The PREDICT program (Figure 1) is the largest, in-depth nutritional research program of its kind in the world. PREDICT encompasses a collection of rigorously designed clinical trials led by scientists from Massachusetts General Hospital, King’s College London, Stanford Medicine, and Harvard T.H. Chan School of Public Health.
These studies were designed to allow us to quantify and predict individual variations in postprandial responses to standardized meals in a real-world setting, while also gathering as much data about lifestyle factors as possible. This research has allowed us to explore many different features of the complex postprandial responses to better understand which factors influence them and how they subsequently impact health outcomes.
All of these studies are designed in such a way that their data can be seamlessly combined in our machine learning models to help us better understand individual responses to any food or meal and provide personalized food recommendations.
Overview of the PREDICT studies
To get a sense of the scale of these studies, we've already collected over:
4 million glucose readings
56,000 triglyceride readings
12 terabytes of microbiome data
...and this is just the beginning!
The PREDICT program commenced with the PREDICT 1 study (IRAS 236407; IRB 2018P002078; NCT03479866) (Jun 2018-May 2019). This study was conducted in partnership with King’s College London, Twins UK, and Massachusetts General Hospital.
PREDICT 1 (Figure 2) enrolled 1,102 healthy individuals (60% of the cohort was recruited from the TwinsUK registry, of which 230 were twin pairs), from the UK (n=1,002) and US (n=100) with the primary aim of examining how genetics, metabolic differences, the gut microbiome, meal context (e.g. exercise, sleep, meal ordering, time of day), meal composition, and individual characteristics (e.g. age, sex, BMI) affect postprandial responses to meals.
This groundbreaking nutrition research project has uniquely benefitted from being carried out largely on the twin population from the Twins UK Study (a 25-year investigation of health and lifestyle in over 14,000 twins). By studying twins, for the first time, we were able to disentangle the impact of genetics from other determinants of our responses to food and discovered that genes are not as important as previously believed for predicting our responses to food.
PREDICT 1 was specifically designed to quantify and predict individual variations in postprandial triglyceride, glucose, and insulin to 8 carefully designed standardized test meals. Additionally, we measured variations in metabolomic responses to the two standardized clinic-day test meals and variations in glucose responses to hundreds of thousands of meals eaten at home. The test meals, designed by nutritional experts with decades of experience of conducting postprandial intervention studies, consisted of isocaloric muffins with varying macronutrient composition and standard 75g oral glucose tolerance tests (OGTT).
Never before has a study combined postprandial data gathered in a clinical setting, free-living data for all meals consumed over two weeks, and shotgun sequenced microbiome data for this many individuals. This data, alongside the data from PREDICT 2, has provided core training data for the ZOE scores.
The first findings from this study are published in Nature Medicine.
PREDICT 1 Plus
PREDICT 1 Plus (IRAS 236407; NCT03479866) (Figure 3) is the ongoing, second phase of the PREDICT 1 study that is recruiting a further 900 individuals from the Twins UK registry to undertake a similar protocol to PREDICT 1 to further enhance our prediction models. Building on the learnings from PREDICT 1, protocol modifications in PREDICT 1 Plus allows us to explore the lipemic dose-response and effects of meal order on glycemic and lipemic responses in more depth.
PREDICT 1 sub-studies
A subset of highly compliant PREDICT 1 participants (n=100) was recruited to take part in the entirely remote PREDICT-Carbs study (IRAS 236407; NCT03479866) (Jan-May 2019), which aimed to compare the glycemic responses elicited by different sources of carbohydrate within and between individuals (shown in Figure 4). Participants in this study received a standardized dietary intervention of breakfasts, lunches, and snacks based on various carbohydrate-rich staple foods.
The study also tested glycemic response to a high-carbohydrate snack when eaten after preloads of various non-nutritive sweeteners, to capture inter-individual and inter-sweetener differences. The breakfast-lunch intervention design also allowed us to explore and untangle the influence of meal sequence and time of day on the resulting glycemia.
PREDICT-Cardio (IRAS 236407; NCT03479866) (Oct 2019-Feb 2020) aimed to gather data on intermediary cardiometabolic risk measures to better understand the link between postprandial metabolic responses and cardiometabolic disease. These risk measures (shown in Figure 5) included vascular function (as measured by Pulse Wave Velocity (PWV)) and atherosclerosis (measured by Carotid Intima-Media Thickness (CIMT) and plaque).
We also measured liver fat (measured by MRI), which is central to metabolic dysregulation and is emerging as an important measure of metabolic health that is modulated by diet. This study involved a sub-cohort (n=50) of PREDICT 1 Plus participants who were females, aged >55 years and predicted to be either low or high postprandial responders to dietary fat.
(IRB Pro00033432; NCT03983733)
The PREDICT 2 study (June 2019-March 2020) (Figure 6) was conducted in collaboration with researchers from Massachusetts General Hospital and Stanford University, which aimed to help us further understand postprandial responses to dietary intake as well as their modulation by meal sequence and time of day. Unlike PREDICT 1, this study was designed to be carried out entirely remotely ‘at-home’. This allowed us to focus on recruiting a diverse range of individuals from many geographies and ethnicities and is the most comprehensive at-home nutrition study ever conducted.
A total of 987 volunteers took part in this home-based study from almost every state in the US, in which we demonstrated the efficacy of remote study delivery at an unprecedented depth and scale. This study was able to collect much of the phenotype data that was collected during PREDICT 1, but entirely remotely. The successful delivery of this study has put us at the forefront of remote nutritional research and enabled us to design an at-home testing product for ZOE that is comparable to these cutting edge research studies. Many of the tests in the ZOE product have never previously been available commercially, including the lipid response tests and the detailed output from the microbiome analysis.
The ZOE PREDICT 3 study (NCT04735835) is an ongoing single-arm mechanistic intervention study which commenced in July 2020. The PREDICT 3 study will build on previous research in over 2,000 individuals (PREDICT 1 and 2) to further refine machine learning models that predict individual responses to foods, with the aim of advancing precision nutrition science and individualized dietary advice.
The study incorporates both standardized and controlled dietary intervention, for the purpose of testing postprandial responses to specific mixed meals, in addition to a free-living period with a dietary record for measuring responses to a large variety of meals consumed in a realistic context, where the role of external factors (e.g. exercise, sleep, time of day) on postprandial responses may be determined.
For the first time this PREDICT study is built on top of a commercial product which will allow access to a much larger group of participants who are already collecting large amounts of data through digital and biochemical devices that can contribute to science.
To date more than 45,000 participants have completed PREDICT 3 program.
Data from the PREDICT Studies in more detail
Through the various studies that form part of our PREDICT Program, we have been able to gather more in-depth data than has previously been possible in nutrition studies. This is a result of using novel technologies and high-quality dietary assessment methods, allowing for both scale and precision.
For precision nutrition to be a success, high precision data must be obtained. This enables the discrimination of subtle differences in biomarkers between individuals, using comparable methods so all this data can be fed into machine learning algorithms. To achieve this, we have worked with leading analytical laboratories to refine collection and analysis methodologies, thus improving the precision of data from samples collected remotely and ensuring results are comparable to those collected in controlled, clinical settings.
Berry SE, Valdes AM, Drew DA, et al. Human postprandial responses to food and potential for precision nutrition. Nat Med. 2020;26:964-73. doi: 10.1038/s41591-020-0934-0