Efficacy of federated learning on genomic data: a study on the UK Biobank and the 1000 Genomes Project

Combining training data from multiple sources increases sample size and reduces confounding, leading to more accurate and less biased machine learning models. In healthcare, however, direct pooling of data is often not allowed by data custodians who are accountable for minimizing the exposure of sensitive information. Federated learning offers a promising solution to this problem by training a model in a decentralized manner thus reducing the risks of data leakage. Although there is increasing utilization of federated learning on clinical data, its efficacy on individual-level genomic data has not been studied. This study lays the groundwork for the adoption of federated learning for genomic data by investigating its applicability in two scenarios: phenotype prediction on the UK Biobank data and ancestry prediction on the 1000 Genomes Project data. We show that federated models trained on data split into independent nodes achieve performance close to centralized models, even in the presence of significant inter-node heterogeneity. Additionally, we investigate how federated model accuracy is affected by communication frequency and suggest approaches to reduce computational complexity or communication costs.

Genome-wide association studies and polygenic risk score phenome-wide association studies across complex phenotypes in the human phenotype project

BACKGROUND

Genome-wide association studies (GWASs) associate phenotypes and genetic variants across a study cohort. GWASs require large-scale cohorts with both phenotype and genetic sequencing data, limiting studied phenotypes. The Human Phenotype Project is a longitudinal study that has measured a wide range of clinical and biomolecular features from a self-assignment cohort over 5 years. The phenotypes collected are quantitative traits, providing higher-resolution insights into the genetics of complex phenotypes.

METHODS

We present the results of GWASs and polygenic risk score phenome-wide association studies with 729 clinical phenotypes and 4,043 molecular features from the Human Phenotype Project. This includes clinical traits that have not been previously associated with genetics, including measures from continuous sleep monitoring, continuous glucose monitoring, liver ultrasound, hormonal status, and fundus imaging.

FINDINGS

In GWAS of 8,706 individuals, we found significant associations between 169 clinical traits and 1,184 single-nucleotide polymorphisms. We found genes associated with both glycemic control and mental disorders, and we quantify the strength of genetic signals in serum metabolites. In polygenic risk score phenome-wide association studies for clinical traits, we found 16,047 significant associations.

CONCLUSIONS

The entire set of findings, which we disseminate publicly, provides newfound resolution into the genetic architecture of complex human phenotypes.

FUNDING

E.S. is supported by the Minerva foundation with funding from the Federal German Ministry for Education and Research and by the European Research Council and the Israel Science Foundation.

GRAPE: genomic relatedness detection pipeline

Classifying the degree of relatedness between pairs of individuals has both scientific and commercial applications. As an example, genome-wide association studies (GWAS) may suffer from high rates of false positive results due to unrecognized population structure. This problem becomes especially relevant with recent increases in large-cohort studies. Accurate relationship classification is also required for genetic linkage analysis to identify disease-associated loci. Additionally, DNA relatives matching service is one of the leading drivers for the direct-to-consumer genetic testing market. Despite the availability of scientific and research information on the methods for determining kinship and the accessibility of relevant tools, the assembly of the pipeline, which stably operates on a real-world genotypic data, requires significant research and development resources. Currently, there is no open source end-to-end solution for relatedness detection in genomic data, that is fast, reliable and accurate for both close and distant degrees of kinship, combines all the necessary processing steps to work on a real data, and is ready for production integration. To address this, we developed GRAPE: Genomic RelAtedness detection PipelinE. It combines data preprocessing, identity-by-descent (IBD) segments detection, and accurate relationship estimation. The project uses software development best practices, as well as Global Alliance for Genomics and Health (GA4GH) standards and tools. Pipeline efficiency is demonstrated on both simulated and real-world datasets. GRAPE is available from: https://github.com/genxnetwork/grape.

BREAst Cancer Personalised NuTrition (BREACPNT): dietary intervention in breast cancer survivors treated with endocrine therapy - a protocol for a randomised clinical trial

INTRODUCTION

Breast cancer survivors treated with adjuvant endocrine therapy commonly experience weight gain, which has been associated with low adherence to therapy and worse breast cancer prognosis. We aim to assess whether a personalised postprandial glucose targeting diet will be beneficial for weight management as compared with the recommended Mediterranean diet in this patient population METHODS AND ANALYSIS: The BREAst Cancer Personalised NuTrition study is a phase-2 randomised trial in hormone receptor positive patients with breast cancer, treated with adjuvant endocrine therapy. The study objective is to assess whether dietary intervention intended to improve postprandial glycaemic response to meals results in better weight and glycaemic control in this population as compared with the standard recommended Mediterranean diet. Consenting participants will be assigned in a single blinded fashion to either of two dietary arms (Mediterranean diet or an algorithm-based personalised diet). They will be asked to provide a stool sample for microbiome analysis and will undergo continuous glucose monitoring for 2 weeks, at the initiation and termination of the intervention period. Microbiome composition data will be used to tailor personal dietary recommendations. After randomisation and provision of dietary recommendations, participants will be asked to continuously log their diet and lifestyle activities on a designated smartphone application during the 6-month intervention period, during which they will be monthly monitored by a certified dietitian. Participants' clinical records will be followed twice yearly for 5 years for treatment adherence, disease-free survival and recurrence.

ETHICS AND DISSEMINATION

The study has been approved by the ethics committee in the Sheba medical centre (file 5725-18-SMC, Ramat Gan, Israel) and the Weizmann Institutional Review Board (file 693-2, Rehovot, Israel). The findings of this study will be published in a peer reviewed publication.

TRIAL REGISTRATION NUMBER

NCT04079270.

The Gut Microbiome of Adults With Type 1 Diabetes and Its Association With the Host Glycemic Control

OBJECTIVE

Previous studies have demonstrated an association between gut microbiota composition and type 1 diabetes (T1D) pathogenesis. However, little is known about the composition and function of the gut microbiome in adults with longstanding T1D or its association with host glycemic control.

RESEARCH DESIGN AND METHODS

We performed a metagenomic analysis of the gut microbiome obtained from fecal samples of 74 adults with T1D, 14.6 ± 9.6 years following diagnosis, and compared their microbial composition and function to 296 age-matched healthy control subjects (1:4 ratio). We further analyzed the association between microbial taxa and indices of glycemic control derived from continuous glucose monitoring measurements and blood tests and constructed a prediction model that solely takes microbiome features as input to evaluate the discriminative power of microbial composition for distinguishing individuals with T1D from control subjects.

RESULTS

Adults with T1D had a distinct microbial signature that separated them from control subjects when using prediction algorithms on held-out subjects (area under the receiver operating characteristic curve = 0.89 ± 0.03). Linear discriminant analysis showed several bacterial species with significantly higher scores in T1D, including Prevotella copri and Eubacterium siraeum, and species with higher scores in control subjects, including Firmicutes bacterium and Faecalibacterium prausnitzii (P < 0.05, false discovery rate corrected for all). On the functional level, several metabolic pathways were significantly lower in adults with T1D. Several bacterial taxa and metabolic pathways were associated with the host's glycemic control.

CONCLUSIONS

We identified a distinct gut microbial signature in adults with longstanding T1D and associations between microbial taxa, metabolic pathways, and glycemic control indices. Additional mechanistic studies are needed to identify the role of these bacteria for potential therapeutic strategies.

Prediction of Personal Glycemic Responses to Food for Individuals With Type 1 Diabetes Through Integration of Clinical and Microbial Data

OBJECTIVE

Despite technological advances, results from various clinical trials have repeatedly shown that many individuals with type 1 diabetes (T1D) do not achieve their glycemic goals. One of the major challenges in disease management is the administration of an accurate amount of insulin for each meal that will match the expected postprandial glycemic response (PPGR). The objective of this study was to develop a prediction model for PPGR in individuals with T1D.

RESEARCH DESIGN AND METHODS

We recruited individuals with T1D who were using continuous glucose monitoring and continuous subcutaneous insulin infusion devices simultaneously to a prospective cohort and profiled them for 2 weeks. Participants were asked to report real-time dietary intake using a designated mobile app. We measured their PPGRs and devised machine learning algorithms for PPGR prediction, which integrate glucose measurements, insulin dosages, dietary habits, blood parameters, anthropometrics, exercise, and gut microbiota. Data of the PPGR of 900 healthy individuals to 41,371 meals were also integrated into the model. The performance of the models was evaluated with 10-fold cross validation.

RESULTS

A total of 121 individuals with T1D, 75 adults and 46 children, were included in the study. PPGR to 6,377 meals was measured. Our PPGR prediction model substantially outperforms a baseline model with emulation of standard of care (correlation of R = 0.59 compared with R = 0.40 for predicted and observed PPGR respectively; P < 10-10). The model was robust across different subpopulations. Feature attribution analysis revealed that glucose levels at meal initiation, glucose trend 30 min prior to meal, meal carbohydrate content, and meal's carbohydrate-to-fat ratio were the most influential features for the model.

CONCLUSIONS

Our model enables a more accurate prediction of PPGR and therefore may allow a better adjustment of the required insulin dosage for meals. It can be further implemented in closed loop systems and may lead to rationally designed nutritional interventions personally tailored for individuals with T1D on the basis of meals with expected low glycemic response.

Effects of personalized diets by prediction of glycemic responses on glycemic control and metabolic health in newly diagnosed T2DM: a randomized dietary intervention pilot trial

BACKGROUND

Dietary modifications are crucial for managing newly diagnosed type 2 diabetes mellitus (T2DM) and preventing its health complications, but many patients fail to achieve clinical goals with diet alone. We sought to evaluate the clinical effects of a personalized postprandial-targeting (PPT) diet on glycemic control and metabolic health in individuals with newly diagnosed T2DM as compared to the commonly recommended Mediterranean-style (MED) diet.

METHODS

We enrolled 23 adults with newly diagnosed T2DM (aged 53.5 ± 8.9 years, 48% males) for a randomized crossover trial of two 2-week-long dietary interventions. Participants were blinded to their assignment to one of the two sequence groups: either PPT-MED or MED-PPT diets. The PPT diet relies on a machine learning algorithm that integrates clinical and microbiome features to predict personal postprandial glucose responses (PPGR). We further evaluated the long-term effects of PPT diet on glycemic control and metabolic health by an additional 6-month PPT intervention (n = 16). Participants were connected to continuous glucose monitoring (CGM) throughout the study and self-recorded dietary intake using a smartphone application.

RESULTS

In the crossover intervention, the PPT diet lead to significant lower levels of CGM-based measures as compared to the MED diet, including average PPGR (mean difference between diets, - 19.8 ± 16.3 mg/dl × h, p < 0.001), mean glucose (mean difference between diets, - 7.8 ± 5.5 mg/dl, p < 0.001), and daily time of glucose levels > 140 mg/dl (mean difference between diets, - 2.42 ± 1.7 h/day, p < 0.001). Blood fructosamine also decreased significantly more during PPT compared to MED intervention (mean change difference between diets, - 16.4 ± 37 μmol/dl, p < 0.0001). At the end of 6 months, the PPT intervention leads to significant improvements in multiple metabolic health parameters, among them HbA1c (mean ± SD, - 0.39 ± 0.48%, p < 0.001), fasting glucose (- 16.4 ± 24.2 mg/dl, p = 0.02) and triglycerides (- 49 ± 46 mg/dl, p < 0.001). Importantly, 61% of the participants exhibited diabetes remission, as measured by HbA1c < 6.5%. Finally, some clinical improvements were significantly associated with gut microbiome changes per person.

CONCLUSION

In this crossover trial in subjects with newly diagnosed T2DM, a PPT diet improved CGM-based glycemic measures significantly more than a Mediterranean-style MED diet. Additional 6-month PPT intervention further improved glycemic control and metabolic health parameters, supporting the clinical efficacy of this approach.

TRIAL REGISTRATION

ClinicalTrials.gov number, NCT01892956.

Personalized Postprandial Glucose Response-Targeting Diet Versus Mediterranean Diet for Glycemic Control in Prediabetes

OBJECTIVE

To compare the clinical effects of a personalized postprandial-targeting (PPT) diet versus a Mediterranean (MED) diet on glycemic control and metabolic health in prediabetes.

RESEARCH DESIGN AND METHODS

We randomly assigned adults with prediabetes (n = 225) to follow a MED diet or a PPT diet for a 6-month dietary intervention and additional 6-month follow-up. The PPT diet relies on a machine learning algorithm that integrates clinical and microbiome features to predict personal postprandial glucose responses. During the intervention, all participants were connected to continuous glucose monitoring (CGM) and self-reported dietary intake using a smartphone application.

RESULTS

Among 225 participants randomized (58.7% women, mean ± SD age 50 ± 7 years, BMI 31.3 ± 5.8 kg/m², HbA_1c, 5.9 ± 0.2% [41 ± 2.4 mmol/mol], fasting plasma glucose 114 ± 12 mg/dL [6.33 ± 0.67 mmol/L]), 200 (89%) completed the 6-month intervention. A total of 177 participants also contributed 12-month follow-up data. Both interventions reduced the daily time with glucose levels >140 mg/dL (7.8 mmol/L) and HbA_1c levels, but reductions were significantly greater in PPT compared with MED. The mean 6-month change in "time above 140" was -0.3 ± 0.8 h/day and -1.3 ± 1.5 h/day for MED and PPT, respectively (95% CI between-group difference -1.29 to -0.66, P < 0.001). The mean 6-month change in HbA_1c was -0.08 ± 0.19% (-0.9 ± 2.1 mmol/mol) and -0.16 ± 0.24% (-1.7 ± 2.6 mmol/mol) for MED and PPT, respectively (95% CI between-group difference -0.14 to -0.02, P = 0.007). The significant between-group differences were maintained at 12-month follow-up. No significant differences were noted between the groups in a CGM-measured oral glucose tolerance test.

CONCLUSIONS

In this clinical trial in prediabetes, a PPT diet improved glycemic control significantly more than a MED diet as measured by daily time of glucose levels >140 mg/dL (7.8 mmol/L) and HbA_1c. These findings may have implications for dietary advice in clinical practice.

10 K: a large-scale prospective longitudinal study in Israel

The 10 K is a large-scale prospective longitudinal cohort and biobank that was established in Israel. The primary aims of the study include development of prediction models for disease onset and progression and identification of novel molecular markers with a diagnostic, prognostic and therapeutic value. The recruitment was initiated in 2018 and is expected to complete in 2021. Between 28/01/2019 and 13/12/2020, 4,629 from the expected 10,000 participants were recruited (46 %). Follow-up visits are scheduled every year for a total of 25 years. The cohort includes individuals between the ages of 40 and 70 years. Predefined medical conditions were determined as exclusions. Information collected at baseline includes medical history, lifestyle and nutritional habits, vital signs, anthropometrics, blood tests results, Electrocardiography, Ankle-brachial pressure index (ABI), liver US and Dual-energy X-ray absorptiometry (DXA) tests. Molecular profiling includes transcriptome, proteome, gut and oral microbiome, metabolome and immune system profiling. Continuous measurements include glucose levels using a continuous glucose monitoring device for 2 weeks and sleep monitoring by a home sleep apnea test device for 3 nights. Blood and stool samples are collected and stored at - 80 °C in a storage facility for future research. Linkage is being established with national disease registries.

Rationale and design of a randomised controlled trial testing the effect of personalised diet in individuals with pre-diabetes or type 2 diabetes mellitus treated with metformin

INTRODUCTION

Metformin and diets aimed at promoting healthy body weight are the first line in treating type 2 diabetes mellitus (T2DM). Clinical practice, backed by clinical trials, suggests that many individuals do not reach glycaemic targets using this approach alone. The primary aim of the Personalised Medicine in Pre-diabetes-Towards Preventing Diabetes in Individuals at Risk (PREDICT) Study is to test the efficacy of personalised diet as adjuvant to metformin in improving glycaemic control in individuals with dysglycaemia.

METHODS AND ANALYSIS

PREDICT is a two-arm, parallel group, single-masked randomised controlled trial in adults with pre-diabetes or early-stage T2DM (with glycated haemoglobin (HbA1c) up to 8.0% (64 mmol/mol)), not treated with glucose-lowering medication. PREDICT is conducted at the Clinical Research Facility at the Garvan Institute of Medical Research (Sydney). Enrolment of participants commenced in December 2018 and expected to complete in December 2021. Participants are commenced on metformin (Extended Release, titrated to a target dose of 1500 mg/day) and randomised with equal allocation to either (1) the Personalised Nutrition Project algorithm-based diet or (2) low-fat high-dietary fibre diet, designed to provide caloric restriction (75%) in individuals with body mass index >25 kg/m². Treatment duration is 6 months and participants visit the Clinical Research Facility five times over approximately 7 months. The primary outcome measure is HbA1c. The secondary outcomes are (1) time of interstitial glucose <7.8 mmol/L and (2) glycaemic variability (continuous glucose monitoring), (3) body weight, (4) fat mass and (5) abdominal visceral fat volume (dual-energy X-ray absorptiometry), serum (6) low-density lipoprotein cholesterol (7) high-density lipoprotein cholesterol and (8) triglycerides concentrations, (9) blood pressure, and (10) liver fat (Fibroscan).

ETHICS AND DISSEMINATION

The study has been approved by the St Vincent's Hospital Human Research Ethics Committee (File 17/080, Sydney, Australia) and the Weizmann Institutional Review Board (File 528-3, Rehovot, Israel). The findings will be published in peer-reviewed open access medical journals.

TRIAL REGISTRATION NUMBER

NCT03558867; Pre-results.

Comparison of asymptotic confidence sets for regression in small samples

In case of small samples, asymptotic confidence sets may be inaccurate, with their actual coverage probability far from a nominal confidence level. In a single framework, we consider four popular asymptotic methods of confidence estimation. These methods are based on model linearization, F-test, likelihood ratio test, and nonparametric bootstrapping procedure. Next, we apply each of these methods to derive three types of confidence sets: confidence intervals, confidence regions, and pointwise confidence bands. Finally, to estimate the actual coverage of these confidence sets, we conduct a simulation study on three regression problems. A linear model and nonlinear Hill and Gompertz models are tested in conditions of different sample size and experimental noise. The simulation study comprises calculation of the actual coverage of confidence sets over pseudo-experimental datasets for each model. For confidence intervals, such metrics as width and simultaneous coverage are also considered. Our comparison shows that the F-test and linearization methods are the most suitable for the construction of confidence intervals, the F-test - for confidence regions and the linearization - for pointwise confidence bands.

Analysis of the efficacy of SGLT2 inhibitors using semi-mechanistic model

The Renal sodium-dependent glucose co-transporter 2 (SGLT2) is one of the most promising targets for the treatment of type 2 diabetes. Two SGLT2 inhibitors, dapagliflozin, and canagliflozin, have already been approved for use in USA and Europe; several additional compounds are also being developed for this purpose. Based on the in vitro IC₅₀ values and plasma concentration of dapagliflozin measured in clinical trials, the marketed dosage of the drug was expected to almost completely inhibit SGLT2 function and reduce glucose reabsorption by 90%. However, the administration of dapagliflozin resulted in only 30–50% inhibition of reabsorption. This study was aimed at investigating the mechanism underlying the discrepancy between the expected and observed levels of glucose reabsorption. To this end, systems pharmacology models were developed to analyze the time profile of dapagliflozin, canagliflozin, ipragliflozin, empagliflozin, and tofogliflozin in the plasma and urine; their filtration and active secretion from the blood to the renal proximal tubules; reverse reabsorption; urinary excretion; and their inhibitory effect on SGLT2. The model shows that concentration levels of tofogliflozin, ipragliflozin, and empagliflozin are higher than levels of other inhibitors following administration of marketed SGLT2 inhibitors at labeled doses and non-marketed SGLT2 inhibitors at maximal doses (approved for phase 2/3 studies). All the compounds exhibited almost 100% inhibition of SGLT2. Based on the results of our model, two explanations for the observed low efficacy of SGLT2 inhibitors were supported: (1) the site of action of SGLT2 inhibitors is not in the lumen of the kidney's proximal tubules, but elsewhere (e.g., the kidneys proximal tubule cells); and (2) there are other transporters that could facilitate glucose reabsorption under the conditions of SGLT2 inhibition (e.g., other transporters of SGLT family).