https://orcid.org/0000-0002-0520-7604
Abstract
The increasing prevalence of diabetes, obesity, and their cardiometabolic sequelae present major global health challenges and highlight shortfalls of current approaches to the prevention and treatment of these conditions. Representing the largest global burden of morbidity and mortality, the pathobiological processes underlying cardiometabolic diseases are in principle preventable and, even when disease is manifest, sometimes reversable. Nevertheless, with current clinical and public health strategies, goals of widespread prevention and remission remain largely aspirational. Application of precision medicine approaches that reduce errors and improve accuracy in medical and health recommendations has potential to accelerate progress towards these goals. Precision medicine must also maintain safety and ideally be cost-effective, as well as being compatible with an individual’s preferences, capabilities, and needs. Initial progress in precision medicine was made in the context of rare diseases, with much focus on pharmacogenetic studies, owing to the cause of these diseases often being attributable to highly penetrant single gene mutations. By contrast, most obesity and type 2 diabetes are heterogeneous in aetiology and clinical presentation, underpinned by complex interactions between genetic and non-genetic factors. The heterogeneity of these conditions can be leveraged for development of approaches for precision therapies. Adequate characterization of the heterogeneity in cardiometabolic disease necessitates diversity of and synthesis across data types and research methods, ideally culminating in precision trials and real-world application of precision medicine approaches. This State-of-the-Art Review provides an overview of the current state of the science of precision medicine, as well as outlining a roadmap for study designs that maximise opportunities and address challenges to clinical implementation of precision medicine approaches in obesity and diabetes.
Approaches to medicine and public health: from conventional population strategies through precision, personalized, and genomic medicine.
Introduction
Precision medicine aims to reduce errors and improve accuracy in medical and health recommendations by using information about the individual and their circumstances to optimize the prediction, prevention, diagnosis, and/or treatment of disease.1 Intended to deliver optimal healthcare to the right person at the right time, every time, precision medicine has potential to drive improvements in many aspects of healthcare and medicine. This includes tailoring prevention modalities and treatments using person-level information, as well as improving accuracy and precision in diagnostics and prognostics.
Clinical applications of precision medicine are currently best known in the case of rare diseases, for which genomics is often integral. In monogenic diseases, for example, such as some forms of cystic fibrosis and various cancers, for which driver mutations have been established, precision medicine owes much to the Human Genome Project.2 However, genomics has so far proven much less impactful in the prediction, prevention and treatment of common complex diseases, such as common obesity and diabetes, than first anticipated.3 In these diseases, precision medicine is a powerful tool that harnesses the inherent heterogeneity in clinical presentation, aetiology and treatment response. This may involve information about a person’s genomic and other molecular drivers, although other basic information can also be used to derive successful precision medicine solutions.
Underweight, healthy weight, overweight and obesity are defined by body mass index (BMI) thresholds, with obesity generally considered to be BMI ≥30 kg/m2 in populations of European ancestry. The thresholds for clinical weight classifications determined by BMI and risk of T2D vary across populations with different ethnicities.4 Despite its widespread adoption in public health and clinical practice, BMI is an imprecise and inaccurate estimate of adiposity, which contributes to the heterogeneous clinical presentation and risk profile for T2D and other obesity-related comorbidities. Nevertheless, even after accounting for measurement error and hereditary factors, the disease risk profile associated with excess adiposity is highly variable5 (see Text Box 1).
While disease heterogeneity can complicate the interpretation of risk factors and treatment effects, it is the foundation for precision medicine. In the ‘palette model’ of diabetes, McCarthy6 used the metaphor of the artist's palette to explain how processes interact to determine an individual’s path to diabetes. The model dictates that in some instances there are simple, yet powerful drivers of the disease, which are likened to the bright hues around the edges of the palette; yet for many other people, diabetes emerges through more complex processes that are much harder to map, involving the interplay of numerous biological and environmental factors, which are likened to the colours mixed in the middle of the palette. Understanding and delineating heterogeneity in the palette of diabetes and obesity is essential to developing precision medicine studies with potential for clinical translation.
The cardinal diagnostic feature of diabetes is chronically elevated blood glucose concentrations, with an estimated 96% of the world’s 521 million diabetes cases being T2D.7 The remaining proportion is primarily type 1 diabetes, an autoimmune disease, as well as several very rare manifestations caused by single gene mutations (monogenic diabetes). Diabetes can also result from damage to the pancreas caused by blunt force trauma,8 toxicity (induced by checkpoint inhibitors, for example9) or surgery; however, these forms of diabetes are uncommon. There is currently no definitive diagnostic test for T2D; rather, a diagnosis is reached after all other explanations for chronic hyperglycaemia have been excluded. Accordingly, a diagnosis of T2D is often idiopathic. Current clinical practice guidelines for T2D10 emphasize the individualization of diabetes care, ensuring the healthcare recipient is at the centre of the process. Hence, tailoring therapies (surgical, behavioural, and pharmaceutical) to the individual is key to success, with success being determined by the extent to which complications are prevented and quality of life optimized. This is, essentially, the ‘art of medicine’. The tailoring of healthcare in this way occurs at the patient-practitioner interface and is quite different from the concepts of precision or personalized medicine, which should be implemented earlier in the clinical process and are more automated, measurable, and scalable.
The inherent heterogeneity in T2D diagnoses and clinical presentation combined with variation in age of onset are major barriers to preventing and managing T2D. The aetiology of T2D is also heterogeneous, with complex interactions between genetic and environmental factors acting on a diversity of systemic processes that drive hyperglycaemia. These processes usually involve multiple tissues and organs,11 and may begin at conception.12 These factors combined make precision medicine approaches attractive options for improving clinical outcomes for people with obesity and/or T2D, with the potential for precision interventions in pregnancy.13–16
There exists a vast literature on precision diabetes medicine, which has been systematically reviewed, published as a series of 15 papers13–26 and summarized in the 2nd International Consensus Report on Precision Diabetes Medicine.27 Rather than reiterate this effort, this State-of-the-Art Review provides a nuanced overview of the current state of development of precision medicine research in obesity and diabetes, as well as highlighting opportunities and pitfalls for clinical implementation of precision medicine approaches in these heterogeneous disorders. We provide a step-by-step guide to leveraging heterogeneity in study design for translation of precision medicine solutions into clinical settings (Figure 1).
Step-by-step guide to developing precision medicine solutions for clinical translation. 1. Determine the clinical practice setting(s) where the approach will eventually be deployed; 2. Select appropriate tests to detect and characterize heterogeneity, determining whether signal is sufficient for statistical and/or clinical purposes. 3. Identify valid and reliable biomarkers (tags) of the underlying signal. Determine if the biomarker is causal (only essential if an intervention target). 4. Determine how well the prediction model performs in other populations; 5. Benchmark the prediction model against current standards, ensuring appropriate statistical tests are performed (e.g. prediction accuracy, sensitivity/specificity, reclassification); 6. Assess the scalability of the approach, considering multiple factors
Precision medicine
The term ‘precision medicine’ was coined in the late 1970s in the field of acupuncture medicine28 but has since been used extensively across many medical disciplines. Oncology provides some of the most tractable examples of precision medicine in clinical practice. In breast cancer, for example, four intrinsic molecular subtypes of breast cancer (luminal A, luminal B, HER2-enriched and basal-like) have been identified on the basis of gene expression patterns.29 These intrinsic subtypes provide increased biological and prognostic information beyond that afforded by the traditional pathology-based markers. Investigation of the intrinsic molecular subtypes within specific groups, such as HR-positive tumours, demonstrates differences in overall survival, as well as sensitivities to specific therapies.30 Additionally, about 50% of human cancers are caused by mutations in the tumour suppressor protein p53 gene, resulting in uncontrolled cell proliferation and tumorigenesis. There are at least 10 drugs that target p53 mutations currently approved by the FDA for cancer therapeutics, with more in the research pipeline. These drugs function to either restore p53 function or stabilize the wildtype form of the protein, to degrade mutate or missense p53 in cells, or to induce death in cells with mutant p53.31
‘Precision medicine’ is often used interchangeably with other terms such as personalized medicine, individualized medicine, genomic medicine, and stratified medicine. Rather than view these as synonyms, it is more constructive to consider them as components of precision medicine, each having different roles in optimization processes (see: Graphical abstract). Iterative consensus work on definitions and standardization of nomenclature in diabetes precision medicine is described in detail elsewhere.27,32
In monogenic disorders, equating ‘genomic’ with ‘precision’ medicine is logical, as the primary drivers of monogenic disorders are genetic mutations. Unlikely polygenic disorders, where the effects of genes can be modified through therapeutic interventions, for monogenic disorders there is currently little or no scope for disease prevention, due to the highly penetrant nature of the mutations.33 Hence, most precision medicine for monogenic disorders is genomic medicine, which can be powerful for disease diagnosis and treatment optimization, such as gene therapies.34
Heterogeneity in obesity and T2D arises because their aetiologies are multifactorial, with both polygenic and non-genetic drivers6 (see Text Box 2). While a person’s genotype may raise risk of disease, modifiable environmental factors are usually the catalysts11; thus, intervening on these modifiable factors can help delay or prevent progression to disease even when the genetic load is relatively high.36 There are numerous examples of precision medicine where genetic data plays little or no role in the algorithms used for precision diagnostics,39 prevention,40 treatment,41,42 and prognosis.43 Hence, conflating ‘precision’ and ‘genomic’ medicine in complex disease scenarios can be misleading, as the inclusion of genetic data is not always informative.
Precision medicine approaches are useful not only for stratification of population subsets for potential intervention. They can also be leveraged to ascertain new insights into disease pathogenesis. For example, at a population-level, obesity and T2D appear tightly coupled; however, with the exception of populations with underweight, T2D risk rises linearly with increasing body mass,35 and most people with T2D have overweight or obesity at the time their diabetes is diagnosed. Moreover, intervention trials (lifestyle, pharmacotherapy, and bariatric surgery) in people with pre-diabetes show that risk of diabetes can be substantially reduced by losing weight.36 Nevertheless, about 1:10 people with a healthy BMI develop T2D and about 7:10 people with obesity remain diabetes-free.37
By focusing on scenarios and population subgroups where obesity and T2D risk are discordant, novel insights into obesity and diabetes pathophysiology can be obtained. For example, we recently described an approach using discordant polygenic risk scores to study the pathophysiological features that distinguish obesity and diabetes concordance (SNPs associated with both higher risk of T2D and higher BMI) from discordance (SNPs associated with lower risk of T2D and higher BMI).38 PRSs characterizing the concordant and discordant profiles were used to elucidate distinct clinical and molecular phenotypic patterns linked to fat distribution, liver metabolism, blood pressure, lipid metabolites, and blood levels of proteins involved in extracellular matrix remodelling, as well as premature cardiovascular mortality. Causal inference analyses highlighted the causal roles of excess visceral adiposity, elevated blood pressure, and elevated cholesterol content of high-density lipoprotein particles in the development of T2D in obesity. Furthermore, 17 genetic loci from the discordant signature were identified as possible T2D drug targets.
Precision medicine should not be viewed as replacing or superseding contemporary evidence-based medicine. Rather, contemporary medicine can be used as the foundation upon which key aspects of precision medicine, such as probabilistic scoring or stratification, can be developed, so that more precise and accurate predictions related to risk factors, treatment effects, and disease prognosis can be derived. There are two levels of personalization: the first relates to features of the individual that can be objectively assessed and quantified. This may include readouts from continuous blood glucose, heart rate, or blood pressure monitors, for example. The second level relates to an individual’s relevant subjective characteristics that cannot be meaningfully quantified such as a person’s preferences, capabilities, and attitudes toward health advice and/or interventions. The latter feature cuts across the translational process to ensure any concomitant medical decisions, interventions or health recommendations are those that are most compatible for the recipient.
Some of the most tractable examples of precision medicine in cardiometabolic diseases are those associated with diagnosis and treatment of conditions in which specific pathogenic pathways have been identified using genetics, most notably monogenic disorders. The most common monogenic disorders of this type include: (i) familial hypercholesterolemia (FH), an autosomal dominant genetic disorder characterized by elevated blood concentrations of low-density lipoprotein cholesterol (LDL-C) caused by molecular defects in LDL particle clearance44; (ii) maturity-onset diabetes of the young (MODY), with first onset usually in adolescence or early adulthood. Of the 40 established forms of monogenic diabetes,23 about a dozen are MODY, which collectively account for 1%–6% of all diabetes cases in Europe and the US45; (iii) mutations in the melanocortin 4 receptor gene (MC4R) are the most common form of monogenic obesity, affecting roughly 3:1000 people.46 In some instances, genomic-based diagnoses have proven useful in targeting therapies. For example, insulin therapy is often initially prescribed to people with monogenic diabetes, yet with successful diagnosis for KATP neonatal diabetes, HNF1A-diabetes and HNF4A-diabetes, substantial improvements in blood glucose control can be achieved by switching to sulfonylurea agents.23 In monogenic obesities, the MC4R-agonist, setmelanotide, is approved for treating pro-opiomelanocortin, proprotein convertase subtilisin/kexin type 1, and leptin receptor deficiencies, as well as Bardet-Biedl syndrome, Alström syndrome, and Prader-Willi syndrome.47 Nevertheless, genetic diagnoses do not always indicate specific treatments. This is true of heterozygous FH, which, like other common forms of dyslipidaemia, is treated-to-target with LDL-lowering agents.
Tools of precision medicine
Initial characterization of molecular heterogeneity in obesity and T2D was driven by advances in genetics and the emergence of massively-parallel array-based sequencing technologies applied to relatively small datasets. Contemporary genome-wide association studies (GWAS) now often include data from hundreds of thousands, sometimes millions, of participants, and many large studies are performed using more granular genome (and other molecular) sequencing technologies.
The advances in human health technologies and research methods that underpinned GWAS have since fuelled a diversity of epidemiological research focused on molecular biomarkers such as proteins, metabolites, and epigenetic marks, as well as metagenomic analyses of gut microbiota. Array and sequencing technologies are being deployed at scale in cohorts worldwide, generating complex molecular readouts to understand cardiometabolic disease aetiology and treatment response.48 Vertical integration of these multidimensional, participant-level datasets has facilitated the development of many new and powerful prediction models and diagnostic subclassification algorithms that might prove useful for a range of precision medicine applications. However, precision medicine does not necessarily need to leverage high-throughput, often expensive, technologies, as patient stratification can at times be done using more readily available clinical parameters, such as BMI, waist circumference, and estimated glomerular filtration rate, sometimes by juxtaposing these variables to define ‘discordant’ subtypes using machine learning methods.49
Signal and noise in heterogeneity
The presence of heterogeneity within health datasets allows for identification of substructures that might enable patient stratification for improved diagnoses and the optimization of health interventions. However, heterogeneity is almost always a combination of ‘signal’ and ‘noise’, and sometimes purely ‘noise’, which is a major barrier to the successful development of precision medicine.
‘Signal’ reflects causal processes that influence disease manifestation, susceptibility to risk factors, treatment response, and prognosis. Identifying reliable markers of these causal processes is the objective of most research in precision medicine, with considerable emphasis during the past two decades placed on the discovery of omic biomarkers.48 ‘Noise’, on the other hand, reflects any error (differential or non-differential) linked to the generation, interpretation or application of the data derived from these measurements. Observing heterogeneity per se, without accounting for error, may lead investigators to incorrectly motivate a case for precision medicine, potentially wasting valuable resources.
In the case of randomized controlled trials, randomization and masking help prevent confounding by ensuring any potential confounding factors (measurable and unmeasurable, known and unknown) are evenly distributed across the trial’s arms, allowing unbiased comparisons of treatment effects between placebo and intervention. When seeking to establish treatment effect heterogeneity in trials, heterogeneity observed in the placebo arm should be subtracted from the heterogeneity observed in the active treatment arm; as with the assessment of treatment effects, a failure to do so would risk biasing this assessment of heterogeneity.
Most lifestyle and surgical interventions, however, are not amenable to masking and placebo-control trial designs. Take the Diabetes Prevention Program,36 for example, where the median weight losses in the intensive lifestyle intervention, metformin and control arms were 5.6, 2.1, and 0.1 kg, respectively, during the 2.8-years (median) intervention period. While many participants who received the intensive lifestyle intervention achieved the 1-year weight loss goal (5% reduction), those who did not (about 1/3 of the lifestyle intervention group) went on to gain weight over the ensuing years, despite receiving the lifestyle intervention.50 Elsewhere, the HERITAGE trial delivered a 20-week structured exercise intervention, showing improvements in aerobic fitness in many participants, and reductions in fitness in others.51 At first glance, data from these studies suggest that these interventions caused weight gain and reduced fitness in some people, conclusions that defy consensus and warrant scrutiny.
Behavioural compensation may account for some of the apparent treatment response heterogeneity seen in lifestyle intervention trials. In a trial of adults aged 58–73 years undergoing a 10-day endurance training programme, objectively assessed total energy expenditure declined during the intervention period owing to a reduction in physical activity outside the exercise intervention sessions.52 In a study in which ad libitum energy intake was assessed in a metabolic ward multiple times over a two-year period, participants often over-consumed food relative to their normal consumption.53 The extent of over-consumption was highly correlated within-individual (intraclass correlation = 0.90), with similarly high correlations for macro-nutrient intake, a source of bias that is challenging to control without masking and placebo.
In situations where the placebo interventions cannot be masked, careful monitoring of behaviours and appropriate use of these data in subsequent analyses, may help minimize false heterogeneity. Causal inference analyses using genetics as instrumental variables can also be helpful when seeking to partition signal from noise.
Leveraging signal heterogeneity
Statistical tests of heterogeneity (variance heterogeneity tests) have been used in epidemiological settings to identify population subgroups that vary in their genetic susceptibility to weight gain conditional on differences in environmental exposures. One of the top-ranking loci is FTO, a gene well-known to harbour obesogenic variants that interact with lifestyle exposures.54 In a meta-analysis of genome-wide SNP-BMI heterogeneity signals (N = 170 000 participants), the top-ranking variant (SNP rs7202116) localized to FTO.55 No other genome-wide significant signals were detected. A 7% difference in BMI variance between the two homozygous genotypes at rs7202116 was observed, which can be broadly interpreted as the total heterogeneity in BMI attributable to gene-environment interactions at this locus. A genome-wide variance heterogeneity analysis using a more powerful statistical method (Levene’s test) examined heterogeneity in BMI and lipid traits (total cholesterol, LDL-c, HDL-c, triglycerides) in a meta-analysis of up to 44 211 participants’ data.56 Multiple variants showed statistically significant variance heterogeneity, of which an FTO SNP ranked top for BMI heterogeneity. Some of these signals were underpinned by specific environmental modifiers, strongly suggesting that individual variations in risk-factor susceptibility are conditional on genetic characteristics. At an individual SNP level, these heterogeneity signals appear too small to meaningfully affect weight change in lifestyle intervention trials,57 although the effects of multiple SNPs in aggregate may generate a clinically relevant effect.58
An observational analysis of roughly 3000 adults from the USA who had logged physical activity using a smartwatch (summarized as steps per day) examined the interaction between physical activity and a BMI polygenic burden score in relation to incident obesity.59 An 81% higher (P = 3.57 × 10−20) and 43% lower incident obesity risk were reported when comparing the 75 and 25th percentiles of the PRS and step counts per day, respectively, during the five-year follow-up period. The statistical test of interaction (gene score × steps per day) was not statistically significant (P = .37), suggesting these findings may be false positives. Nevertheless, the authors concluded that people with higher polygenic scores (75th percentile) would need to walk about 2280 steps further each day (of 11 020 total) than people with lower polygenic scores (50th percentile) for their risk of incident obesity to be comparable.
Feasibility of precision medicine for T2D has also been explored by examining treatment response heterogeneity and its clinical predictors.60 Meta-regression analysis of clinical trial data (N = 174 placebo-controlled trials, N = 86 940 participants), focusing on variability in glycaemic control post-treatment, showed a slight increase in variability in HbA1c levels after treatment with active drugs compared with placebos, with GLP-1 receptor agonists showing the largest differences. Subsequent analyses61 focused on body weight variability in response to pharmacological treatment in participants with T2D. Studies comparing glucose-lowering drugs to placebo were analysed across 120 RCTs, with 43 663 participants. While slightly larger treatment response heterogeneity was observed in groups receiving active drugs, after statistical adjustment, the difference in body weight variability between treatment and placebo groups was negligible. It was concluded that there is little potential for precision medicine to improve glycaemic control or weight loss in T2D.
A fundamental limitation of these analyses is that most pharmacological RCTs have strict inclusion/exclusion criteria and often undertake a run-in phase to identify participants who are unlikely to adhere to or tolerate the intervention(s), thereby minimizing between-individual differences in phenotypic characteristics that underpin treatment response heterogeneity. Moreover, assessing treatment response variability at the individual participant-level requires that each participant is exposed to both the treatment and control interventions (i.e. randomized cross-over trials),62 which is rarely the case.
Analyses of real-world data derived from populations that are more diverse in clinical characteristics than most trial cohorts reveal more promising results. Analysis of data from the UK Clinical Practice Research Datalink database, treatment response heterogeneity was modelled in patients initiating SGLT2 and DPP-4 inhibitor therapies.63 Separate treatment response heterogeneity prediction models for the two drug classes were built using multiple clinical features (e.g. baseline HbA1c, age, BMI, estimated glomerular filtration rate, and alanine aminotransferase). The model identified a subgroup, representing ∼40% of patients, who responded especially well to SGLT2 inhibitors. These results were used to inform the design of TriMASTER,42 a precision medicine trial of diabetes medications. This trial of 525 adults with T2D tested two hypotheses: (i) that participants with BMI >30 kg/m2 would experience greater glucose lowering with thiazolidinediones compared with DPP4 inhibitors; (ii) participants with an estimated glomerular filtration rate of 60–90 mL/min/1.73 m2 would experience greater glucose lowering with DPP4 inhibitors compared with SGLT2 inhibitors. Despite each of the three drugs achieving similar overall changes in HbA1c, specific subgroups showed differential responses, supporting the pre-specified hypotheses.
Analyses of real-world data63 and TriMASTER42 highlight the power of precision medicine approaches for enhancing treatment efficacy. By contrast, as the post-hoc analyses of clinical trials discussed above60,61 demonstrate, re-analysis of conventional drug trials may be less informative. Despite TriMASTER providing eloquent proof-of-concept, the results are of limited practical value in many high-income settings (such as in the UK, where the trial was based), because prescribing practices have advanced since TriMASTER was designed, and some commonly prescribed drugs today were not tested in the trial. However, in low- and middle-income countries, the TriMASTER patient-stratification concept could transform healthcare.64
Identifying hidden structure within heterogeneous data
There are many ways to identify substructures within datasets that might be leveraged for precision medicine approaches.
The first is a hypothesis-driven approach, where disease stratification is guided by knowledge of the disease process, and combinatorial analyses are performed to establish whether disease risk or other clinical outcomes vary by strata. Latent class mixed-effects modelling using OGTT data obtained in nearly 6000 adults without diabetes65 identified four distinct patterns of glycaemic response that differed in incident diabetes risk and mortality, but not CVD risk. A profile characterized by high 30-min and low 120-min glucose concentrations conveyed a higher risk of developing diabetes (HR 4.1 [95% CI 2.2, 7.6]) than a profile characterized by low 30-min and high 120 min glucose concentrations (HR 1.5 [95% CI 1.0, 2.2]).
The second approach involves the use of variables that are established clinical markers of disease, upon which analyses are performed that help define structures within an index population. To this end, machine learning methods have been used extensively to cluster individuals into subgroups sharing similar characteristics. Ahlqvist et al.39 used k-means hard clustering in data from 8980 patients, many of whom had been recently diagnosed with diabetes. A total of 5 clusters/subtypes were derived from 6 clinical variables (glutamate decarboxylase antibodies, age at diagnosis, BMI, HbA1c, and estimates of β-cell function and insulin resistance). The 5 subtypes were early-onset severe autoimmune diabetes (6.7% of patients), severe insulin-deficient diabetes (17.5% of patients), severe insulin-resistant diabetes (15.3% of patients), mild obesity-related diabetes (21.6% of patients), and mild age-related diabetes (39.1% of patients). The 5 subtypes bore different associations with the development of diabetes complications, achievement of treatment goals, and medication prescriptions, with replication in 3 independent cohorts. A follow-on study exploring genetic drivers of these clusters within the same cohort of patients with diabetes and others without diabetes,66 showed that family history of diabetes and the frequency of specific genetic variants differed across the subgroups.
Around 30 replication analyses of the initial phenotypic clustering work from Ahlqvist et al. have been reported across diverse populations,22,27 illustrating the reproducibility of the approach. However, the interpretation of the study’s findings has generated controversy. While some have implied direct clinical value of this diagnostic subclassification,67 others have highlighted inherent limitations that undermine translation.68 One of the barriers to effective translation of these findings is that many individuals cannot be assigned definitively to a given cluster, meaning that such labelling will likely to be incorrect in many patients. Moreover, even individuals assigned with high probability to a given cluster may migrate to another cluster later in the course of their disease.69 A further limitation of the approach is that it requires variables (e.g. blood insulin concentrations) that are not routinely available in clinical settings and for which assays lack international standardization.
Various alternative clustering methods have been developed to overcome these challenges, some of which use machine learning methods.22 In an analysis of the IMI DIRECT cohort, a soft-clustering method was used to map clinical heterogeneity in T2D to individual aetiological processes. Specifically, disease ‘archetypes’ were defined using 32 clinical variables in ∼800 people with newly diagnosed T2D.70 Quantitative clustering scores were derived and assigned to each participant. These scores capture the complexity of T2D presentation without the need for categorization and were associated with glycaemic deterioration, genetic risk profile, circulating omics biomarkers, and phenotypic stability over 36-months follow-up. One archetype was linked with obesity, insulin resistance, dyslipidaemia, and impaired β-cell glucose sensitivity, and had the fastest disease progression and highest demand for drug therapy. The analysis showed that about a third of the population could be adequately subclassified into archetypes, defining distinct diabetes subtypes. The remainder of the cohort could not be adequately subclassified, highlighting that subclassification of T2D cannot be done accurately in most people with the disease, given the methods and data currently available. Improvements in the quality, depth, and scope of data, as well as innovations in data analysis methods, are likely to strengthen future T2D subclassification attempts.
A third approach is to partition heterogeneity of T2D using genetic data. These analyses focus on mapping gene variants and epigenetic marks to pathophysiological processes that cause T2D; such processes typically span eight organ-specific pathways.71 Using genetics to segment disease is advantageous because a person’s nuclear genome remains unchanged throughout the life-course. Thus, genetic variants are uniquely powerful instruments in causal inference analyses because confounding, common in many association studies of biological variants and disease outcomes, is largely neutralized.
Udler and colleagues72 used Bayesian nonnegative matrix factorization to cluster 94 genetic variants into five process-specific PRSs (pPRSs) associated with distinct tissue-specific enhancer enrichment across 28 cell types, representing mechanistic pathways of disease. The clinical consequences of these pathways were then examined in people with T2D. Two clusters were associated with diminished beta cell function, while the other 3 were characterized by obesity-mediated lipodystrophy-like fat distribution, and liver lipid dysmetabolism, all indicative of cellular insulin resistance. Higher levels of the pPRS were associated with elevated blood pressure, coronary artery disease, and stroke.
Similarly, Mahjahan et al.73 used summary statistics as the basis for soft clustering 94 T2D-associated gene variants to identify multitrait patterns of association, from which 3 core physiological processes were determined. The first cluster comprised nine variants strongly associated with BMI and dyslipidaemia, with 3 novel coding signals (at PNPLA3, POC5, and BPTF) impacting T2D risk through adiposity-mediated pathways. The second cluster comprised 39 variants associated with insulin secretion, while the third category consisted of 23 variants associated with insulin action.
Suzuki et al.74 took aggregated GWAS data from about 400 000 people with and 2.1 million people without T2D, 40% of whom were not of predominantly European ancestry. The analyses extended the set of published loci with statistically robust signals for T2D from 466 to 611 loci, the latter harbouring 1289 genome-wide significant (P < 5 × 10−8) variants. Using machine learning (k-means clustering combined with iterative multiple imputation of missing SNV-phenotype associations), these signals were clustered into 8 sets with distinct associations across 37 cardiometabolic traits and open chromatin regions specific to cells in pancreatic islets, adipocytes, endothelium and enteroendocrine tissue. Thereafter, cluster-specific partitioned polygenic scores were derived within clusters in an independent dataset and tested for associations with T2D-related cardiovascular traits. These analyses yielded pPRS associations with coronary artery disease, peripheral artery disease and end-stage diabetic nephropathy across ancestries.
Although clinical practice often requires disease classification for diagnosis and treatment, information can be lost when continuous variables are categorized, often reducing statistical power. Thus, categorization of variables should be done as late in the data analysis pipeline as possible.75 Machine learning methods have been developed to help minimize the need for diagnostic classification, instead using continuous data. Reverse graph embedding approaches, for example, were used to characterize phenotypic diversity in 23 137 Scottish patients newly diagnosed with T2D. pPRSs, comprised of T2D-associated SNPs, were used to elucidate underlying biological processes. These polygenic scores were used to map the risk of disease progression, focusing on time-to-insulin therapy, and development of chronic kidney disease, referable diabetic retinopathy, and major adverse cardiovascular events. Replication was successfully undertaken in the UK Biobank, and data from ADOPT was used to extend the results to include drug response.41
A major challenge to the clinical evaluation and implementation of precision medicine is that research and methods are poorly standardized, impeding data synthesis and comparison.76 Moreover, health equity impact, patient-public involvement and engagement, and cost-effectiveness are rarely considered, and benchmarking against contemporary standards and approaches are rarely undertaken in the published literature. To address these challenges, we published the BePRECISE reporting guidelines for precision medicine research of clinical relevance.76
Individual responses vs. population averages
Contemporary evidence-based medicine assumes that the average effects of risk factors and therapies at a population level approximate those that can be expected within individuals. Precision medicine, conversely, assumes that there is sufficient heterogeneity in these effects to justify the optimization of interventions to population subgroups who share similar characteristics (Figure 2). Nevertheless, an average effect implies that within the overall population some individuals benefit from the standard medical approach, even though there may be many who do not benefit. Under a Gaussian distribution, half of the population would experience above average treatment effects, and the other half would experience below-average effects. In this setting, one could argue that at least half of those receiving the treatment experience effects ranging from adequate to excellent (i.e. standard medicine approaches are appropriate for this subgroup). In this scenario, a precision medicine solution might be sought only for the below-average subgroup. Establishing the shape of such distributions should be done early in the process of designing precision medicine interventions, as this will help determine stratification cut-points. It is also worth considering that even if there is considerable risk-factor heterogeneity, if the therapy is homogeneous in its effects, there would be scant justification for using precision medicine.
Precision strategies have the potential to deliver more equitable intervention effects compared with conventional population strategies when treatment effects are heterogeneous
Precision medicine in the clinic
Although most progress in precision medicine has been in distinct disease areas, common diseases tend to develop in concert with other diseases. An analysis of population-based claims data from ∼31 million Medicare fee-for-service beneficiaries revealed that one in every two people aged <65-years had at least two chronic health conditions, rising to 62% of those aged 65–74 years, and 81.5% in people aged ≥85 years. The siloization of medicine, where focus is placed primarily on treating diseases in isolation rather than the person’s overall condition, has long been understood to be sub-optimal.77 Nevertheless, most contemporary medicine, both research and practice, focuses on specialization based on categories of disease. This approach has two main limitations: the first is that diseases often interact, particularly with regard metabolic disease, with shared affected pathways and even shared treatments, such as GLP-1 receptor agonists for obesity, heart failure and glycaemic control.78,79 The second challenge relates to information loss, which occurs when arbitrary classifications are imposed on continuous data (e.g. thresholds for blood markers, body corpulence, or blood pressure).
Innovations in precision medicine will need to build on contemporary evidence-based medicine and be embedded into existing medical infrastructures. Even so, the transition towards precision medicine in obesity, T2D and their sequalae presents an opportunity to shift from a disease-centric diagnostic process to one that is initially, at least, disease agnostic. This may involve the use of person-specific molecular and non-molecular data to query biological ‘systems’ to determine function/dysfunction under basal (e.g. rested or fasted) and perturbed (e.g. with metabolic or cognitive challenge) conditions, as well as personalized monitoring of the biological state over extended durations. Mapping a person’s dynamic health state has been made possible through innovative technological developments primarily in: (i) high-throughput molecular phenotyping and genome sequencing; (ii) high-dimensionality computing and AI, and (iii) digital devices including those for imaging and bio-monitoring.80
Conclusion
Precision medicine in obesity and diabetes has enormous, near-term, potential to elucidate and address the global burdens of these conditions. New technologies and data types coupled with widely used, simple clinical and anthropometric parameters can be leveraged to reveal substructures in heterogenous conditions, to increase accuracy and to reduce error in contemporary medicine. Genomics and other high throughput data-types can also be used to further understanding and elucidate pathophysiological mechanisms that underpin obesity and T2D, including interactions between the conditions and identifying new therapeutic targets. Efficient study design in precision medicine research that minimizes bias and confounding within the innate heterogeneity of these conditions is needed to ensure high-fidelity translation of research findings into the clinic.
Supplementary data
Supplementary data are not available at European Heart Journal online.
Declarations
Disclosure of Interest
Within the past 5 years, P.W.F. has received consulting honoraria from Eli Lilly Inc., Novo Nordisk Foundation, Novo Nordisk A/S, UBS, Qatar Foundation, and Zoe Ltd. This work was done outside these roles and does not necessarily reflect the opinions of any of these organizations. P.W.F. has also received investigator-initiated grants (paid to institution) from numerous pharmaceutical companies as part of the Innovative Medicines Initiative of the European Union. Within the past 5 years, J.L.S. has received consulting fees from the World Health Organization and the University of Bergen. This work was done outside these roles, and the opinions expressed herein do not necessarily reflect those of the World Health Organization or the University of Bergen. J.L.S. is the founder of BabelFisk, which provides consulting services in health sciences, public and global health.
Data Availability
No data were generated or analysed for or in support of this paper.
Funding
P.W.F. was supported by grants from the European Commission (ERC-CoG_NASCENT—681742) Swedish Research Council (#2019-01348), and Swedish Foundation for Strategic Research (Stiftelsen for Strategisk Forskning) (LUDC-IRC, 15-0067).
