Complexity of Type 2 Diabetes Mellitus Data Sets Emerging from Nutrigenomic Research: A Case for Dimensionality Reduction?

Abstract

Nutrigenomics promises personalized nutrition and an improvement in preventing, delaying, and reducing the symptoms of chronic diseases such as diabetes. Nutritional genomics is the study of how foods affect the expression of genetic information in an individual and how an individual's genetic makeup affects the metabolism and response to nutrients and other bioactive components in food. The path to those promises has significant challenges, from experimental designs that include analysis of genetic heterogeneity to the complexities of food and environmental factors. One of the more significant complications in developing the knowledge base and potential applications is how to analyze high-dimensional datasets of genetic, nutrient, metabolomic (clinical), and other variables influencing health and disease processes. Type 2 diabetes mellitus (T2DM) is used as an illustration of the challenges in studying complex phenotypes with nutrigenomics concepts and approaches.

Introduction

Type 2 diabetes (T2DM) is an example of a complex trait – that is, it results from the contribution of many genes [1], many environmental factors including diet [2], and the interactions among those genes and environmental factors. Various but differing combinations of these factors can produce the same clinical features. The key clinical feature characterizing T2DM, high blood glucose levels [3], may be caused by aberrations in one or more different molecular pathways. T2DM also presents with related physiological responses of hyperinsulinemia, insulin resistance, and other complications. Each of these physiological responses may result from gene – nutrient (and environment) interactions. Hence, the study of nutritional genomics applies to T2DM [4], and indeed almost all chronic diseases, because subsets of nutrient – gene interactions contribute to health or cause disease [5].

The overarching challenges for understanding these disease processes, and indeed, all biological processes including health, are the genetic heterogeneity of humans, the complications of overlapping and varied disease mechanisms, and the complexity of diet and other environmental variables. High-throughput technologies developed in the past 15 years now permit the analysis of hundreds of thousands of genes and their variants, hundreds to thousands of clinical markers such as metabolites, and, theoretically, large numbers of different nutrients and bioactives in foods. The ability to generate high-dimensional datasets however, presents another significant challenge: how to find patterns among genetic, environmental, and clinical symptoms that define and explain complex biological processes. The focus of this review is on the concepts underlying the complexities of the gene – nutrient interactions and emerging approaches for analyzing interacting genetic, molecular, dietary, and clinical data.

The Characteristics of T2DM: Clinical Complexity

The mechanisms, etiology, epidemiology, and genetics of T2DM have been extensively reviewed elsewhere [6-16]. The key diagnostic indicator of T2DM is a fasting blood glucose level above 126 mg/dL (normal range: 70 to 100) on at least two occasions or random blood glucose level of more then 200 mg/dL with symptoms of polyuria and polydipsia (see [17] or [18,19]). Further classification of individuals with impaired fasting glucose levels is done with an oral glucose tolerance (OGT) test. Subjects consume a high-glucose drink (75 g of glucose) administered in the fasted state. Although there are gradations of responses to biological tests, individuals are nevertheless grouped into three classes: normal, impaired, and diabetic.

In addition to the abnormally high circulating glucose and the OGT response, individuals may also be obese, hypertensive, have dyslipidemia, insulin resistance, and/or hyperinsulinemia [6,11,20,21]. These abnormalities may have overlapping molecular and genetic causes to further complicate diagnosis and treatment options. During the course of the disease, many but not all patients develop co-morbidities of the disease including retinopathy, nephropathy, neuropathies, and cardiovascular disease [18]. The potential for these manifestations of the disease cannot be assessed during initial diagnosis, potentially leading to sub-optimal management of the disease, further complications, and increased healthcare costs.

While the varying complications of T2DM are well known, the majority of individuals with diabetic symptoms are treated similarly [18], first with lifestyle changes and then with drugs. Table 1 lists the 6 major drugs used to treat T2DM, 3 of which stimulate insulin production in the pancreas, and the others affect glucose production in the liver, glucose uptake in the intestine, and glucose utilization (through PPAR activity that alters insulin resistance) in the peripheral tissues. The final result is that only ∼20% of patients control symptoms through lifestyle changes [22], about 50% of T2DM patients take oral medications only (Table 1), about 11% take combinations of oral agents with insulin, and 16% take insulin alone [22]. Drug responsiveness alone demonstrates the molecular and lifestyle heterogeneity of T2DM. Optimizing the medications for each patient can be a lengthy trial and error process, involving significant amounts of time and considerable expense.

Table 1. Drug Classes for the Treatment of Type 2 Diabetes.

The number of subtypes of T2DM can be estimated by the different drugs used to treat different clinical indications of type 2 diabetes. In addition to changes in diet and physical activity levels (lifestyle), there are 6 major classes of drugs, 3 targeting the pancreas, one the liver, other pathways in the intestine, and other classes of drugs the adipose and muscle. Some patients require multiple classes of drugs including insulin. Effectiveness is the percent of patients responding to treatment (from http://www.aafp.org/PreBuilt/monograph_diabetestreatment.pdf; Accessed 2 February 2006). See text for details.

Treatment	Target Tissue	Indications	Effectiveness
Lifestyle	All	All	15%
Sulfonylurea	Pancreas	T2DM < 5 yr	∼50%
Meglitinides	Pancreas	T2DM < 5 yr & ⇑ PPG2	?
Exenatide	Pancreas	T2DM	2nd line
Biguanides	Liver	Obese, insulin resistant	∼75%
alpha-glucosidase	Intestine	⇑ PPG	2nd line
Thiazolidinediones	Adipose, muscle	Obese, insulin resistant	2nd line

Genetic Complexity of T2DM

The contributions of many genes [23 and the numerous environmental factors that alter genetic expression confound simple experimental designs for identifying gene, nutrient, or gene – nutrient interactions that cause (rather than affected by) T2DM. The experimental methods successfully used for identifying the genetic basis of monogenic (single gene) diseases {Jimenez-Sanchez, 2001 #1678], such as Huntington disease or cystic fibrosis, can not be directly applied to identifying the genetic basis of complex traits such as T2DM [24].

However, a variation on association studies, quantitative trait locus (QTL) analysis, can identify regions of chromosomes that contribute to a complex trait [25,26]. QTLs are found by statistical analysis of how frequently a region of a chromosome is associated with a measurable phenotype, e.g., insulin levels or glucose response for T2DM. QTLs identify chromosomal regions containing one or more genes contribute different amounts to the phenotype. For example, one QTL may contribute 20% to the trait, while another may contribute only 1%. The contributions will vary depending upon gene – nutrient interactions for the gene responsible for the QTL and whether that gene interacts with other genes in the genome (epistasis – see below). The sum of the contributions from causative alleles in different QTLs produces the specific trait or disease (rev. in [27]). The concept of multiple genes and multiple environmental influences contributing to a complex trait can be illustrated by examining what is currently known about the chromosomal regions containing genes that contribute to T2DM.

T2DM QTLs in Humans

Seven QTLs with LOD (log of the odds, a measure of significance) greater than 3.6 [14] have been found to contribute to T2DM. These loci were found in specific and differing ancestral groups and are on chromosomes 1q25.3, 2q37.3, 3p24.1, 3q28, 10q26.13, 12q24.31, and 18p11.22 [14]. To illustrate the potential complexity of T2DM, assume that there are only 3 alleles at each QTL, with one providing protection against developing the disease, or being neutral, or contributing to the disease. With three possible alleles at 7 loci, the number of potential combinations is 2187. The actual number found in human populations is not as large because allele frequencies differ among ancestral groups. That is, chromosomes of European ancestry are likely to have a different proportion of the alleles at a given locus than do chromosomes of African ancestry. The genes, their quantitative contributions in different ancestral backgrounds, and the combination of alleles at the 7 loci that produce different age of onsets and severities of T2DM are not known. To add to the complexity, seventeen other “suggestive” QTLs found on regions of chromosomes 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 20, and X have also been identified. These QTLs have LOD scores between 2.0 and 3.6 [14], values below the accepted threshold of 3.6.

Some of these genes are likely to be regulated by diet, since certain diets are risk factors for T2DM [28-30]. This means that the susceptibility to disease in each of these individuals will also vary depending upon nutrient intakes that alter expression of genetic information. Some examples of other nutrient and non-nutrient environmental factors affecting the T2DM phenotype are energy intake and balance (reviewed in [31]), sleep patterns ([32,33]), physical activity [34-38], liquid intake [39], and psychological factors, such as stress [40]. Aging and the changes induced by aging, produce health and chronic diseases that are not discrete, dichotomous states, but rather are a continuum. Populations have individuals at different points on this continuum from health to disease. For convenience and treatment, individuals with similar phenotypes are grouped regardless of the differing underlying molecular genetic causes of their condition. The inability to categorize individuals into classes or groups with similar genetic makeups and environmental exposures is a key limitation for identifying genes that cause from those that are affected by the disease.

Genes Associated with Type 2 Diabetes Mellitus

Genetic studies have identified over 50 genes (Table 2) as being involved in T2DM, primarily by association analyses (rev in [4]). While a full description of each of these genes is beyond the scope of this review, they participate in a variety of biochemical, regulatory, and signal transduction pathways, many of which are known to be involved in producing the phenotypes associated with T2DM.

Table 2. Candidate Type 2 Diabetes Mellitus Genes.

Incomplete list of candidate genes identified in literature searches for gene – T2DM genetic association studies.

Genetic	Common Name	Function	Chromosome	References
ABCC8	Sulfonylurea receptor	Potassium channel	11p15.1	[1,14,87],88]
ACP1	Acid Phosphatase1, soluble	Phosphatase	2p25	[89]
ADA	Adenosine deaminase	Enzyme, Purine catabolic pathway	20q13.11 a	[90]
ADRB2	2-Adrenergic Receptor	Receptor linked to catecholamine, obesity	5q32-q34 a	[91-93]
ADRB3	3-Adrenergic Receptor	Receptor, lipolysis regulation	8p12-p11.2	[9]
AGRP	Agouti related protein (homolog of mouse agouti)	Signaling, melanocortin antagonist	16q22	[94]
APM1	Adiponectin (ACDC)	Adipocyte hormone	3q27 b	[9,95,96]
CAPN10	Calpain 10	Cysteine protease	2q37.3 b	[97,98]
ENPP1	Glycoprotein PC-1	Inhibits insulin signaling	6q22 - q23	[99, 93]
FABP2	Liver fatty acid binding protein	Long chain fatty acid transport protein	4q28-q31	[100]
FATP4	Fatty acid transporter, SLC27A4	Long chain fatty acid transport protein (RBC)	Chr. 9	[101]
FOXc2	Transcription Factor	Regulator of adipocyte metabolism	16q24.3	[9,102]
FRDA	Frataxin	Mitochondrial ion metabolism	9q13	[103]
GC	Group specific component, Vitamin D binding protein	Vit D involved in regulating insulin levels	4q12	[93,104]
GCGR	Glucagon receptor	Glucose homeostasis	17q25 a	[14]
GCK	Glucokinase, liver	Enzyme, first step in glycolysis	7p15-p13 a	[105-107]
GFPT2	Glutamine:fructose 6-phosphate amidotransferase 2	Hexosamine biosynthesis	5q34-q35 a	[108]
GHRL	Ghrelin	Hormone, Energy homeostasis and feeding	3p26-p25	[109,110]
GNB3	Guanine nucleotide binding protein 3	Signaling, obesity	12p13	[111]
GYS1	Glycogen synthase	Enzyme, impaired glycogen synthesis	19q13.3 a	[9]
HNF1	Hepatic nuclear factor 1	Transcription factor, cholesterol homeostasis	12q24.2 a	[112]
HNF4A	Hepatic nuclear factor 4	Transcription factor, hepatic glycogen stores	20q12 - q13.1 c	[87]
IAPP	Insulin amyloid protein, Amylin	Hormone, glucose uptake pancreas	12p12.3-p12.1	[113,114]
IGF1	Insulin growth factor 1	Hormone, growth	12q22-q24.1 a	[115]
IL6	Interleukin 6	Cytokine	7p21	[116-119]
INS	Variable number tandem repeat in the insulin gene	Glucose regulation	11p15.5	[120-123]
INSR	Insulin receptor	Receptor	19p13.2	[87]
IPF1	Insulin promotor factor 1	Binds to promoters	12q12.1	[124-126]
IRS1	Insulin receptor substrate 1	Signal transduction	2q36 b	[120]
IRS2	Insulin receptor substrate 2	Signal transduction	13q34	[127]
KCNJ11	Potassium inward rectifier channel Kir6.2	Potassium channel	11p15.1	[128]
LIPC	Hepatic lipase	Lipid, lipoprotein regulation	15q21-q23	[129,130]
LIPE	Hormone sensitive lipase	Mobilization of fatty acids	19q13.1-q13.2	[9]
LPL	Lipoprotein lipase	Enzyme; chylomicrons and triglyceride	8p22	[131]
MAPK8IP1	Mitogen activated protein kinase 8 – interacting protein	Signal transduction	11p12p11.2 a	OMIM 604641
NeuroD1	NeuroD/BETA2	Transcription factor, development	2q32 b	[132]
PAI1	Plasminogen activating inhibitor	Control point in coagulation	7q21.2-q22	[133-135]
PC1	Pachonychia congenita 1	Inhibits insulin signaling	5q15-q21	[99]
PCK1	Phosphoenolpyruvate carboxykinase 1	Enzyme, regulation of gluconeogenesis	20q13.31 a	[136]
PGC1	Peroxisome proliferators-activated receptor coactivator - 1	Transcriptional coactivator	4p15.1	[137-139]
PIK3R1	Phosphoinositide-3-kinase regulatory subunit p85	Glucose clearance	5q13	[140]
PON2	Paraoxonase 2	High fasting glucose	7q21.3	[141]
PPARG	Peroxisome proliferator-activated receptor- gamma 2	Lipid and glucose regulation	3p25 b	[142-144]
PPP1R3A	Protein phosphatase 1, regulatory (inhibitor) subunit 3A	Glycogen metabolism	7q11.23-q21.11	[13,145,146]
RORC	RAR-related orphan receptor C	Nuclear hormone, immune response	1q21	[147]
RRAD	Ras-related associated with diabetes	Insulin sensitivity	16q22	[93,148]
SLC2A2	GLUT2 glucose transporter	Glucose transporter	3q26.1-q26.3 b	[87]
SLC2A4	GLUT4 glucose transporter	Glucose transporter	17p13 a	OMIM 138190
SOS1	Son of sevenless homolog	Guanine nucleotide exchange factor	2p22-p21	[87]
TCF7L2	Transcription factor	Blood glucose homeostasis	T of rs7903146 & T of rs12255372	[149]
TNF	Tumor necrosis factor	Proinflammatory cytokine	6p21.3	[118]
UCP2	Uncoupling protein 2	Mitochondrial transporter	11q13	[150-152]

^aMap position overlaps QTL from diabetes search at http://www.NCBI.nlm.nih.gov/mapview

^bMap position overlaps QTL shown in Figure 1

^cMap position overlaps QTL with near suggestive LOD score [14].

Adapted from [4]

The list in Table 2 is deceiving because many of the genes associated with T2DM in one population fail to be associated in other population, raising the question of whether these genes cause T2DM or are simply affected by disease process. In the absence of obvious flaws in study design, execution, or data analysis [41], lack of association of genes among populations may be due to (i) chronic diseases that are caused by contributions of several genes that may differ among individuals of different ancestral background, (ii) different individuals may have one or more complications such as dyslipidemia, insulin resistance, or obesity which confound statistical analyses, (iii) many cases in case-control studies are molecularly heterogeneous – that is, the same phenotype can result from alterations in different genes and pathways, and (iv) the environmental variables of diet and physical activity were not analyzed.

Gene-gene interaction (epistasis) may also affect gene–disease association studies. Epistatic interactions can occur through protein-protein, protein-gene, RNA-protein interactions, or RNA silencing [42-45]. Epistasis occurs because genes, RNA, proteins, and enzymes are usually part of a pathway or pathways, many of which are interconnected. The concept of biochemical buffering [46,47] or homeostasis, provides a plausible explanation: biological systems try to maintain balance. Inheriting one predisposing allele, therefore, may not alter the final phenotype (such as disease) because an allele of another gene may buffer the effect of the first allele. Gene – gene interactions can be allele-specific and may be additive, negative, or multiplicative. To illustrate epistasis, subjects carrying the adiponectin G allele and the PPARγ2 Ala12 allele seem to be more insulin sensitive than those that encode the adiponectin T allele, Interactions between adiponectin and PPARγ2 genotypes also contribute to fasting insulin concentrations, insulin concentrations in oral glucose tolerance tests, and insulin resistance index [48].

Genetic Ancestry Matters

Epistatic interactions occur because of the gene variants that an individual inherits. The HapMap project has demonstrated that 85 – 90% [49,50] of human variation occurs within a population and only about 10 – 15% is unique to one of the geographical ancestral populations. However, the frequency of alleles differs among ancestral populations such that the probability of inheriting causative or interacting gene variants varies. Genetic drift, small populations existing for extended periods of time (population bottlenecks), or selective pressure all contribute to selection of various alleles. Even small differences in allele frequencies among populations, and therefore among individuals, will lead to differences in biological responses, which include responses to diet.

A specific example that illustrates this point has recently been published. The HapK haplotype (a collection of single nucleotide polymorphisms (SNPs) within a chromosomal region) in the leukotriene 4 hydrolase (LTH4A) gene) is a greater risk factor for myocardial infarction (MI) in African Americans than in European Americans. This is presumably caused by (i) LTH4A interacting differently with one or more gene variants in either African versus European chromosomal regions and/or (ii) different environmental factors altering the influence of LTH4A on myocardial infarction [51]. Thus, the effect of a given allele on a trait or disease must be considered in the context of the other genes [52] in the individual and the environmental factors that may influence its expression and/or function [53,54].

Epigenesis and Chromosome Structure Affect Expression of Genetic Information

Epigenetic interactions also influence the statistical and real association of a SNP with a disease or response to diet. Epigenetics are heritable alterations in gene activity that occur without a change in the sequence of nuclear DNA. X-chromosome inactivation and gene silencing (imprinting) are examples of epigenesis [55]. The two primary mechanisms of epigenesis are DNA methylation and chromatin remodeling, which are separate but related pathways. DNA is methylated at (primarily [56]) CpG dinucleotides which often are clustered in CpG islands near promoters of genes [57]. Proteins which bind to methyl-CpG, (MCPs [58]) interact with other proteins, including chromatin protein modifying enzymes such as histone deacetylases (HDAC) [59]. HDACs in turn alter chromatin structure by changing the activators, co-activators, repressors, and co-repressors that bind to DNA. The net result of DNA methylation and chromatin remodeling is to change the accessibility of DNA to regulatory proteins and transcriptional complexes.

Nutrients alter epigenetic mechanisms through several distinct pathways [60,61]. Dietary deficiencies of choline, methionine, folate, vitamin B₁₂, vitamin B₆, and riboflavin [62,63] alter one carbon metabolism and the concentration of one of its final products, S-adenosylmethionine (S-AM). S-AM is the immediate precursor of DNA (and other) methylation reactions. Chromatin remodeling is a related mechanism that contributes to the control of gene expression. Chromatin structure is regulated in part by the energy balance in a cell: changes in calorie intake alters the NADH:NAD+ (reduced nicotinamide adenine dinucleotide:nicotinamide adenine nucleotide) ratio (reviewed in [64]) and the activity of SIRT1 (sirtuin 1), an NAD⁺-dependent histone deacetylase. The soy – derived peptide, lunasin, may also influence chromatin remodeling by affecting the acetylation – deacetylation balance in chromatin [65]. It is not known whether other peptides can alter acetylation – deacetylation reactions. Changing transcriptional profiles by these methods would alter the phenotype independently of the presence of gene variants (SNPs or collections of SNPs).

Long-term exposure to diets that influence chromatin remodeling (e.g., calories and soybean protein) and DNA methylation (one carbon metabolites) could induce permanent epigenetic changes in the genome. Such changes might explain why certain individuals can more easily control symptoms of chronic diseases by changing lifestyle but many seem to pass an irreversible threshold. Epigenetic changes may also explain “developmental windows”— key times during development, such as in utero, where short-term environmental influences may produce long-lasting changes in gene expression and metabolic potential (reviewed in [21]). Developing experimental approaches for dissecting the environmental influences and the critical genes and pathways will be essential and challenging.

Genotype X Environment Interactions

The concept and definition of gene X environment interactions in nutrition were proposed by Young in 1979 [66]. The precise, statistical definition of gene X environment interaction is “a different effect of an environmental exposure on disease risk in persons with different genotypes,” or, alternatively, “a different effect of a genotype on disease risk in persons with different environmental exposures” [67]. In other words, nutrients affect expression of genetic information and genetic makeup affects how nutrients are metabolized.

Most studies examining candidate gene-disease associations (see Table 2) usually do not account for differences in nutrient intakes among study participants. Gene-diet-phenotype association studies have focused primarily on intermediate risk factors, particularly for cardiovascular disease [68]. Fewer such studies have been conducted for T2DM or the metabolic syndrome. The primary exception has been the association of total and saturated dietary fats (e.g., [69]) with the Pro12Ala variant of peroxisome proliferator activated receptor gamma 2 (PPAR- γ2). Other gene – nutrient interactions, which are not necessarily involved in T2DM, are listed in Table 3. These nutrient – gene – phenotype associations are not always replicated in separate studies. Lack of replication may be attributed to small effect size, population stratification and/or too few study participants (rev. in [70,71]). Well-designed and highly-powered studies are needed to unravel the complexity of gene-nutrient interactions underlying T2DM and its precursor, the metabolic syndrome.

Table 3. Examples of Gene – Nutrient Interactions.

Examples of nutrient – gene interactions.

Gene	Nutrient Variable	Ref
Adiponectin	Mediterranean diet	[153]
	low fat diets	[154]
	macronutrient intake	[155]
Adrenergic receptors	sodium intake	[156]
AGRP	macronutrient intake	[157]
GNB3	sodium intake	[158]
Hepatic lipase	fat	[159]
INS	glucose intake	[160]
PON	alcohol	[161]
PPAR γ	fat	[69]
UCP	energy and body weight	[162]
	chronic overfeeding	[163].

The Analytical Challenge: Finding the Patterns in High-Dimensional Datasets

A key challenge facing high throughput technologies is discovering the functional relationships among elements in genomics, transcriptomics, proteomics, and metabolomics data sets. How are the various levels of quantitative information (DNA, RNA, protein, and metabolite) that define biological processes related to each other and how does diet alter their levels, relationships, interactions, and affect on biological processes [72,73] High-dimensional “-omics” datasets offer additional analytical challenges, for example, the need to control missing values and the high false discovery rate.

The fundamental statistical challenge occurs because each of these “-omics” datasets have high dimensionality – that is, quantitative measurements of the amount (or relative amount) of all (or many) individual components. In a typical genomic, transcriptomic, of proteomic dataset, the number of variables (dimensions) often exceeds the number of samples. For example, many genomic studies measure hundreds of thousands of single nucleotide polymorphism (SNPs). Typical gene expression studies measure tens of thousands of mRNA expression levels; and high-resolution matrix assisted laser desorption ionization time of flight (MALDI-TOF) mass spectrometry proteomic spectra may contain as much as 500,000 data points. In spite of the high-dimensional “-omics” datasets, the number of samples is limited by economic reasons. It is a common challenge in nutritional genomics, that practically each gene has some level of change of expression relative to a nutrient intake or treatment. Each gene – level measurement may have direct, indirect, or inverse correlative relationships to the levels of other gene transcripts, and the sum of these relationships define the transcriptional response to the nutrient or other perturbation, such as the presence or absence of disease or complications of disease. High-dimensional datasets may be represented in a lower dimensional coordinate system by creating a topological map of all gene – level values. The shape of low-dimensional models of any given high-dimensional “-omics” dataset will be influenced and functionally related to the other quantitative measurements of biological components (transcript – protein levels and/or metabolite levels).

A three dimensional topological map has both linear and nonlinear relationships among data elements. Traditional statistical methods are based on regression analysis, that is, they assume linear relationships. Standard statistical methods are often limited in analyzing complex patterns because they omit known interactions among genes and metabolites produced by interacting or competing intracellular pathways and physiological processes [74,75]. Rather, nonlinear analytical methods are increasingly being applied to the analyses of complex biological data (Table 4). One such approach is dimensionality reduction (DR). DR is the concept for analyzing complex data by transforming the original high-dimensional data set (n dimensions) into a lower dimensional data space (m dimensions, where m < n). Instead of graphing 20,000+ individual expression levels (for example), dimensionality reduction methods identify, group, and graph genes or samples with similar properties into 1, 2, 3, or more dimensions [74,75]. Dimensionality reduction algorithms can also discover low-dimensional structures in high-dimensional datasets. The discovered low-dimensional structures may identify relatedness of groups or clusters of genes, proteins, or metabolites that define disease processes, genetic regulatory networks, or disease subtypes. Since the structure of high-dimensional datasets can not be known a priori, the most appropriate method is determined as the one that can best explain most of the variance in the data.

Table 4. Selected Methods of Complex, High Dimensional Data.

See text for details.

Unsupervised Learning (structure detection)
Hierarchical Clustering (HC)	Clusters items based on a similarity metric and organizes them into leaves along branches of a dendogram
Factor analysis	A multivariate technique that aims to summarize a large number of variables with a small number of factors	[164]
Multi-dimensional Scaling (MDS)	MDS is a dimentsionality reduction method that tries to preserve the interpoint distance between points in the low-dimensional space	[165]
Principal Component Analysis (PCA)	Find the n dimensions that explain most of the variance.	[166]
Singular Value Decomposition (SVD)		[167]
Laplacian Eigenmaps	A computationally effective algorithm that has a natural connection to clustering	[168]:
Isomap	Nonlinear embedding; uses geodesic distances	[169].
Locally Linear Embedding (LLE)	Similar to ISOMAP but differs in focus on analyzing distances to nearest neighbors and mapping them to a smooth nonlinear manifold of lower dimensionality	[170]
Hessian Eigenmaps (HLLE)	Refinement of the original Isomap algorithm	[171]
Kohonen Self-organizing Maps (SOM)	A single-layer feed-forward ANN used for visualizing high-dimensional data	[86]
Supervised Learning (structure confirmation)
Linear Discriminant Analysis (LDA)	LDA aims to identify the best linear combination of features that can separate classes of samples	[172]
Support Vector Machines (SVM)	SVMs maximize the geometric margin while minimizing the classification error (maximum margin classifers)
Classification analysis and regression tree (CART)	Analytic procedure for predicting the values of a continuous response variable or categorical response variable from continuous or categorical predictors	[173]
Artificial Neural Networks (ANN)	A form of multiprocessor computer system, with simple processing elements, a high degree of interconnection, simple scalar messages, and adaptive interaction between elements.	[174, 175]

The concept of dimensionality reduction differs from hypothesis – driven, or supervised, approaches. While hypothesis driven experiments are typically considered the gold standard for scientific research, they introduce subjective bias by selecting the possible outcome, do not consider or test for unexpected results, and often filter noise that is considered “unimportant” but which may have un-analyzed information. Factor analysis, classification analysis and regression trees (CART) methods, and artificial neuronal networks (ANN) are examples of supervised techniques (Table 4). Unsupervised analyses, or data-driven approaches, have the disadvantage of being descriptive, but these procedures avoid subjective bias of selecting analyses of a presumed result and they allow for class (or group) discovery. The disadvantage of unsupervised analyses, however, are that they high false discovery rate (type 1 error) and prior knowledge might not explain the results.

The known biological interactions among measured variables (e.g., known gene-gene interactions in gene networks) can be used to further refine results of the unsupervised analyses. Concentrations of one or more metabolites, genetic variations, and alterations in dietary intake may have specific patterns associated with physiological outcomes and responses to treatments. Examples of these interconnections among RNA, protein, and metabolite levels have been published [76]. The goal is to use one or more dimensionality reduction methods to discover interactions and patterns among these datasets that explain complex traits, including health and disease. That is, complexity can be considered an asset rather than an impediment [77], since biological systems are inherently complex, and reductionist approaches fail to account for that complexity. Multidimensional data for T2DM are quantitative measurements of serum or plasma levels of metabolites and proteins, single nucleotide polymorphisms in genes, ancestry-informative markers (AIMs for analyzing genetic background), and dietary variables.

Summary

Modern analytical methods have the ability to generate large datasets of DNA variations, RNA levels, protein amounts, and metabolite measurements. Aberrations in physiology, such as type 2 diabetes and other chronic diseases, introduce additional variables. Many of the standard analytical tools assumed that biological responses were based on linear relationships, an assumption which can not explain phenotypes based on gene – gene, gene – nutrient, or gene – nutrient interactions. While the number of reports describing the analyses of biological data with data reduction methods is increasing [74,75,78-86], it appears that no “rule-of-thumb” can be used to chose the most appropriate analytical method. Rather, the algorithm that best describes the data and relationships, usually as determined by the smallest variance, is the method of choice for that dataset. Discovery of novel patterns among nutrient intakes, genetic ancestry, candidate genes, and physiological measurements may allow us a greater understanding of disease pathologies, and a path to developing ways to maintain health and prevent or delay diseases.