Updated 24th June 2024

ZOE’s PREDICT studies: What we’ve learned

Share this article

  • Share on Facebook
  • Share on Twitter
  • Print this page
  • Email this page

The ZOE product is based on rigorous research. That’s something we’re very proud of, and we want to provide healthcare professionals insight into our process.

In this article, we’ll outline the original research that forms the core of ZOE’s approach to personalised nutrition: the PREDICT programme.

PREDICT is a series of clinical trials led by scientists from Massachusetts General Hospital, King’s College London, Stanford Medicine and the Harvard T.H. Chan School of Public Health.

These studies helped us quantify and predict individual variations in postprandial responses to standardised meals in a real-world setting. At the same time, we gathered a wealth of data on how lifestyle factors influence these responses. 

Why personalise?

Links between diet and health are well-established, and studies show that metabolic responses to food are related to long-term health outcomes

However, while standard nutrition guidelines are based on averages, we know that there’s significant between-person variability in responses to food. 

A 12-month weight loss intervention study comparing a low-fat and a low-carbohydrate diet underlines these individual differences. 

Overall, there was no difference in the average weight loss between the two groups. Interestingly, though, in both groups, some people lost a significant amount while others gained weight.

So, focusing on a more personalised way to address nutrition is vital if we wish to support all individuals successfully.

Understanding and predicting individual postprandial lipid and glucose responses is an important component of personalised nutrition recommendations. This is what we set out to achieve in the PREDICT studies.

PREDICT overview

Before we outline the results from PREDICT 1, here’s a brief overview of the PREDICT studies to date:


Participants: 1,102 healthy individuals, including 230 twin pairs from the TwinsUK registry.

Aim: To explore how genetics, metabolic differences, the gut microbiome, meal context, meal composition and individual characteristics affect postprandial responses.


Participants: As above, plus another 900 from the TwinsUK registry.

Aim: Similar to PREDICT 1, this was carried out to help us refine our prediction models. We also investigated lipemic dose-response and the effects of meal order on glycemic and lipemic responses in more detail. 


Participants: A subset of 100 highly compliant PREDICT 1 participants.

Aim: To compare the glycemic responses elicited by different carbohydrate sources within and between individuals. The study also assessed glycemic response to a high-carb snack eaten following various non-nutritive sweeteners. We also investigated the influence of meal sequence and timing on glycemia.


Participants: 50 female participants from PREDICT 1 aged 55 or older who we predicted to be either low or high postprandial responders to dietary fat.

Aim: We gathered data on cardiometabolic risk measures to better understand the link between postprandial metabolic responses and cardiometabolic disease. This included measures of vascular function and atherosclerosis. We also measured liver fat using MRI.


Participants: 987 participants from almost every state in the U.S.

Aim: Unlike PREDICT 1, we conducted this study remotely so we could recruit a diverse range of individuals. It allowed us to further understand postprandial responses to dietary intake and the effects of meal sequence and time of day. 


Participants: 3,000 participants

Aim: Designed to further refine our machine learning models to predict individual responses to foods.

These studies combined allowed us to gather more in-depth data than had ever been possible in nutrition studies. 

We could dig into the many features of the postprandial response and pinpoint which factors are linked to these features and their relationship with health outcomes.

Importantly, our use of novel technologies and high-quality dietary assessment methods allows for both scale and precision.

Using our machine-learning models, we can combine data from these studies to help us better understand individual responses to any food or meal and provide personalised food recommendations.

PREDICT 1 in more depth

In PREDICT 1, we wanted to understand the between-person variability in triglyceride, glucose and insulin responses to eight standardised test meals. Our primary goal was to derive algorithms to predict an individual’s postprandial metabolic responses to specific foods.

We explored how various factors influenced these responses, including genetics, the gut microbiome, meal composition, age, sex and BMI. We also took meal context – exercise, sleep, meal ordering and time of day – into account.

Our results were published in Nature Medicine in 2021.

On day 1, the participants consumed two carefully designed meals, which varied in macronutrients. The scientists measured metabolic responses to these meals.

Then, in the 13 days that followed, participants ate more of our specially designed meals. They monitored blood lipids with a dried blood spot assay and glucose responses using a continuous glucose monitor (CGM).

Each participant wore sleep and physical activity monitors throughout the study and provided a stool sample for gut microbiome profiling.

The findings

Below, we summarise some of the major findings from PREDICT 1.


Firstly, we found that genes only partly explain differences in metabolic responses to food. Even identical twins often had very different responses to the exact same meal.

Genes accounted for 30% of the variation in glucose responses and just 4% of the variance in triglycerides. 

Intraindividual consistency

Broadly, each individual had similar responses to identical meals when consumed on different days. Some experienced only relatively small responses to all meals, while others had large postprandial excursions after most meals.

This is important, as it suggests that once we’ve learned about an individual’s postprandial responses to specific foods, we can infer their responses to other foods. 

Gut microbiome

Interestingly, for postprandial lipemia, the gut microbiome explained a greater proportion of the variance than the macronutrient composition of the meal. 

Overall, the gut microbiome accounted for 6.4% of postprandial glucose response variation and 7.5% of postprandial triglyceride response variation. 

Meal composition

Meal composition had a significant impact on postprandial responses. For instance, as fat, fibre and protein content increased, blood glucose responses decreased.

However, we also showed that meal timing, exercise and sleep had just as great an impact on postprandial responses as macronutrient composition.

Timing counts

To investigate the importance of meal timing, we compared glycemic responses to the same meal eaten at breakfast and lunch.

We found that individuals had, on average, a 2-fold higher glycemic response to the same meal at lunch – although there was significant individual variation.

Translating the data

Importantly, using machine learning, we could predict an individual’s postprandial triglyceride and glycemic responses by taking into account meal composition, habitual diet, meal context, anthropometry, genetics, microbiome, and clinical and biochemical parameters. 

Wrapping up

The results from PREDICT 1 form the basis of the ZOE product. And the later studies in the PREDICT series helped us firm up and improve our predictions.

So far, we’ve collected more than 4 million blood glucose recordings, 56,000 triglyceride readings and 12 terabytes of gut microbiome data.

And this is just the beginning. We continue to amass data on postprandial responses to food and the gut microbiome. Over time, our predictions and recommendations can only improve.

Share this article

  • Share on Facebook
  • Share on Twitter
  • Print this page
  • Email this page