The prevalence of type 2 diabetes is becoming epidemic and is expected to increase in the next decades (1). Compelling prevention strategies must, therefore, be implemented. Metformin, one of the possible choices for first-line glucose-lowering therapy for hyperglycemia in type 2 diabetes (2), is also effective in preventing the disease (3). However, as with all drugs, there are patients who do not respond well to metformin (4,5) and/or who report side effects, including gastrointestinal distress (6) and reduced vitamin B12 serum concentration (7). In an ideal world, clinicians would have models that predict how each patient reacts to metformin (or any drug) so that they primarily treat the best responders (i.e., maximize treatment efficacy) and exclude patients at high risk of side effects (i.e., minimize personal costs). Such a precision therapeutics approach would also reduce the economic burden of any treatment.

A step toward precision therapeutics has been made by the genome-wide association study of people with prediabetes (from the Diabetes Prevention Program [DPP]) for metformin response, published in this issue of Diabetes (8). The incidence of diabetes and several 1-year changes in quantitative traits, including HbA1c and body weight, were the outcomes investigated. A very conservative experiment-wide statistical significance threshold (P < 9.0 × 10−9) was used, taking into account both the genome-wide design and several comparisons. Importantly, the gene–treatment interaction was also tested to identify possible associations due specifically to metformin. Unfortunately, no significant findings were observed concerning the incidence of diabetes. Conversely, two variants were significantly associated in the metformin arm, but not the placebo arm, with quantitative traits: rs144322333 was associated with changes in HbA1c (this single nucleotide polymorphism [SNP] is near ENOSF1 in chromosome 19), and rs145591055 was associated with changes in body weight (this SNP is near OMSR in chromosome 5). In addition, in the interaction model, two additional SNPs reached the experiment-wide P value, namely, rs6838493 (near LINC01093 in chromosome 4) and rs148219263 (near CDH20 in chromosome 18), with differential changes in HbA1c and body weight, respectively. None of the genes closest to the four SNPs has a biological function easily correlated with metformin efficacy. Unfortunately, despite several efforts, no replication of the association with HbA1c change was obtained, and no attempt was made with body weight reduction.

The design of the DPP uniquely addresses the pharmacogenetics of metformin efficacy in diabetes prevention and represents one of the strengths of this genome-wide association study. However, such uniqueness also represents a limitation, as attempting to replicate the associations in a second independent diabetes prediction study is not actionable. To try to overcome this obstacle, the authors looked for replication in people with type 2 diabetes, but this makes it possible to hypothesize that the failure to replicate results is due to the different phenotype used rather than pointing to a false-positive result.

As usually happens in pharmacogenomics (9), the sample size is also a limitation of this study, which can lead to false-negative findings, especially when testing gene–treatment interaction, an indispensable approach to identify individual responses to drugs (10). Therefore, we agree with the authors when they indicate the need for large pharmacological studies based on omics data in different populations.

A further limitation concerns the analysis of the DPP sample as a whole, even though data were derived from individuals of multiple ancestries. While this approach was somewhat unavoidable, given the small number of individuals of each ethnicity, it is biased toward variants with concordant effects across different ancestries, thus reducing the likelihood of identifying genetic markers associated with the effect of metformin in particular ethnic groups. Indeed, the findings reported do not appear to be ancestry specific, and some are found only in certain ethnicities as a likely sole consequence of the higher minor allele frequencies of the associated SNPs.

Studies like this are important because they discover markers that could unravel new pathways of a given drug's mechanism of action. Indeed, new biomarkers have recently emerged through omics investigation that are associated with several metformin effects (both efficacy and side effects) in different populations. Among these are genetic variants that are clearly advantageous because they are stable and constant throughout life (11), lipid metabolites (12), molecules from gut microbiota (13), and blood-based epigenetic markers (14). All these discoveries may open new avenues in the understanding of metformin’s mechanisms of action. Epidemiologically, the small individual effects that these or similar markers usually have on treatment response can be exploited to create subgroups of different average responses (e.g., low vs. high), but these effects cannot help to make accurate predictions at the individual level, thus providing little progress in precision therapeutics (15). Indeed, models are needed to implement a precision treatment approach, possibly based on common and readily available clinical variables that predict the benefits or harms of a given treatment in each patient. This effect must be examined objectively by specific measures (16) related to the discriminatory capacity of the different individual effects. In recent years, much attention has been paid to the methodology that aims to provide prediction models that take advantage of patient demographic, clinical, and laboratory characteristics, as well as their interactions, while limiting overly optimistic results and obtaining parsimonious algorithms (17). Methodological guidance for modeling treatment effect heterogeneity was recently published (18). Eventually, omics markers should be used to improve preexisting clinical risk models (19). The crucial difference between identifying subgroups at different average levels of risk and predicting risk at the individual level is detailed in Fig. 1.

Figure 1

Difference between subgroup analysis and individual prediction. To identify individuals with different degrees of drug response, two main prediction approaches based on variables associated with treatment effects can be schematically described as follows. Color intensity indicates drug response (either benefits or harms) from none or very mild (white or light blue) to very strong (deep blue). Subgroup analysis is depicted on the lower left side of the figure. Individuals are categorized into subgroups according to one or a few variables that are associated with the treatment effect (e.g., a priori stratification, tree-based analyses, and/or principal component analyses). While the average response to treatment (either benefits or harms) will be greater in subgroup B than subgroup A, some low responders will be in subgroup B and vice versa (i.e., some high responders will be in subgroup A), thus limiting the usefulness of this approach for prediction purposes. Rather, subgroup analysis is very useful for epidemiological studies aimed at addressing drug mechanisms of action or heterogeneity in the pathophysiology of a given disease. Individual predictions are depicted on the lower right side of the figure. Each patient is assigned a numerical score of predicted response to treatment, derived from many variables associated with the drug effect (e.g., multivariable, stepwise analysis and least absolute shrinkage and selection operator). This approach makes it possible to discriminate between responses to treatments of different magnitudes and therefore is the gold standard for implementing all possible applications of precision medicine (not just precision therapeutics), which, as a whole, is based on optimal risk stratification.

Figure 1

Difference between subgroup analysis and individual prediction. To identify individuals with different degrees of drug response, two main prediction approaches based on variables associated with treatment effects can be schematically described as follows. Color intensity indicates drug response (either benefits or harms) from none or very mild (white or light blue) to very strong (deep blue). Subgroup analysis is depicted on the lower left side of the figure. Individuals are categorized into subgroups according to one or a few variables that are associated with the treatment effect (e.g., a priori stratification, tree-based analyses, and/or principal component analyses). While the average response to treatment (either benefits or harms) will be greater in subgroup B than subgroup A, some low responders will be in subgroup B and vice versa (i.e., some high responders will be in subgroup A), thus limiting the usefulness of this approach for prediction purposes. Rather, subgroup analysis is very useful for epidemiological studies aimed at addressing drug mechanisms of action or heterogeneity in the pathophysiology of a given disease. Individual predictions are depicted on the lower right side of the figure. Each patient is assigned a numerical score of predicted response to treatment, derived from many variables associated with the drug effect (e.g., multivariable, stepwise analysis and least absolute shrinkage and selection operator). This approach makes it possible to discriminate between responses to treatments of different magnitudes and therefore is the gold standard for implementing all possible applications of precision medicine (not just precision therapeutics), which, as a whole, is based on optimal risk stratification.

Close modal

Unfortunately, in the field of type 2 diabetes, there are very few validated models to predict the efficacy of some second- and third-line therapies (20,21), so much so that today most treatments for hyperglycemia, including first-line treatments, are still recommended by expert panels based on clinical considerations (2). Indeed, a couple of models have been proposed to predict the efficacy of metformin in people with diabetes, but these are either simple and straightforward preliminary pilot attempts (22) or are based on complex equations with nonlinear terms that are not easy to interpret and implement clinically (23). More importantly, both models lack validation in independent samples, an inescapable prerequisite for implementing prediction medicine approaches (24).

In conclusion, to make the dream of precision therapeutics in the field of diabetes come true, we need a deeper understanding of how each of the molecules that are currently used works. To this aim, comprehensive omics measurements in large pharmacological studies with samples from different ethnicities will play a role. Concurrently, but ideally before omics studies are completed, we need to develop models that can predict the benefits or side effects of a given drug at the individual level. Because the majority of people with type 2 diabetes live in resource-limited countries, these models should be easy to use and based on inexpensive common variables. New markers emerging from omics (or even non-omics) studies will then be added to these models, with the aim of further improving their predictive ability so that a precision therapeutic approach in diabetes can finally be implemented.

See accompanying article, p. 1161.

Funding. This study was supported by the Italian Ministry of University and Research, “Progetti di Ricerca di Interesse Nazionale” 2015 and PON_2017_BIO-D ARS01_00876 (V.T.), Italian Ministry of Health, Ricerca Corrente 2021–2022 (V.T. and C.M.) and RF-2013-02356459 (C.M.), and European Foundation for the Study of Diabetes/Sanofi grant 2017 (C.M.).

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

1.
International Diabetes Federation
.
IDF Diabetes Atlas
, 10th edition.
Brussels, Belgium
,
International Diabetes Federation
,
2021
.
2.
Davies
MJ
,
Aroda
VR
,
Collins
BS
, et al
.
Management of hyperglycemia in type 2 diabetes, 2022. A consensus report by the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD)
.
Diabetes Care
2022
;
45
:
2753
2786
3.
Knowler
WC
,
Fowler
SE
,
Hamman
RF
, et al.
Diabetes Prevention Program Research Group
.
Ten-year follow-up of diabetes incidence and weight loss in the Diabetes Prevention Program Outcomes Study
.
Lancet
2009
;
374
:
1677
1686
4.
Kahn
SE
,
Haffner
SM
,
Heise
MA
, et al.
ADOPT Study Group
.
Glycemic durability of rosiglitazone, metformin, or glyburide monotherapy
.
N Engl J Med
2006
;
355
:
2427
2443
5.
Zeitler
P
,
Hirst
K
,
Pyle
L
, et al.
TODAY Study Group
.
A clinical trial to maintain glycemic control in youth with type 2 diabetes
.
N Engl J Med
2012
;
366
:
2247
2256
6.
Rena
G
,
Hardie
DG
,
Pearson
ER
.
The mechanisms of action of metformin
.
Diabetologia
2017
;
60
:
1577
1585
7.
de Jager
J
,
Kooy
A
,
Lehert
P
, et al
.
Long term treatment with metformin in patients with type 2 diabetes and risk of vitamin B-12 deficiency: randomised placebo controlled trial
.
BMJ
2010
;
340
:
c2181
8.
Li
JH
,
Perry
JA
,
Jablonski
KA
, et al.
Diabetes Prevention Program Research Group
.
Identification of genetic variation influencing metformin response in a multi-ancestry genome-wide association study in the Diabetes Prevention Program (DPP)
.
Diabetes
2023
;
72
:
1161
1172
9.
McInnes
G
,
Yee
SW
,
Pershad
Y
,
Altman
RB
.
Genomewide association studies in pharmacogenomics
.
Clin Pharmacol Ther
2021
;
110
:
637
648
10.
Murcray
CE
,
Lewinger
JP
,
Conti
DV
,
Thomas
DC
,
Gauderman
WJ
.
Sample size requirements to detect gene-environment interactions in genome-wide association studies
.
Genet Epidemiol
2011
;
35
:
201
210
11.
Florez
JC
.
The pharmacogenetics of metformin
.
Diabetologia
2017
;
60
:
1648
1655
12.
Sharma
A
,
Krick
B
,
Li
Y
, et al
.
The use of ceramides to predict metabolic response to metformin in women with PCOS
.
J Endocr Soc
2022
;
6
:
bvac131
13.
Brunkwall
L
,
Orho-Melander
M
.
The gut microbiome as a target for prevention and treatment of hyperglycaemia in type 2 diabetes: from current human evidence to future possibilities
.
Diabetologia
2017
;
60
:
943
951
14.
García-Calzón
S
,
Perfilyev
A
,
Martinell
M
, et al
.
Epigenetic markers associated with metformin response and intolerance in drug-naïve patients with type 2 diabetes
.
Sci Transl Med
2020
;
12
:
eaaz1803
15.
Dennis
JM
.
Precision medicine in type 2 diabetes: using individualized prediction models to optimize selection of treatment
.
Diabetes
2020
;
69
:
2075
2085
16.
van Klaveren
D
,
Steyerberg
EW
,
Serruys
PW
,
Kent
DM
.
The proposed “concordance-statistic for benefit” provided a useful metric when modeling heterogeneous treatment effects
.
J Clin Epidemiol
2018
;
94
:
59
68
17.
Rekkas
A
,
Paulus
JK
,
Raman
G
, et al
.
Predictive approaches to heterogeneous treatment effects: a scoping review
.
BMC Med Res Methodol
2020
;
20
:
264
18.
Kent
DM
,
Paulus
JK
,
van Klaveren
D
, et al
.
The Predictive Approaches to Treatment effect Heterogeneity (PATH) statement
.
Ann Intern Med
2020
;
172
:
35
45
19.
Steyerberg
EW
,
de Wreede
LC
,
van Klaveren
D
,
Bossuyt
PMM
.
Personalized decision making on genomic testing in early breast cancer: expanding the MINDACT trial with decision-analytic modeling
.
Med Decis Making
2021
;
41
:
354
365
20.
Dennis
JM
,
Young
KG
,
McGovern
AP
, et al.
MASTERMIND Consortium
.
Development of a treatment selection algorithm for SGLT2 and DPP-4 inhibitor therapies in people with type 2 diabetes: a retrospective cohort study
.
Lancet Digit Health
2022
;
4
:
e873
e883
21.
Shields
BM
,
Dennis
JM
,
Angwin
CD
, et al.
TriMaster Study Group
.
Patient stratification for determining optimal second-line and third-line therapy for type 2 diabetes: the TriMaster study
.
Nat Med
2023
;
29
:
376
383
22.
Ludovico
O
,
Farina
MG
,
Copetti
M
, et al
.
ENPP1 mRNA levels in white blood cells and prediction of metformin efficacy in type 2 diabetic patients: a preliminary evidence
.
Nutr Metab Cardiovasc Dis
2012
;
22
:
e5
e6
23.
Goswami
S
,
Yee
SW
,
Xu
F
, et al
.
A longitudinal HbA1c model elucidates genes linked to disease progression on metformin
.
Clin Pharmacol Ther
2016
;
100
:
537
547
24.
Trischitta
V
,
Copetti
M
.
Moving toward the implementation of precision medicine needs highly discriminatory, validated, inexpensive, and easy-to-use prediction models
.
Diabetes Care
2020
;
43
:
701
703
Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at https://www.diabetesjournals.org/journals/pages/license.