Objectives: Many common diseases and metabolic disturbances are characterised by the development of micro- and macrovascular complications. Early detection of high risk individuals would aid in preventing clinical consequences.

Methods: A high-throughput nuclear magnetic resonance (NMR) metabonomics approach is introduced to characterise individual systemic metabolic phenotypes. The methodology is based on two molecular windows for native serum – LIPO and LMWM. The LIPO window is giving information, for example, on the lipoprotein subclass distribution and the LMWM window on various low-molecular-weight metabolites such as glucose, alanine, and creatinine. The combination of these two molecular windows is likely to contain most of the information available by NMR metabonomics of native serum. In addition, a recently established NMR protocol for serum lipid extracts is presented to obtain detailed molecular information on serum lipid constituents such as free and esterified cholesterol, phosphatidylcholine, sphingomyelin, and ω-3 fatty acids. The experimentation is fully automated with a current analysis capacity of up to 180 LIPO + LMWM or LIPID windows in 24h. In total, approximately 140 metabolic measures can be quantified from each serum sample.

Results: Besides the introduction of the new instrumental set-up for high-throughput serum NMR metabonomics, an overview of the most recent results for the metabolic phenotyping of the serum samples in the Northern Finland Birth Cohort 1966 (n~6,000) are presented. The metabolic phenotypes are characterised and discussed with respect to various epidemiological variables, risk assessment, diagnostics, and also genome-wide data. We demonstrate the inherent suitability of serum NMR metabonomics to identify subtle changes in lipoprotein subclass-related metabolism as well as in various other metabolites of high interest.

Conclusions: Our work reveals the diffuse genotypic and phenotypic nature of vascular problematics and the limitations of single ‘diagnostic’ biomarkers. However, it also promises cost-effective solutions via high-throughput analytics and advanced computational methods.

Objective: To determine associations of systemic metabolites with subclinical atherosclerosis and assess the existence of distinct metabolic phenotypes associated with early cardiovascular risk.

Methods: 4,407 serum samples were measured by nuclear magnetic resonance (NMR) spectroscopy from the Cardiovascular Risk in Young Finns Study and carotid intima-media thickness (IMT) assessed by ultrasound. Numerous lipoprotein lipids subclasses as well as low-molecular-weight metabolites were quantified from the NMR data.

Results: In these healthy young adults data-driven analysis using self-organizing maps revealed three distinct metabolic phenotypes associated with elevated IMT. The phenotypes were characterized by varying combinations of metabolic disturbances like elevated VLDL and LDL subclasses, but also several low-molecular-weight metabolites.

Conclusion: The study revealed different metabolic phenotypes inherently associated with subclinical atherosclerosis. The findings substantiate developments toward the use of multi-metabolic risk phenotypes in cardiovascular risk assessment.

We have recently set up a highly automated, high-throughput platform for serum NMR metabonomics [1] that has now (end of May 2010) been operational for ~17 months with around ~60,000 samples from various epidemiological studies analyzed. Thus, a new era is at our hands [2] to study large-scale epidemiological studies of common public health problems such as vascular diseases and the related metabolic conditions such as diabetes and metabolic syndrome.

The requirement of true high-throughput demands not only a fine tuned and automated hardware, but calls also for fast and automated methods to process and quantify the spectra. The standard protocol involves two molecular windows for each sample: a standard 1H spectrum (LIPO) and a CPMG spectrum (LMWM) for quantification of lipoprotein subclasses and low-molecular-weight metabolites, respectively. For the latter, an automatic line shape fitting method was originally used to establish metabolite quantifications. However, reliable automation of line shape fitting is a tricky issue, particularly for rather complex biological spectra, and due to the iterative character of the analysis also time concerns arise. We illustrate at the poster how substantial increase in speed was achieved by converting the line shape fitting procedures into linear regression models [3] without compromising the quantification accuracy, but in fact increasing the reliability of the automated analyses.

Briefly, 300 spectra were carefully quantified by manual line shape fitting in TLS (a part of PERCH software) [4]. These spectra were then used as training data to build metabolite specific linear regression models, which perform the quantification by using the spectral data points as input only. A pre-processing procedure, including, e.g., baseline zeroing and peak alignments, was also created to increase the model performance.

The prediction accuracy was tested by Monte Carlo type cross-validation procedures. The regression models proved to be able to quantify at least 22 metabolites from the LMWM spectra, all of them with greater than R2=0.9 cross-validation accuracy. The increase in quantification performance was from one minute per spectrum with the previous iterative line shape fitting method to approximately one minute per 1000 spectra with the new regression models (on a standard desktop computer). This development is crucial for the epidemiological metabonomics applications.

• Soininen P; Kangas AJ; Würtz P; Tukiainen T; Tynkkynen T; Laatikainen R; Järvelin MR; Kähönen M; Lehtimäki T; Viikari J; Raitakari OT; Savolainen MJ; Ala-Korpela M. Analyst 2009, 134, 1781-1785.
• Inouye M; Kettunen J; Soininen P; Silander K; Ripatti S; Kumpula L S; Hämäläinen E; Jousilahti P; Kangas A J; Männistö S; Savolainen MJ; Palotie A; Salomaa V; Perola M; Ala-Korpela M; Peltonen L. Submitted. 2010
• Ala-Korpela M; Changani KK; Hiltunen Y; Bell JD; Fuller BJ; Bryant DJ; Taylor-Robinson SD; Davidson BR. Magn. Reson. Med. 1997, 38, 840-844.
• Soininen P; Haarala J; Vepsäläinen J; Niemitz M; Laatikainen R. Anal. Chim. Acta. 2005, 542, 178-185.

Objective: To determine associations of systemic metabolites with subclinical atherosclerosis and assess the existence of distinct metabolic phenotypes associated with early cardiovascular risk.

Methods: 4,407 serum samples were measured by nuclear magnetic resonance (NMR) spectroscopy from the Cardiovascular Risk in Young Finns Study and carotid intima-media thickness (IMT) assessed by ultrasound. Numerous lipoprotein lipids subclasses as well as low-molecular-weight metabolites were quantified from the NMR data.

Results: In these healthy young adults data-driven analysis using self-organizing maps revealed three distinct metabolic phenotypes associated with elevated IMT. The phenotypes were characterized by varying combinations of metabolic disturbances like elevated VLDL and LDL subclasses, but also several low-molecular-weight metabolites.

Conclusion: The study revealed different metabolic phenotypes inherently associated with subclinical atherosclerosis. The findings substantiate developments toward the use of multi-metabolic risk phenotypes in cardiovascular risk assessment.

Objective – Diabetic nephropathy and other complications are influenced by numerous genetic and environmental factors. Accurate knowledge of the complex inter-dependencies between these factors is critical for pinpointing the best targets for research and treatment. Therefore, the aim of this study was to describe the association patterns between clinical and biochemical features of type 1 diabetes (T1DM) complications.

Design and patients – Baseline clinical data on 4,197 patients with T1DM were collected at the 92 FinnDiane Study Centers in Finland. The study visit included a clinical examination, registering of complications, and questionnaires on education, smoking, alcohol consumption, working status (disabled vs. employed/unemployed), asthma, rheumatoid arthritis and thyroid disease. Blood and 24h urine samples were also collected. The mean age and T1DM duration were 38 and 22 years, respectively and 52% were male. A total of 875 patients (21%) had nephropathy (AER > 300 mg/24h or end-stage renal disease) at the time of study. Vitality status was obtained prospectively from the national population registry after an average of 6.5 years of follow up (25,714 patient years).

Methods – The dataset comprised 28 clinical and 25 biochemical variables that were regarded as the nodes of a network; the associations between the variables were quantified by pair-wise correlation coefficients (continuous data only) or by linear regression modeling (continuous and binary data combined). The resulting structures were visualized by a force-directed node-positioning algorithm.

Results – Heavily connected cliques of dependent variables such as HDL-related lipid markers, indicators of kidney failure and indices of body mass were observed. The links between the cliques showed biologically relevant interactions: an inverse relationship between HDL cholesterol and the triglyceride clique (P < 10-16), a connection between triglycerides and body mass via C-reactive protein (P < 10-16) and estimated IDL cholesterol as the connector between lipoprotein metabolism and macroalbuminuria (P < 10-16). The regression-correlation network showed strong associations between blood pressure, working ability and complications.

Conclusions – There was modularity in the dataset and several biologically relevant connections were visible, which supported the use of networks in clinical data analysis. In particular, accelerated vascular aging was seen in blood pressure correlations and IDL emerged as a putative lipoprotein covariate in the pathogenesis of diabetic nephropathy.

Objective – Diabetic nephropathy and other complications are influenced by numerous genetic and environmental factors. Accurate knowledge of the complex inter-dependencies between these factors is critical for pinpointing the best targets for research and treatment. Therefore, the aim of this study was to describe the association patterns between clinical and biochemical features of type 1 diabetes (T1DM) complications.

Design and patients – Baseline clinical data on 4,197 patients with T1DM were collected at the 92 FinnDiane Study Centers in Finland. The study visit included a clinical examination, registering of complications, and questionnaires on education, smoking, alcohol consumption, working status (disabled vs. employed/unemployed), asthma, rheumatoid arthritis and thyroid disease. Blood and 24h urine samples were also collected. The mean age and T1DM duration were 38 and 22 years, respectively and 52% were male. A total of 875 patients (21%) had nephropathy (AER > 300 mg/24h or end-stage renal disease) at the time of study. Vitality status was obtained prospectively from the national population registry after an average of 6.5 years of follow up (25,714 patient years).

Methods – The dataset comprised 28 clinical and 25 biochemical variables that were regarded as the nodes of a network; the associations between the variables were quantified by pair-wise correlation coefficients (continuous data only) or by linear regression modeling (continuous and binary data combined). The resulting structures were visualized by a force-directed node-positioning algorithm.

Results – Heavily connected cliques of dependent variables such as HDL-related lipid markers, indicators of kidney failure and indices of body mass were observed. The links between the cliques showed biologically relevant interactions: an inverse relationship between HDL cholesterol and the triglyceride clique (P < 10-16), a connection between triglycerides and body mass via C-reactive protein (P < 10-16) and estimated IDL cholesterol as the connector between lipoprotein metabolism and macroalbuminuria (P < 10-16). The regression-correlation network showed strong associations between blood pressure, working ability and complications.

Conclusions – There was modularity in the dataset and several biologically relevant connections were visible, which supported the use of networks in clinical data analysis. In particular, accelerated vascular aging was seen in blood pressure correlations and IDL emerged as a putative lipoprotein covariate in the pathogenesis of diabetic nephropathy.

Background and aims: Risk assessment of diabetic complications is challenging due to the gradual disease processes and the lack of accurate metabolic phenotypes. Here the focus is on the latter problem: to reveal the multi-variate biochemical features behind the clinical outcomes and premature deaths in a set of Finnish patients with type 1 diabetes.

Materials and methods: Baseline biochemical and clinical data were collected for 2173 men and 2024 women with type 1 diabetes (age at onset below 35 years, insulin treatment within a year of onset) from the Finnish Diabetic Nephropathy Study. The vitality status was obtained after an average of 6.5 years of follow-up (25,714 patient years, 295 deaths). Clinical records and measurements of 20 biochemical variables (lipoprotein subfractions, apolipoproteins, serum and urine creatinine, inflammatory markers and 24h-urine albumin) were collected by standardized questionnaires and laboratory methods. The data were analysed by a self-organising map (SOM), and visualized separately for males and females with computational significance estimates and confidence intervals. The metabolic phenotypes revealed by the SOMs were then compared against the observed all-cause mortality.

Results: Two sides of the diabetic complications cluster were revealed: a metabolic syndrome profile of insulin resistance, low HDL2C, and abdominal obesity, and a high-adiponectin, high-LDLC, high HDL2C profile associated with microvascular complications. A 10.1-fold [CI95%: 7.3-13.1%] population adjusted risk of death in males and a 10.7-fold [CI95%: 7.9-13.7] risk in females was observed at the intersection of the two metabolic states.

Conclusion: The SOM approach offers a visual way of incorporating multi-variate biochemical datasets into clinical decision making, as demonstrated in this study by the dissection of diabetic complications into a metabolic continuum between two characteristic phenotypes. We expect this type of computational medicine to be crucial not only for cultivating the scientific understanding of multifactorial metabolic disorders, but also fro enabling tailored treatments of complex diseases.

Altered lipid metabolism and subsequent vascular complications are the underlying causes for clinical symptoms in common diseases like dementias, atherothrombosis and diabetes [1]. The non-selective nature and metabolic specificity are advantageous features of 1H NMR spectroscopy to study lipid composition of body fluids. NMR spectroscopy can yield data on all components in a mixture, without the need to fractionate or derivatise the sample. Sophisticated analysis, however, is the key for reliability and extensive molecular coverage.

Our poster will focus on the optimisation of the 1H NMR lipidomics approach for metabonomics use. The figure below illustrates a part from a 1H NMR spectrum of serum lipid extracts together with line fitting using the total-line-shape fitting tool of the PERCH NMR Software [2]. We have applied the presented methodology for over 500 serum samples from two different clinical studies. At the poster we will present the lipid descriptors available and discuss the suitability of this approach for clinical metabonomics.

• Mäkinen, VP; Soininen, P; Forsblom, C; Parkkonen, M; Ingman, P; Kaski, K; Groop, PH; Ala-Korpela, M. Mol. Syst. Biol. 2008, 4, 167.
• Soininen, P; Haarala, J; Vepsäläinen, J; Niemitz, M; Laatikainen, R. Anal. Chim. Acta. 2005, 542, 178.

Towards personalised medicine: Understanding the factors that influence human health and cause diseases has always been a driving force of research. With the exciting progress in highthroughput analytical techniques and the profound integration of experimental and computational approaches, medicine has newly got hold of new technological and conceptual tools for holistic investigations of living organisms at the system level. The still young discipline of systems biology has mostly been applied to study well-characterised model organisms. However, the first human studies also report on remarkable opportunities that combined molecular and computational technologies can have for the progress of personalised and predictive medicine:

"This application of 1H NMR metabonomics of serum demonstrates the diffuse nature of complex vascular diseases and the limitations of single diagnostic biomarkers, but it also promises cost-effective solutions through high-throughput analytics and advanced computational methods, as illustrated here for patients with type 1 diabetes in a real clinical situation." [1]

Metabonomics – a new field of 'omics': Genomics, transcriptomics and proteomics, represent the 'genomistic' main discipline in life sciences. The phenotype of a biological system, however, is principally reflected by its metabolite composition and their interactions. Therefore, a key 'omics' in understanding of biomolecular function is metabonomics: the measurements of multi-metabolic responses to (patho)physiological stimuli or genetic modifications. Mass spectrometry (MS) and 1H NMR spectroscopy have become the two key technologies in this area. An appealing feature of NMR is its specific yet non-selective nature:

"1H NMR spectroscopy techniques are rather fast and straightforward to apply to all biofluids in vitro and also to various tissues ex vivo and in vivo – approaches combining data on various biofluids and/or tissues of the same individuals (integrated metabonomics) are thus increasingly used to study systems level biochemistry." [2]

Towards new technological platforms: The metabonomic applications are leading to massive, complex data sets. We are systematically developing and applying various bioinformatics methods in this area, for example, self-organising maps will be introduced as an appealing tool for visual metabolic profiling [1]. Our collaborative medical applications focus on metabolic phenotyping and risk assessment of vascular complications. Please come and discuss the future of health care with us at the poster – does NMR really have a role here?

• Mäkinen, VP; Soininen, P; Forsblom, C; Parkkonen, M; Ingman, P; Kaski, K; Groop, PH; Ala-Korpela, M. Mol. Syst. Biol. 2008, 4, 167.
• [2] Ala-Korpela, M. Expert Rev. Mol. Diagn. 2007, 7, 761.

Objective – Subclinical risk assessment of diabetic kidney disease is challenging due to the gradual disease processes and the lack of accurate metabolic phenotypes. Here the focus is on the latter problem: to reveal the multivariate biochemical features behind the clinical outcomes and premature deaths in a populationbased set of Finnish patients with type 1 diabetes.

Design and patients – Baseline biochemical and clinical data were collected for 2,176 males and 2,121 females with type 1 diabetes from the Finnish Diabetic Nephropathy Study, after excluding those patients with age of onset ≥35, or more than 50% of data missing. The vitality status was obtained after an average of 6.5 years of followup (24,589 patient years, 271 deaths).

Methods – Clinical records and measurements of 19 biochemical variables (lipoprotein subfractions, apolipoproteins, serum and urine creatinine, inflammatory markers and 24hurine albumin) were collected by standardized questionnaires and laboratory methods. The data were analysed by self-organizing maps (SOMs) of males and females separately, with computational significance estimates and confidence intervals. The metabolic phenotypes revealed by the SOMs were compared against the observed allcause mortality.

Results and conclusions – Two sides of the diabetic complications cluster were revealed: a metabolic syndrome profile of insulin resistance, low HDL2C, and abdominal obesity, and a highadiponectin, highLDLC, highHDL2C profile associated with microvascular complications and longer diabetes duration. An 11.4fold [CI95%: 9.713.4] population adjusted risk of death in males, and a 9.1fold [CI95%: 6.911.3] risk in females was observed at the intersection of the two metabolic states, where also the prevalence of advanced kidney disease was the highest (67% for males, 56% for females). At subclinical level, this highrisk metabolic phenotype was associated with a 3.7fold [CI95%: 1.36.0] risk in males with no kidney disease (normoalbuminuria), and a 4.3fold [CI95%: 0.57.8] risk in males with unclassified albuminuria status. The SOM approach offers a visual way of incorporating multivariate biochemical datasets into clinical decision making, as demonstrated in this study by the dissection of albuminuria into a metabolic continuum between two characteristic phenotypes. We expect this type of computational medicine to be crucial not only for cultivating the scientific understanding of complex diseases, but also for enabling tailored treatments before clinical signs emerge.

Background and aim: Low plasma adiponectin is related to atherosclerosis. Interestingly, alcohol usage increases adiponectin concentrations. Here we investigated the link between adiponectin, conventional lipoprotein risk factors, and alcohol consumption.

Methods: Plasma adiponectin and lipids in VLDL, IDL, LDL and HDL of 70 subjects were measured and analysed using self-organizing maps, an unsupervised neural network methodology forming an overall picture of the biochemical situation.

Results: The analyses revealed four different lipoprotein phenotypes (PH). PH-I in which HDL particles were rich in cholesterol and phospholipids but poor in triglycerides reflecting a rise in the concentration of large HDL particles. Mean plasma adiponectin (14.6 µg/ml), apoA-I (1.84 g/l), and alcohol consumption (182 g/day) were high. VLDL, IDL and LDL concentrations were low. PH-II was characterized by an increased concentration of large LDL particles, medium HDL but low VLDL, and IDL. Mean adiponectin was 12.4 µg/ml and alcohol consumption 51 g/day. In PH-III, both LDL and HDL particles were poor in cholesterol but rich in triglycerides reflecting an increase in small dense LDL and HDL. VLDL and IDL triglycerides were also high. Mean adiponectin was 10.2 µg/ml and alcohol consumption 101 g/day. PH-IV had the lowest adiponectin (9.3 µg/ml) and low HDL, but high concentrations of VLDL, IDL, and LDL. Mean alcohol consumption was 42 g/day.

Conclusions: The results revealed a clearly anti-atherogenic lipoprotein phenotype (PH-I) that was associated with the highest adiponectin and alcohol consumption. In contrast, the atherogenic lipoprotein phenotypes (PH-III/IV) were associated with significantly lower adiponectin and alcohol consumption.

Background and Aims: Macrovascular disease in connection to diabetic nephropathy accounts for most of the premature deaths in type 1 diabetic patients. Metabolic syndrome is common in the subset that develop these serious complications, which suggests that the risk factors are similar to those in the general population. The definitions of risk criteria, however, may not be well suited for type 1 diabetes nor be accurate enough to detect subtle metabolic changes. For this reason, we have measured proton nuclear magnetic resonance (NMR) spectra of serum for a subset of type 1 diabetic patients from the Finnish Diabetic Nephropathy (FinnDiane) study. Our goal is to determine the multi-variate metabolic characteristics of type 1 diabetes that will serve as the baseline for the prospective phase of FinnDiane, and to evaluate the role of proton NMR as an analytical tool in this effort.

Materials and Methods: Proton NMR spectra of serum for 613 type 1 diabetic patients was measured on a Bruker AVANCE 500Mhz spectrometer. Age- and sex-matched patients were selected from the FinnDiane cohort based on nephropathy status. Proton NMR data was obtained from two molecular windows: a standard spectrum with wide lipoprotein lipid and albumin signals and a T2-filtered spectrum that better reveals the smaller metabolites with lower molecular weights. A double-tube system for reference substance was used to enable absolute quantification of metabolites. The complex spectral data was analysed by constructing a self-organising map, an unsupervised neural network analysis method.

Results: The self-organising map of the spectra identified multi-variate spectral profiles that are linked to clinical criteria such as nephropathy, retinopathy, metabolic syndrome and macrovascular complications. It also revealed specific connections between the diagnoses (e.g. nephropathy) and the corresponding serum biochemistry. The observations were validated by non-spectroscopic data from the FinnDiane database and interpreted with the help of statistical map colourings (figure).

Conclusion: A single proton NMR analysis of serum for type 1 diabetic patients gave enough information for a multi-metabolite characterisation of diabetic complications that is likely to be of great value in the sub-clinical screening of metabolically vulnerable patients.

Atherosclerosis and diabetic kidney disease are examples of gradually developing complex conditions that lead and contribute to increased occurrence of macrovascular events such as myocardial infarction and stroke. Patients with a long history of diabetes are especially at risk, although not all develop these complications. To distinguish the high-risk individuals, we applied 1H NMR spectroscopy as a high-throughput metabolic characterization tool in a large clinical study and investigated the multi-metabolite nature of various clinical classifications. 1H NMR spectra of serum were measured for 613 patients with type 1 diabetes from the FinnDiane study. The spectral shapes were then analysed by constructing a self-organizing map, and coloring the map based on the clinical criteria and several metabolites that were directly quantified from the spectra. As expected, a decrease in kidney function (elevated serum creatinine) was related to increased mortality (figure). Increased triglycerides and decreased HDL2 were related to macrovascular diseases, although there was also overlap with kidney disease. The combination of 1H NMR and self-organizing map provided a simple and easily interpretable, yet powerful and efficient method to reveal subtle metabolic disturbances. It is likely that this type of approach is crucial to monitor the disease progression and when deciding treatements at an individual basis.

Early identification of physiological conditions that relate to increased risk of atherosclerosis would be advantageous to facilitate individual primary prevention. Metabolic syndrome (MetS), accompanied with characteristic changes in lipoprotein subclasses, is one such condition with a steeply increasing prevalence in the general population. 1H NMR spectroscopy is an alternative approach to quantify lipoprotein subclasses and can also provide a holistic overview of other serum metabolites. Here we studied if the metabonomic tactic, i.e., 1H NMR spectroscopy together with a newly proposed data analysis methodology, namely affinity propagation [1] would enable characterisation of lipoprotein subclass-related metabolism in a clinically relevant context. Using experimentally derived model signals for VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b, HDL2a, HDL3a, HDL3b and HDL3c two biochemically characteristic categories of spectra were simulated, representing lipoprotein subclass profiles for a normolipidaemic (healthy) and a metabolic syndrome (high risk) status. Sets of spectra representing a metabolic pathway between these two main categories were also included using three representative categories between the healthy and MetS lipoprotein profiles. The analyses were performed with thousands of spectra. Affinity propagation finds exemplars exchanging messages between data points, taking as input the measures of similarity between pairs of samples. The message passing algorithm can find clusters without the need of a pre-specified number of clusters. The latter is a particularly important aspect in biomedical applications. The results show that affinity propagation is capable of correctly classifying from 68% to 81% of the spectra, and, importantly, confusion only occurs between the biologically adjacent categories. Additionally, it should be emphasised that the simulated spectra included both realistic statistical population variation as well as individual variations in the lipoprotein subclass signals. Therefore, these results demonstrate the inherent suitability of 1H NMR metabonomics to identify subtle changes in lipoprotein subclass-related metabolism. This work appears as the first application of affinity propagation to analyse NMR metabonomics data. Different executions, using only different subsets of the 1H NMR spectra, show that it is possible to optimise the measure of similarity between data points in order to improve the clustering results. In general, our results suggest that affinity propagation might have many application areas in metabonomics.

• B. J. Frey, et al. Science. 2007, 315, 972–976.

Early identification of physiological conditions that relate to increased risk of atherosclerosis would be advantageous to facilitate individual primary prevention. Metabolic syndrome (MetS), accompanied with characteristic changes in lipoprotein subclasses, is one such condition with a steeply increasing prevalence in the general population. 1H NMR spectroscopy is an alternative approach to quantify lipoprotein subclasses and can also provide a holistic overview of other serum metabolites. Here we studied if the metabonomic tactic, i.e., 1H NMR spectroscopy together with a newly proposed data analysis methodology, namely affinity propagation [1] would enable characterisation of lipoprotein subclass-related metabolism in a clinically relevant context. Using experimentally derived model signals for VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b, HDL2a, HDL3a, HDL3b and HDL3c two biochemically characteristic categories of spectra were simulated, representing lipoprotein subclass profiles for a normolipidaemic (healthy) and a metabolic syndrome (high risk) status. Sets of spectra representing a metabolic pathway between these two main categories were also included using three representative categories between the healthy and MetS lipoprotein profiles. The analyses were performed with 1600 spectra. Affinity propagation finds exemplars exchanging messages between data points, taking as input the measures of similarity between pairs of samples. The message passing algorithm finds clusters with much lower classification error than many other methods, and does not need a pre-specified number of clusters. The latter is a particularly important aspect in biomedical applications. The results show that affinity propagation is capable of correctly classifying from 68% to 81% of the spectra, and, importantly, confusion only occurs between the adjacent categories. Additionally, it should be emphasised that the simulated spectra included both realistic statistical population variation as well as individual variations in the lipoprotein subclass signals. Therefore, these results demonstrate the inherent suitability of 1H NMR metabonomics to identify subtle changes in lipoprotein subclass-related metabolism. This work appears as the first application of affinity propagation to analyse NMR metabonomics data. Different executions, using only different subsets of the 1H NMR spectra, show that it is possible to optimise the measure of similarity between data points in order to improve the clustering results. In general, our results suggest that affinity propagation might have many application areas in metabonomics.

• B. J. Frey, et al. Science. 2007, 315, 972–976.

The presence of a metabolic syndrome (MetS) relates to particular changes in the lipoprotein subclass profile and to an increased risk for atherosclerosis. Some subclasses are known to be more important than others with respect to the risk for atherosclerosis, but the current clinical risk assessment methodology cannot take this properly into account. The use of 1H NMR spectroscopy seems to have significant potential in clinical use since the technique enables a fast measurement of the lipoprotein profile directly from a serum sample. Many computational methods have been developed for this purpose [1]. However, the intrinsic capability of this approach to identify the different lipoprotein subclass profiles has never been clearly evaluated in a probabilistic manner. To accomplish this, we have applied Fuzzy Artmap algorithms to classify individuals at the borderline of health and disease based on their 1H NMR spectra of serum. Fuzzy Artmaps are neural network classification algorithms that are trained in a supervised way. Compared to many other neural paradigms, these networks can learn quickly from a reduced dataset, they are easy to program (they require less computational power than many other paradigms), and they also have the ability to learn rare events easily. Moreover, using the so-called "voting strategy" they can give a confidence value to the category assignment given to new samples not seen during the training [2]. We have used Fuzzy Artmap algorithms to classify samples from an extensive simulated dataset of serum 1H NMR spectra. This dataset was built based on experimental lipoprotein subclass information and it includes typical lipoprotein profiles for normolipidaemic (healthy) and MetS (high risk) individuals. Sets of spectra representing a metabolic pathway between the two categories were also included using three representative categories between the healthy and MetS lipoprotein profiles. The analyses were performed with 2500 spectra. In the Fuzzy Artmap analyses 50% of the samples were used for training and 50% for independent evaluation. Fifty voting iterations were performed to obtain the confidence values. The results show that Fuzzy Artmaps are capable of correctly classifying 81% of the spectra. Moreover, confusion only occurs between adjacent categories (i.e., 4% of healthy individuals are erroneously classified as belonging to the first metabolic pathway group). These results clearly indicate that the use of 1H NMR spectra for atherosclerosis risk assessment is feasible. Thus, these results warrant further work in this direction to pursue the measurement of statistically representative data on real populations.

• M. Ala-Korpela et al. Atherosclerosis. 2007, 190, 352–358.
• Carpenter et al. IEEE Transactions on neural networks. 1992, 3, 698–713.

Identification of (patho)physiological conditions that relate to increased risk of atherosclerosis would be advantageous for individual primary prevention. Metabolic syndrome (MetS), accompanied with characteristic changes in lipoprotein subclasses, is one such condition with an increasing prevalence. 1H NMR spectroscopy is an alternative approach to lipoprotein subclass measurements and can also provide an overview of other serum metabolites1,2. Here we studied if the metabonomic strategy, i.e., 1H NMR together with holistic data analyses, would enable the characterisation of lipoprotein subclass-related metabolism in clinically relevant contexts. First, using experimentally derived model signals for 11 lipoprotein subclasses (VLDL1-2, IDL, LDL1-3, HDL2b&2a, and HDL3a-3c) two biochemically typical categories of spectra were simulated, representing subclass profiles for a normolipidaemic and a MetS status. A set of spectra representing a metabolic pathway between the two categories was also generated. Various analyses of thousands of simulated 1H NMR spectra clearly identified the lipoprotein subclass profiles and their changes. These findings are supported by comparable analyses for a representative set of experimental 1H NMR spectra of 69 serum samples with a wide range of lipoprotein lipid concentrations3. In addition, we discuss the role of MetS in a cohort of over 700 type 1 diabetic patients and illustrate how the multi-metabolite data via 1H NMR reveal the continuum of diabetic complications, i.e., MetS, diabetic kidney disease and atherosclerosis4. All the results demonstrate the inherent suitability of 1H NMR metabonomics to identify subtle changes in lipoprotein subclass-related metabolism. Accordingly, 1H NMR can be seen as a new methodology for screening individuals at high risk for atherosclerosis and also being potentially useful in prospective assessment of individual 'health paths' in the borderline of, for example, normal lipoprotein metabolism (health) and the MetS (disease).

• M. Ala-Korpela. Progr Nucl Magn Reson Spectr. 1995;27:475.
• M. Ala-Korpela, N. Lankinen, A. Salminen, T. Suna, P. Soininen, R. Laatikainen, P. Ingman, M. Jauhiainen, M.-R. Taskinen, K. Heberger, K. Kaski. Atherosclerosis. 2007;190:352.
• T. Suna, A. Salminen, P. Soininen, R. Laatikainen, P. Ingman, S. Mäkelä, M. J. Savolainen, M. L. Hannuksela, M. Jauhiainen, M.-R. Taskinen, K. Kaski, M. Ala-Korpela. NMR Biomed. 2007 Nov;20(7):658-72.
• V.-P. Mäkinen, P. Soininen, C. Forsblom, M. Parkkonen, P. Ingman, K. Kaski, P.-H. Groop, M. Ala-Korpela. Magn Reson Mater Phy. 2006;19:281.

Identification of (patho)physiological conditions that relate to increased risk of atherosclerosis would be advantageous for individual primary prevention. Metabolic syndrome (MetS), accompanied with characteristic changes in lipoprotein subclasses, is one such condition with an increasing prevalence. 1H NMR spectroscopy is an alternative approach to lipoprotein subclass measurements and can also provide an overview of other serum metabolites. We have studied if the metabonomic strategy, i.e., 1H NMR together with holistic data analyses, would enable the characterisation of lipoprotein subclass-related metabolism in clinically relevant contexts. First, using experimentally derived model signals for 11 lipoprotein subclasses two biochemically typical categories of spectra were simulated, representing subclass profiles for a normolipidaemic and a MetS status. A set of spectra representing a metabolic pathway between the two categories was also generated. Various analyses of thousands of simulated 1H NMR spectra, e.g., using self-organising maps (SOMs), fuzzy artmaps and affinity propagation, clearly identified the lipoprotein subclass profiles and their changes. These findings are supported by comparable analyses for a representative set of experimental 1H NMR spectra of 69 serum samples with a wide range of lipoprotein lipid concentrations [1]. In addition, we discuss the role of MetS in type 1 diabetic patients and illustrate how the multi-metabolite data via 1H NMR reveal the continuum of diabetic complications, i.e., MetS (see the figure below illustrating statistical SOM analysis).

All the results demonstrate the inherent suitability of 1H NMR to identify subtle changes in lipoprotein subclass-related metabolism. Accordingly, 1H NMR can be seen as a new methodology for screening individuals at high risk for atherosclerosis and also being potentially useful in prospective assessment of individual 'health paths' in the borderline of health (normal lipoprotein metabolism) and disease (MetS).

• T. Suna, A. Salminen, P. Soininen, R. Laatikainen, P. Ingman, S. Mäkelä, M. J. Savolainen, M. L. Hannuksela, M. Jauhiainen, M.-R. Taskinen, K. Kaski, M. Ala-Korpela. NMR Biomed. 2007 Nov;20(7):658-72.

Cholesterol in atherosclerotic lesions originates mostly from lipid droplets formed from modified low-density lipoprotein (LDL) particles. Understanding of these modifications calls for dynamic physicochemical studies on LDL modifications. In general, there is a lack of techniques that would facilitate a non-destructive and dynamic follow-up of molecular processes in a native LDL sample under an enzymatic attack in a physiological environment.

However, 1H NMR seems a feasible choice. Particularly, since our recent studies on the effects of phospholipase A2 (PLA2) on LDL structure at 800 MHz distinguish all major phospholipids, including lysophosphatidylcholine (lyso-PC), at the LDL particles as illustrated in the figure on the right.

The signal assignment for the lyso-PC is novel and we illustrate the applicability of the methodology in the case of lipid peroxidation that is generally considered as one of the key pro-atherogenic modifications of LDL (see the figure below). It was found, somewhat surprisingly, that the LDL-associated PLA2 is activated in the very beginning of the formation of PC-hydroperoxides. Together with the methodology, the (patho)physiological rationale of the resulting lyso-PC generation will also be discussed in the presentation [1].

• P. Soininen, K. Öörni, H. Maaheimo, R. Laatikainen, P. T. Kovanen, K. Kaski, M. Ala-Korpela. Biochem Biophys Res Commun. 2007 Aug 17;360(1):290-4.

A key challenge in metabonomics is to uncover quantitative associations between multidimensional spectroscopic data and biochemical measures used for disease risk assessment and diagnostics. Here we focus on clinically relevant estimation of lipoprotein subclasses by 1H NMR spectroscopy of serum. Some subclasses are more important than others with respect to the risk for atherosclerosis, but the current clinical risk assessment methodology cannot take this properly into account. 1H NMR enables a fast measurement of the lipoprotein profile, and several computational methods have already been developed for this purpose [1]. We have recently presented a novel Bayesian approach to quantify clinical variables and to determine their spectroscopic counterparts in 1H NMR data [2]. The analysis focused on lipid concentrations of major lipoprotein fractions, namely VLDL, IDL, LDL and HDL to test the automated Markov chain Monte Carlo (MCMC) Bayesian inference with a well-known biochemical background and spectroscopic characteristics. The results illustrated a high-quality quantification ability of the presented Bayesian approach [2].

In this work we study if the MCMC Bayesian inference could also be used to quantify the lipoprotein subclasses from the 1H NMR spectra of serum. Using experimentally derived model signals for VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b, HDL2a, HDL3a, HDL3b and HDL3c we simulated a representative set of spectra that contained extensive variation in the lipoprotein subclass profiles. The simulated spectra also included individual variations in the lipoprotein subclass signals. Our preliminary Bayesian MCMC analysis results for the quantitative performance in the case of all the abovementioned 11 lipoprotein subclasses are very good for the simulated data set. This suggests that the Bayesian approach, while computationally demanding, could provide a framework for quantitative modelling of metabonomic NMR data. In line with our earlier analysis [1], the quantification accuracy is subclass dependent as illustrated in the adjacent figure. Even though the preliminary results show very good generalisation abilities with the simulated data, it is possible that the simulation process itself may not optimally represent the inter-subject variance of the spectral features for a realistic population. Thus, the results must be verified with extensive real data. The Bayesian approach is presented at the poster together with critical considerations of the pros and cons of the methodology.

• M. Ala-Korpela; Lankinen, N; Salminen, A; Suna, T; P. Soininen; R. Laatikainen; Ingman, P; Jauhiainen, M; Taskinen, MR; Heberger, K; K. Kaski. Atherosclerosis. 2007, 190, 352–358.
• Vehtari, A; Mäkinen, VP; P. Soininen; Ingman, P; Mäkelä, SM; Savolainen, MJ; Hannuksela, ML; K. Kaski; M. Ala-Korpela. BMC Bioinformatics. 2007, 8(Suppl 2), S8.

Alcohol has a major role in public health care and is of social importance to various communities. Thus, unsurprisingly, it is also an ever-lively issue in Finnish politics. Alcohol has multiple effects on the health of an individual on its own and also as modifying the response of the body to various other metabolites. Ethanol has an immediate effect on the neural system as a psychoactive drug. Sustained or excessive use might permanently impair cognitive functions. Alcohol abuse also damages the liver and increases the probability for progressive liver diseases such as fatty liver and cirrhosis. Alcohol affects energy and substrate metabolism in addition to having high energy levels (especially beverages with added sugar) and thus promote obesity. However, moderate alcohol consumption is known to have beneficial effects, one of which is the decreased risk for coronary heart disease. This is partly manifested via the increase of serum HDL cholesterol levels, but the exact molecular mechanisms behind these effects are yet to be established. More information on the effects of alcohol on serum lipoprotein subclasses would also be desirable.

NMR metabonomics is a readily utilisable method for quantifying body's metabolic response to changing environment, diet & drinking or (patho)physiological state. 1H NMR spectroscopy can be routinely applied to, for example, tissue extracts as well as to urine and serum samples. Data can be recorded in multiple molecular windows in which different metabolites are prominent. The resulting information rich datasets call for various statistical data analyses to retrieve the metabolic information and associations.

Here we report preliminary results from a 1H NMR metabonomics study of serum samples from 71 individuals with low, moderate and high alcohol consumption. We applied two molecular windows: a LIPO window (dominated by lipoprotein lipids) and a LMWM window (low-molecular-weight metabolites). These NMR data are complemented with an extensive set of biochemical and clinical measurements. We welcome all the symposium participants to our poster to discuss the pros and cons of alcohol consumption and to look at the colour-coded pseudo-2D NMR spectra resulted from the covariance analyses of the metabolite associations in the abovementioned molecular windows and groups of individuals. There will be no permillage limit to enter the poster.

Lineshape fitting is a commonly used strategy in obtaining quantitative information from NMR spectra. There is a great variety of software available for the analysis, but they are generally restricted to manual treatment of one spectrum at a time. There is, however, a need for effective ways to analyse large sets of spectra, for example when examining time series or metabonomics data. We introduce software that automates the process of analysis and data handling as well as provides the user with a modern graphical user interface and many timesaving interpretation features.

The software package is programmed around the well-known PERCH software (PERCH Solutions Ltd., Kuopio, Finland) [1] because of its various mathematical capabilities, speed and features, including the elegant options to use prior knowledge in the analyses. Our software works as illustrated in the figure at the bottom. First, the user creates a work queue by stacking the wished combinations of spectra and different running parameters controlling the functions of PERCH's lineshape fitting routine. When the batch process is started, the program automatically writes the input files, starts PERCH and reads in the results after the fitting is complete; a process that is repeated for every spectrum and every specified configuration in the work queue. In the interpretation section of the software, plots and tables of individual parameters from the optimised lineshape models can be made. Furthermore, composite variables can be created, for example, if a progression of a peak's area divided by the area of another peak is of interest (as shown above in the figure). One can also import external case-specific biochemical or other data to be incorporated into the interpretation phase. The gathered information can finally be exported for visualisation or further analysis.

A laptop will be available at the poster for demonstration and evaluation. We acknowledge R. Laatikainen and Matthias Niemitz from the PERCH-team for helpful information and all the important perching details; we also anticipate that they will be around the poster.

Alzheimer's disease (AD) is the most common dementia in the world. It is a progressive brain disorder that gradually destroys a person's memory and other cognitive abilities. The initiation and development of AD are poorly understood and there are no distinct biomarkers allowing for early detection and preventive treatment. The major focus in AD research has traditionally been in proteomics but in recent years there has been an increasing interest in the association of serum cholesterol (TC) and lipoproteins with the development of AD [1].

Cholesterol is a key component of cell membranes and exists in myelin sheath of nerve cell axons. Cholesterol is also known to affect proteins that are associated with the pathophysiological changes in the AD brain. Interestingly, cholesterol lowering drugs, such as statins, decrease the progression of AD [1]. Both higher as well as lower TC levels have been indicated to associate with the development of AD. However, only very limited data are available on the associated serum lipoprotein profiles of the AD patients.

We present here results from comprehensive meta-analyses of the TC studies in this field during 1986–2007 (n~9000 individuals). Indeed, our results reveal that the opposite findings concerning the role of TC and the development of AD can be explained by two distinct lipoprotein profiles: a "conventional" atherosclerotic risk profile (with high TC and LDL-C) and a profile associated with the metabolic syndrome (decreased TC and HDL-C plus elevated triglycerides). These results suggest that disturbed lipoprotein metabolism is connected to the development of AD. Moreover, the connection is not straightforward but two different metabolic pathways; also associated with atherosclerosis, seem to be involved. These new findings that relate lipoprotein profiles to the risk of AD also support the use of 1H NMR; the figure above illustrates an idea of early risk detection. In the presentation we will discuss the meta-analyses together with preliminary results from a prospective study of cognitive decline by 1H NMR.

• L. A. Shobab, G. Y. Hsiung, H. H. Feldman Lancet Neurol. 2005, 4, 841–852.

Atherosclerosis is a major cause of cardiovascular disease that is the most common cause of death in the Western World. Lipoproteins and more specifically the lipoprotein subclass distribution, i.e., the so-called lipoprotein profile, plays a key role in the development of atherosclerosis. Some subclasses are known to be more important than others with respect to the risk for atherosclerosis, but the current clinical risk assessment methodology cannot take this properly into account. 1H NMR spectroscopy enables a fast measurement of the lipoprotein profile from a serum sample, and several computational methods have been developed for this purpose [1]. The use of 1H NMR spectroscopy seems to have significant potential also in clinical use. However, the intrinsic capability and accuracy of this approach to identify and quantify the different lipoprotein subclasses has not been properly addressed.

In this work we studied the intrinsic accuracy of 1H NMR spectroscopy using various computational methods and an extensive, biochemically sound, set of simulated serum spectra. The overall lipoprotein subclass quantification accuracy was found high when i) the spectral variables were selected according to a high correlation to the lipoprotein subclass concentration, and ii) partial least squares (PLS) or an early stop-MLP committee, i.e., a regularized nonlinear function approximator, was used. There were clear differences in the methodological accuracy achieved for different lipoprotein subclasses. A rather surprising new finding was that the non-lipoprotein components, such as albumin, in the 1H NMR spectra of serum do not pose any difficulties for the quantification accuracy, but the largest deviations seem to originate from the interindividual variation in the composition and structure of the lipoprotein subclass particles.

The current results support the potential role of 1H NMR in the individual risk assessment of atherosclerosis. However, further studies will be needed to detail the effects of individual variations and the maximum quantitative information available from the lipoprotein subclass profiles in the 1H NMR spectra of serum.

• M. Ala-Korpela, N. Lankinen, A. Salminen, T. Suna, P. Soininen, R. Laatikainen, P. Ingman, M. Jauhiainen, M.-R. Taskinen, K. Heberger, K. Kaski. Atherosclerosis. 2007, 190, 352–358.

Identification of metabolic conditions for increased risk of atherosclerosis would be advantageous to facilitate individual primary prevention. Metabolic syndrome, with characteristic changes in the distribution of lipoprotein subclasses, is one such condition with an increasing prevalence in the general population. 1H NMR spectroscopy is a rather new approach to quantify lipoprotein subclasses along with a holistic overview of serum metabolites. Consequently, it can also be seen as a new methodology for screening individuals at high risk for atherosclerosis. Here we studied whether 1H NMR metabonomics would enable characterisation of lipoprotein subclass-related metabolism in a clinically relevant context. Using experimentally derived model signals for VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b, HDL2a, HDL3a, HDL3b and HDL3c, two biochemically characteristic categories of spectra were simulated, representing lipoprotein subclass profiles for a normolipidaemic and a metabolic syndrome status. A set of spectra representing a metabolic pathway between these two categories was also generated. The analysis, based solely on these 1H NMR spectra, clearly identified the lipoprotein subclass profiles and their changes. The findings are supported by comparable analyses for a representative set of experimental 1H NMR spectra of 69 serum samples with a wide range of lipoprotein lipid distribution. The results demonstrate the inherent suitability of 1H NMR spectroscopy to study lipoprotein subclass-related metabolic changes. In addition, the metabonomic strategy provides information on subtle changes in lipoprotein subclasses and other serum metabolites, thus being potentially useful in the evaluation and management of an individual's risk for atherosclerosis.

Macrovascular disease in connection to diabetic nephropathy accounts for most of the premature deaths in type 1 diabetes (T1DM). Metabolic syndrome (MetS) is common in the patients that are affected by these complications, which suggests that the risk factors are similar to those in the general population. The definitions of MetS, however, may not work in T1DM nor be accurate enough to detect subtle changes. Our aim is therefore to investigate the metabolic features and their associations to complications in T1DM patients. For this reason, we measured 1H NMR spectra of serum for a subset of T1DM patients (n = 463) from the FinnDiane multicenter study. Since the data are extensive and complex, we used self-organising maps (Figure) to visualise the multidimensional patterns of metabolism. Also, we did classification analysis to assess the descriptive power of 1H NMR, and found strong links between the spectra, metabolic syndrome and nephropathy. Based on this study, 1H NMR metabonomics yields a detailed and comprehensive picture of metabolic risk factors an advantage that is likely to have great value in the detection and treatment of diabetic complications.

The compositional variability within lipoprotein fractions and the metabolic relations between distinct particles are of fundamental interest. Analysis and visualisation of such multi-dimensional data is often hampered in conventional statistical tools. Here we introduce self-organising maps (SOMs) that transform multi-dimensional data into a two-dimensional map of individuals, where neighbours have similar lipid profiles. Each lipid variable can be visualised for the whole data set by colouring the map regions according to the average value of their local inhabitants. In this study VLDL, IDL, LDL, HDL2 and HDL3 particles were isolated and their lipid and protein compositions measured for 146 individuals. Twenty-five input parameters were used in the SOM analysis. An example of the results (for particle compositions per protein) is illustrated in the Figure. A group of individuals (upper boxes) with dominant VLDL1 (high TG and medium CE) also possess small TG-rich LDL and another group of individuals (lower boxes) with dominant VLDL2 (medium TG and high CE) have large CE-rich LDL with low TG. These results, together with those that will be discussed in the presentation, demonstrate that SOM analyses reveal relevant compositional and metabolic associations in lipoprotein data and thus provide a new tool for lipoprotein research.

Background: Adiponectin, an adipocyte-derived cytokine, has been recently suggested to have an important role in the pathophysiology of atherosclerosis. Adiponectin can suppress the secretion of inflammatory cytokines, such as TNF-alpha. Moreover, adiponectin may inhibit the formation of foam cells. Plasma adiponectin concentrations are lower in patients with coronary artery disease than in control subjects. The effects of alcohol consumption on plasma adiponectin levels are still controversial.

Objective: To study the effects of alcohol consumption on plasma adiponectin consentrations in heavy alcohol drinkers and controls. Plasma adiponectin concentration was measured in 49 male alcohol drinkers (median alcohol intake 155 g/day) and 46 control men (11 g/day).

Results: Mean plasma adiponectin concentration was 58% higher in the heavy alcohol drinkers (13.3 ± 4.0 ng/ml) than in the controls (8.4 ± 2.6 ng/ml) (P < 0.001, ANCOVA adjusted for BMI). Plasma adiponectin concentration correlated with HDL-C concentration after adjustment for BMI in the heavy alcohol drinkers (r = 0.444, P < 0.01) and in the controls (r = 0.519, P < 0.001). In addition, adiponectin was inversely correlated with VLDL-TG concentration (r = -0.507, P < 0.01) in the heavy alcohol drinkers after adjustment for BMI

Conclusions: In addition to high plasma HDL cholesterol concentration, high plasma adiponectin concentration is likely one of the anti-atherogenic features in heavy alcohol drinkers.

Identification of physiological conditions relating to increased risk of atherosclerosis would be advantageous to facilitate individual primary prevention. Metabolic syndrome, with characteristic changes in the distribution of lipoprotein subclasses, is such a condition with an increasing prevalence in the general population. 1H NMR spectroscopy is an alternative approach to quantify lipoprotein subclasses and can also provide a holistic overview of other serum metabolites. Here we studied whether the metabonomic tactic, i.e., 1H NMR spectroscopy together with chemometric data analyses, would enable characterisation of lipoprotein subclass-related metabolism in a clinically relevant context. By using experimentally derived model signals for VLDL1, VLDL2, IDL, LDL1, LDL2, LDL3, HDL2b, HDL2a, HDL3a, HDL3b and HDL3c, two biochemically characteristic categories of spectra were simulated representing lipoprotein subclass profiles for a normolipidaemic and a metabolic syndrome status. We also generated a set of spectra representing a metabolic pathway between these two categories. The analysis of the 1H NMR spectra clearly identified the lipoprotein subclass profiles and their changes. The findings are supported by comparable analyses for a representative set of experimental 1H NMR spectra of 69 serum samples with a wide range of lipoprotein lipid concentrations. The results demonstrate the inherent suitability of 1H NMR metabonomics to identify subtle changes in lipoprotein subclass-related metabolism. Accordingly, 1H NMR can be seen as a new methodology for screening individuals at high risk for atherosclerosis and also being potentially useful in prospective assessment of individual 'health paths' in the borderline of, for example, normal lipoprotein metabolism and the metabolic syndrome.

Treatment of LDL with phospholipase A2 (PLA2) leads to hydrolysis of the sn-2 fatty acyl ester bond in LDL phosphatidylcholine (PC), yielding one free fatty acid (FFA) and lysophosphatidylcholine (lysoPC) molecule. At physiological albumin concentration and at pH 7.4, most of the lipolysis products are transferred from LDL to albumin. Earlier studies have shown that the affinity of albumin for FFAs is decreased if the pH is lowered. Recently, it has been demonstrated that during atherogenesis, the pH in atherosclerotic lesions is decreased. Since the hydrolysis products, FFA and lysoPC, promote atherosclerosis, the objective of this study was to investigate the distribution of these hydrolysis products between LDL and human serum albumin (HSA) at various pH values. Samples at pH range 5.5-7.5 with LDL concentrations of 1.0-1.3 mg/ml and fatty acid-free HSA at a final concentration of 2% (w/v) were prepared. LDL particles were separated from HSA by ultracentrifugation and the amount of FFAs was determined in each LDL sample. In addition, distribution of the FFA and lysoPC between LDL and HSA were studied directly in the incubation mixtures by using a novel application of 1H NMR spectroscopy at 800 MHz. We found that the lower the pH, the higher was the amount of the hydrolysis products that were not transferred to HSA but remained in the PLA2-treated LDL particles. The NMR indicated that the amount of HSA-transferred molecules at pH 7.4 was almost double of the amount at pH 5.5. Such PLA2-treated LDL particles act as proinflammatory particles, the effects of which, at local acidic conditions in the intima, may be enhanced.

A novel identification and analysis of mobile cholesterol compounds (Cbs) in an experimental glioma model by 1H MRS in vivo is presented. Previous 1H NMR MAS experiments [1] and chromatographic findings of high cholesteryl ester content in BT4C gliomas [2] gave an idea to investigate whether certain peak height anomalies in 1H MR in vivo spectra of tumours could be explained by the presence of Cbs and their characteristic backbone resonance beneath the fatty acid resonances in the aliphatic region of the spectra.

Female BDIX (n = 10) rats carrying HSV tk+ BT4C tumours were treated with Ganciclovir to induce programmed cell death, in a similar fashion as before [1, 2]. The MR experiments were carried out in a 4.7 T magnet using the STEAM sequence (TE = 3 ms, NT = 256) to obtain single voxel spectra at days 0, 2, 4, 6 and 8 prior to treatment. All spectral analyses were done using the PERCH NMR software.

The lipid fatty acid resonances alone are not capable of explaining the experimental spectra (see Fig.). In fact, the Cb resonance was found to comprise 12.3 ± 0.8 % of the total MRS visible lipid signal in nontreated tumours. We found continual increase for CH= (p<0.05 at day 6, Student's t-test), which is in line with previous studies [1, 2] and a peaking trend for the Cb lipid signals (p<0.05 at day 4, Student's ttest) during the treatment. The likely molecular origin for the Cb resonance is mobile cholesteryl esters within the intracellular lipid vesicles, particularly in the case of untreated tumours. Apparent biochemical process facilitating observed lipid accumulation would be enhanced uptake of serum lipoproteins and degradation of mitochondrial membranes. Whatever cellular processes drive the lipid accumulation and the increase in the pool of mobile lipids, it is evident from our current findings that the contribution of Cbs can not be overlooked.

• J. Griffin, K. Lehtimäki, P. Valonen, O. Gröhn, M. Kettunen, S. Ylä-Herttuala, A. Pitkänen, J. Nicholson, R. Kauppinen. Cancer Res. 2003, 63, 3195.
• J. Hakumäki, H. Poptani, A. Sandmair, S. Ylä-Herttuala, R. Kauppinen. Nat. Med. 1999, 5, 1323.

In a 1H NMR spectrum, one metabolite can manifest several peaks, and the signal intensities are both biochemically and (patho)physiologically related. Furthermore, the data sets are extensive but redundant: one measurement can yield tens of thousands of data points, but the effective dimensionality is much less due to a smaller number of NMR-visible compounds. There is also heavy overlap of metabolite resonances. For these reasons, bridging the gap between biochemistry – as revealed by 1H NMR spectroscopy – and the relevant measures of current clinical practise poses methodological challenges.

Here we propose a hierarchical paradigm for knowledge discovery in NMR metabonomics. The key idea is to successively narrow the focus of attention to the most important spectral regions within the limitations of statistical analysis and computing resources. At the lowest level, unsupervised high-throughput screening algorithms are preferred, such as sample correlation structure, linear component analysis and self-organising map. These methods can quickly verify the integrity of the 1H NMR measurements and reveal their general statistical properties. At the next level, non-NMR data is incorporated into the analysis. This can be achieved by regression models and by comparing the results from the unsupervised learning methods with independent classifications from other sources. Also a network-type of visualisation for parameter correlation structures may prove useful. The emphasis is still on throughput: regression models are fitted separately using small frequency bands and computationally efficient methods such as partial linear least squares are preferred.

Finally, the most interesting non-NMR variables or parameter relationships are selected for detailed inspection. As an example, we have applied advanced Bayesian modelling to the analysis of clinical data on the metabolic effects of alcoholism. The method is based on adaptive kernel parameterisation of the spectra, and a full Bayesian inference of the model parameters. Although computationally heavy, this method exhibited excellent accuracy in predicting biochemical measures from spectral features (see Fig.). The same associations could not be seen as distinctly in the screening phase emphasising the call for advanced approaches for increased insights and the overall benefits of a hierarchical paradigm.

The molecular structure of circulating lipoprotein particles and their modifications are central in the development of atherosclerosis. It is currently established that each lipoprotein subclass has an individual role in lipid transportation and metabolism. Accordingly, each subclass of particles also has a unique role in the prediction of atherosclerosis risk. 1H NMR spectroscopy as a means to quantify lipoprotein subclasses directly from serum has received wide clinical and commercial interest. The experimental part is a fast procedure that contrasts favourably to other lipoprotein measurement protocols. As far as we know, however, this is the first study aiming to define the spectroscopic and molecular characteristics at the subclass level.

In this study we have separated 11 lipoprotein subclasses from twelve individuals using ultracentrifugation procedures. The lipid and protein composition of each subclass was determined using standard biochemical assays. The 1H NMR spectra of all the subclass samples were recorded at 310 K at 600 MHz. A double tube system with an external reference was used to enable absolute quantification. A lineshape fitting model, consistent for all the subclasses, was constructed for the lipid resonances in the main aliphatic spectral region. This kind of analysis lead to absolute quantification of all the main lipid resonances in each subclass spectrum and allowed quantitative comparison with the independent biochemical lipid assays. The PERCH NMR software was used for the spectral analysis.

Preliminary comparisons between the 1H NMR spectroscopy (illustration of some aliphatic resonances on the right) and the biochemical lipid assays show subclass dependent molecular interactions for various lipid molecules in the particles. These experimental findings will be presented together with structural data from our optimised geometric lipoprotein particle model. The results reveal that remarkable portions of triglyceride and cholesterol ester molecules are embedded in the surface phospholipid and protein monolayer: the smaller the lipoprotein particles the more pronounced appears the content of hydrophobic core lipids in the particle surface.

Accordingly, at the poster, some new clues on the molecular structure of lipoprotein particles will be free to discuss and to potentially digest.

Metabonomics is an emerging complementary post-genomic technology. In mammalian studies 1H NMR spectroscopy of body fluids is the most common single methodology applied for the assessment of metabolic phenotypes. Two typical simplifications often made in handling metabonomic data are the use of 'bins', i.e., summed spectral intensities over fairly broad regions (e.g., 0.04 ppm), and the application of rather uninformative chemometric methods, such as principal component analysis. We will present a straightforward and visual approach facilitating the use of full spectral resolution in the data analysis and allowing detection of intra- and intermolecular dependencies in metabonomic data sets.

The approach involves calculation of the correlation and covariance between all the data points in the whole data set and displaying the information in the form of colour-coded pseudo-2D NMR spectra as illustrated on the right in the case of simulated data: The simulated spectra representing the lipoprotein part of human serum were constructed as a sum of experimentally obtained model signals for eleven lipoprotein subclasses. By varying the levels of these subclasses, two clinically distinct categories of sum spectra were generated: a group of spectra corresponding to a normal lipoprotein profile (N) and another group corresponding to a lipoprotein profile in metabolic syndrome (MS), a common disturbance of lipoprotein metabolism related to increased risk for coronary heart disease. Internal heterogeneity was induced to these two categories resulting in 729 spectra for each category. In addition, a metabolic pathway (MP) consisting of 501 spectra combining linearly the N and MS categories was generated. While this simulation of spectra is a simplification, it does provide an extensive dataset, in which the biological variability and measurement errors are excluded, and, importantly, in which the metabolic changes in all the lipoprotein subclasses are exactly known. For example, the associated changes in VLDL (increase) and HDL2 (decrease) along the MP from N towards the MS are, even in the case of heavy signal overlap, immediately evident from the covariance pattern of the pseudo-2D spectrum; the statistical variation within N and MS groups produces a completely different pattern. At the poster we will also present a new finding of dissimilar relation of glucose and lipoprotein metabolism in two separate clinical studies of alcoholism and type 1 diabetes.

In oil chemistry a key question is how the molecular attributes of composite hydrocarbon mixtures can be associated with the macroscopic properties of the oil products. The hydrocarbon mixtures are characterized by legislative and customer defined properties. These requirements can be met with multiple variations in the chemical composition. The amount of aromatic components has been restricted in solvents and diesel fuels due to heath criteria and previous knowledge of their deteriorating influences on fuel quality. Consequently, the aromatic content in fuels has diminished and understanding the effects of paraffinic and naphtenic structures on fuel quality has become of increased value. Traditional product property analyses give only vague information on the chemical composition since the isoparaffin branching structure of middle distillates is not identified. In fact, there is no measure based on which accurate prediction of product properties would be possible. In this work we are aiming at increased insight into the complex and often non-linear relationships between product properties and molecular composition. Our results indicate that the combination of 13C NMR spectroscopy and self-organising neural networks offers some clear advantages over the traditional protocols used in the oil industry and also provides means for predicting the product properties based on the identified molecular attributes.

Self-organising maps are a class of unsupervised neural networks whose characteristic feature is their ability to map non-linear relations in multidimensional data sets into visually more friendly, two dimensional planes of nodes. In this particular case, as shown on the right, the input data to the SOM analysis were 2060 aromatic and 2266 aliphatic data points from the 13C NMR spectra. The SOM algorithm transforms the input data vectors into a two dimensional map in which each node will be represented by a single feature vector representing the original parameter space. The point density of the feature vectors follows roughly the probability density of the data thereby making SOM as a valuable tool for detecting similarities and groupings in a data set. A visual inspection of the grouping in the U-matrix above, demonstrating the distances between the nodes, clearly verifies that the 13C spectra are grouped on the basis of the spectral characteristics: the samples S43 and S20, with similar spectra, locate near in the SOM map in contrast to S19 and S20, with markedly different spectra. At the poster more details will be given on the associations between the 13C NMR data and the product properties of the oils.

It would be advantageous to detect molecular and cellular processes related to the early stages of developing atherosclerosis. This would clinically facilitate early individual primary prevention and also give a personal rationale to comply with lifestyle modifications and potential drug therapies. The recent applications of MR in atherothrombosis research suggest the potential usage of in vitro MRS for risk assessment and of in vivo MRI for direct detection of plaque composition and vulnerability. The scheme presented below can be seen as one option to elucidate the potential of MR in detecting individual intermediate atherothrombotic end points and utilising their prognostic value before the occurrence of a definite end point.