A Novel Bayesian Approach for Uncovering

Potential Spectroscopic Counterparts for

Clinical Variables in 1H NMR Metabonomic Applications

Aki Vehtari1*, Ville-Petteri Mäkinen1,2, Pasi Soininen3, Petri Ingman4,

Sanna Mäkelä5, Markku Savolainen5, Minna Hannuksela5,

Kimmo Kaski1, and Mika Ala-Korpela1*

1Laboratory of Computational Engineering, Systems Biology and Bioinformation Technology, Helsinki University of Technology, P.O. Box 9203, FI-02015 HUT, Finland;

2Folkhälsan Research Center, University of Helsinki, Finland;

3Department of Chemistry, University of Kuopio, Finland;

4Department of Chemistry, University of Turku, Finland;

5Department of Internal Medicine, University of Oulu, Finland.

{*Aki.Vehtari, *Mika.Ala-Korpela}@hut.fi

Protein lipid aggregates

The ‘omics’ Revolution and Systems Biology

Lipoproteins – the lipid transporters in human circulations

Metabo*omics

A T H E R O S C L E R O S I S

Underlies the clinical conditions leading to death of

approximately half of the people in Western countries.

A systemic disease characterised by the local build-up of

lipid-rich plaques within the walls of large arteries.

The Trade-Off between Metabolic Coverageand the Quality of Metabolic Analysis

Fernie, Trethewey, Krotzky and Willmitzer, Nat Rev Molec Cell Biol 5, 1 (2004).Systems Biology & the ‘omics’ Revolution

Feasible, done

To be explored… … metabo*omics…

Lipoprotein subclasses are a key issue in atherothrombosis

5 10 20 40 60 80 1000

Diameter (nm)

1.20

1.10

1.06

1.02

1.006

0.95

Den

sity

(g/

ml)

HDL2

HDL3

ChylomicronRemnants

VLDL

IDL

Chylo-microns

Lp(a)

LDL

http://www.liposcience.com/

> million NMR LipoProfile® tests

J. D. Otvos et al., LipoScience

Inc. -CH3

15 subclasses

1 spectrum at a time

Quantification of Biomedical NMR

Data using Artificial Neural Network

Analysis: Lipoprotein Lipid Profiles

from 1H NMR Data of Human Plasma

Ala-Korpela, Hiltunen and Bell. NMR in Biomedicine 8, 235 (1995)

1H NMR biochemistry versus clinical biochemistry

Metabolic information by 1H NMR spectroscopy of serum

Principal Component Analysis

Lipoprotein Subclass Profiles via 1H NMR SpectraSelf-Organising Maps

SOM – rather easy and rather fast

The SOM clearly organised according

to the lipoprotein subclass profiles,

i.e.,

according to the spectral information

in the lipoplasma spectra.

Lipoprotein Subclass Profiling by 1H NMR

METABOLIC SYNDROME

METABOLIC

PATHWAY

NORMAL

-N(CH3)3 region / SOM U-matrix

Suna, et al., NMR in Biomedicine, submitted.

1H NMR biochemistry versus clinical biochemistry

Metabolic and

other individual

characteristics

Metabolite

profiles of

pre-dose biofluids

Inter-subject

variation in

effects of drugs

Influence

Influence

Predictable…?!

1H NMR biochemistry versus clinical biochemistry

Individual Risk Assessment and Diagnostics (of Atherothrombosis)

• T H E R E I S A C A L L F O R

M E T A B O N O M I C A P P R O A C H E S …

… p a r t i c u l a r l y s i n c e:

The 1H NMR Profile of Serum –in principle– Contains ALL

the Relevant Information for the CHD Risk Assessment

= Lipoprotein Subclasses + many other metabolites…

1H NMR Spectra of Human Serum at 500 MHz

Molecular windows

A Novel Bayesian Approach for Uncovering

Potential Spectroscopic Counterparts for

Clinical Variables in 1H NMR Metabonomic Applications

Aki Vehtari1*, Ville-Petteri Mäkinen1,2, Pasi Soininen3, Petri Ingman4,

Sanna Mäkelä5, Markku Savolainen5, Minna Hannuksela5,

Kimmo Kaski1, and Mika Ala-Korpela1*

1Laboratory of Computational Engineering, Systems Biology and Bioinformation Technology, Helsinki University of Technology, P.O. Box 9203, FI-02015 HUT, Finland;

2Folkhälsan Research Center, University of Helsinki, Finland;

3Department of Chemistry, University of Kuopio, Finland;

4Department of Chemistry, University of Turku, Finland;

5Department of Internal Medicine, University of Oulu, Finland.

{*Aki.Vehtari, *Mika.Ala-Korpela}@hut.fi; vmakine2@lce.hut.fi

Objectives and requirements

● Quantitative target: Estimating the value of a

clinical variable from 1H NMR spectrum.

- Accuracy must be maximized.

● Explanatory target: What are the spectral

features that best explain the clinical variable?

- Results must be easy to interpret.

● The two requirements can be conflictive.

Dataset

●100 serum samples from an ongoing clinical study of the effects of alcoholism (Dept Internal Medicine, Univ Oulu).

●Two 1H NMR molecular windows (LIPO & LMWM) were measured from each sample.

●A 500 MHz NMR-spectrometer with a double-tube system that enables absolute metabolite quantification.

●Automatic sample changer (24 samples in 16h).

VLDL / LDL / HDL

Overlapping resonances

Phospholipid choline headgroup

Lipoprotein spectra

TriglyceridesHDL

particles

Lipid signals

Lactate doublet

Cholesterol

Low-molecular weight metabolites

Glucose peaks

Creatinine

Acetate

Lactate

AlanineValine

Creatinine

Correlation analysis

HDL particles

Triglycerides

Abnormal triglycerides and HDL cholesterol are associated with cardiovascular diseases and are components of the metabolic syndrome, a clinically established condition with increased risk for atherosclerosis.

1H NMR biochemistry versus clinical biochemistry

Bayesian inference*

Regression model- robust against outliers- linearity preferred

Feature extraction- biologically motivated- easy to interpret- relevant wrt to target

Feature selectionFeature weights

Spectral parameterisation

*This is only a schematic view and not a precise description of the posterior density.

Regression model

●Heteroscedastic linear regression with scale-mixture Gaussian noise model (asymptotically Student-t).

●μ is the α2U-scaled predictor, effectively reducing the effect of outliers.

●The purpose of α is to improve convergence of Gibbs’ sampling (Gelman et al. 2004).

Kernel-based feature extraction

width

location3σ

Spectral parameterisation

●Gaussian kernels truncated at 3σ.

●Each kernel is represented by width and location.

●Posterior widths and locations were obtained by slice sampling from [0.008, 0.8] ppm.

●The number of kernels was obtained by reversible jump MCMC (max 20 kernels).

●MCMCStuff software for Matlab was written byAki Vehtari et al.

Additional details

● The degrees of freedom for the residual

model was obtained by slice sampling

within [2, 40].

● Prior for the number of kernels was a

decaying exponential to eliminate the “tail

saturation” effect.

● Ten independent chains of 10 000 samples.

● Predictive replicates were found to closely

match10-fold cross-validation results.

Quantification results

Non-spectroscopic measurement

1H

NM

R

VLDL-TGR2 = 0.97

HDL-CR2 = 0.87

n = 75 n = 67

1H NMR biochemistry versus clinical biochemistry

Relevant spectral features for VLDL-TG

Frequency [ppm]

Best observable VLDL-TG signal is located at the biochemically expected frequencies. -CH3

(-CH2-)n

-N(CH3)3

Relevant spectral features for HDL-C

Frequency [ppm]

Best observable HDL cholesterol signals are in the phospholipid choline headgroup region

-CH3

(-CH2-)n

-N(CH3)3

Conclusions

●Kernel based parameterisation is able to describe the data effectively.

●A Bayesian treatment with linear regression gives both accuracy in quantification and relative ease of interpretation.

●The results are biochemically fully coherent.

Future work

●Continued assessment of usefulness of different approaches within clinical research environment (interpretation).

●Systematic testing of kernel-based regression with Bayesian and other approaches (accuracy).

●Analysis across the full frequency range in both molecular windows.

●Application of Bayesian kernel parameterisation to classification problems.

Good life is antiCHD…!It is most probable that the

Bayesian methodology will

have a crucial role in paving

the way for metabonomics

on the clinical arena.

Life is about probabilities (and lipoproteins)…

THANK YOU