, particularly for postprandial triglyceride — a biomarker that standard dietary advice rarely targets explicitly.
The mechanism of benefit was predominantly personalization rather than the overall healthiness of the recommended foods. ZOE-guided participants did not simply eat “healthier” by conventional metrics — they ate foods calibrated to their individual glucose, triglyceride, and microbiome responses, meaning some participants were guided toward foods that would be rated average on a population-level nutrient density score but were specifically well-matched to their individual metabolic profile.
Limitations and What PREDICT Doesn’t Resolve
PREDICT’s contributions are substantial, but the studies carry important limitations that are underemphasized in popular science coverage.
The participant cohorts were predominantly white, educated, and UK or US-based. Gut microbiome composition differs substantially between populations with different ancestries, traditional diets, and geographic environments.4 The algorithms derived from primarily European and North American participants may not generate accurate food scores for South Asian, East Asian, African, or Latin American individuals whose microbiome compositions diverge from the training population. PREDICT 3, if conducted, would need substantial geographic and ethnic diversity to address this.
Dietary intake was measured by self-report diary — a method subject to recall bias, underreporting of energy-dense foods, and systematic errors in portion size estimation. Even with a dedicated app, free-living dietary assessment captures approximately 70–80% of actual intake accurately in research participants; less in real-world consumer use. The models trained on PREDICT data are therefore models of metabolic response to reported diet, not measured diet — a distinction that matters for the precision of individual food scores.4
CGM accuracy in non-diabetic individuals with glucose ranges below 100 mg/dL is lower than the device manufacturers’ published specifications, which are derived from clinical testing in hyperglycaemic participants. In the euglycaemic range (80–130 mg/dL) where PREDICT’s non-diabetic participants primarily operated, individual CGM readings carry ±15 mg/dL uncertainty at minimum — a limitation that the PREDICT team acknowledged and partially addressed by using AUC (area under the curve) rather than peak glucose values as the primary metabolic outcome, averaging over many readings to reduce point-wise noise.3
PREDICT also did not measure portal vein metabolites or hepatic triglyceride flux directly, leaving open questions about the precise mechanistic pathways from gut microbiome composition to postprandial lipemia. The correlations between specific microbial taxa and triglyceride responses are reproducible in the dataset but do not establish mechanism — a limitation that matters for designing microbiome-targeted dietary interventions.
References
-
Berry SE, Valdes AM, Drew DA, et al. “Human postprandial responses to food and potential for precision nutrition.” Nature Medicine 26, no. 6 (2020): 964–973. (PREDICT 1 primary publication.)
-
Bermingham KM, Linenberg I, Hall WL, et al. “Gemelli microbiome and ZOE personalized dietary advice study.” Nature Medicine 28 (2022): 2344–2351. (Bermingham PREDICT 2 evaluation and ZOE RCT.)
-
Freckmann G, Pleus S, Grady M, et al. “Continuous Glucose Monitoring in the Context of Nonclinical Applications.” Journal of Diabetes Science and Technology 14, no. 3 (2020): 582–593. (CGM accuracy in euglycaemic ranges.)
-
Dahl WJ, Auger J, Alyousif Z. “Gastrointestinal effects of high-fiber diets: impact on the gut microbiome and cardiometabolic health.” Advances in Nutrition 11, no. 5 (2020): 1064–1082. (Population diversity in microbiome research and dietary response variability.)
Frequently asked questions
- What exactly did the PREDICT studies measure that earlier nutrition research didn't?
- PREDICT 1 collected five simultaneous data streams on 1,002 participants: continuous glucose monitoring, serial postprandial triglyceride blood draws, gut microbiome sequencing, accelerometer-measured activity, and detailed diet diaries. No prior nutrition study combined all five in the same participants at the same time.
- How much of my metabolic response to food is determined by my genes?
- Less than expected. The PREDICT twin analysis found genetics explained only 34 % of postprandial triglyceride variance and 49 % of glucose variance. The remaining 51–66 % was attributable to gut microbiome, habitual diet, sleep, and lifestyle factors — all of which are modifiable.
- Why are postprandial triglycerides important if my fasting lipid panel looks normal?
- Postprandial triglyceride responses were at least as variable and as predictive of cardiometabolic risk as glucose responses in PREDICT, yet are invisible to CGMs. High-fat meals that produce modest glucose spikes can still generate large triglyceride excursions, a distinct cardiovascular risk factor that fasting lipid tests frequently miss.
- Does eating the same meal earlier in the day really produce a lower glucose spike?
- Yes. PREDICT 1 found that identical standardised meals consumed after 5 PM produced postprandial glucose AUC approximately 18 % higher and triglyceride AUC approximately 25 % higher than the same meals eaten before noon, driven by the circadian decline in peripheral insulin sensitivity through the day.
- What are the main limitations of applying PREDICT findings to non-Western populations?
- PREDICT cohorts were predominantly white, educated, UK- or US-based adults. Gut microbiome composition differs substantially across ancestries and traditional diets, so food scores derived from these cohorts may be inaccurate for South Asian, East Asian, African, or Latin American individuals whose microbiomes diverge from the training population.