American Cancer Society Study
900,000 adults
Calle et al., NEJM, 348:1625-1638, 2003
Overall Relative Cancer Risk (BMI >40 vs 18.5 to 24.9):
M: 1.52 (1.13 – 2.05): F: 1.62 (1.40 – 1.87)
Increased risk of colorectal, pancreatic, liver, esophagus,
kidney, multiple myeloma, non-Hodgkin’s lymphoma,
gallbladder, prostate, breast, cervical, ovarian, uterine
“Current patterns of overweight and obesity in the United States
could account for 14% of all deaths from cancer in men and 20%
of those in women”
Guidelines for Healthy Weight
Nurses’ Health Study: Willett et al., NEJM, 341:427-434, 1999
BMI of 26 vs 21
Coronary Heart Disease:
Hypertension:
Type II Diabetes:
2x increase
2-3x increase
8x increase
Weight Change of 15 kg
Coronary Heart Disease:
Hypertension:
Type II Diabetes:
2x increase
2-3x increase
6x increase
Caloric Restriction (CR)
CR is an experimental paradigm in
which the dietary/caloric intake of a
group of animals is reduced relative to
that eaten by ad libitum fed controls
Caloric restriction is the most potent,
most robust, and most reproducible
known means of reducing morbidity
and mortality in mammals
Survival Data, 1987 Cohort, Casein Diet
Survival Rate
100
80
AL
DR
60
40
20
0
0
200
400
600
800
Days
1000
1200
1400
DR Reduces Morbidity:
Breast Cancer
• Delays tumor onset (initiation and promotion)
• Slows progression
• Can modulate oncogene penetrance
– v-Ha-ras tumors decreased 67%
• (Fernandes et al., PNAS 92:6494-6498, 1995)
• Can prevent carcinogen-induced tumors
– 7,12-demethyl-benz(a)anthracene (60% AL, 0% DR)
• Kritchevsky et al. Cancer Res 44, 3174-3177, 1984
– Even with high fat diet, tumor yield, size, burden
down 93-98%
• Klurfied et al. Cancer Res 47, 2759-2762, 1987
CR “Beneficial” Effects
• Lower oxidative stress
– “Better” redox balance
• “Improved” glucose metabolism
– Increased insulin sensitivity
– Reduced blood glucose
– Reduced diabetes risk
• Reduced inflammation
How do we study complex
biological/clinical problems?
How do we address such questions in
humans, where our ability to manipulate
and analyze the system is limited?
High Throughput
and/or Data Density Studies
•
•
•
•
Genomics/SNPs
mRNA expression arrays
Proteomics
Small metabolites
Metabolomics:
The –omics face of biochemistry
Measurement of changes in populations of
low molecular weight metabolites under a
given set of conditions
Fiehn
GENOME
TRANSCRIPTOME
PROTEOME
DISEASE
STATE
METABOLOME
ENVIRONMENT
HEALTH
STATE
GENOME
TRANSCRIPTOME
PROTEOME
DISEASE
STATE
METABOLOME
ENVIRONMENT
HEALTH
STATE
Sample Analysis
Sample Collection
Database Curation
Response (µA)
0.80
0.60
0.40
0.20
0.00
0.0
20.0
40.0
60.0
80.0
Retention time (minutes)
1
100.0
Objectively Defining
Class Identity
3 SD
2 SD
Actual
Mechanistic Insight
Drug Development
Toxicology
Classification
Prediction
Functional genomics
Sub-threshold studies
Others
AL8
AL7
AL5
AL1
AL4
AL3
AL2
AL6
DR8
DR6
DR5
DR7
DR1
DR4
DR2
DR3
1.0
Computational Modeling
of Metabolic Serotypes
Observed Values vs.
Predicted Values
0.8
0.6
0.4
0.2
0.0
2 SD
Predicted
Modeling
Metabolic Interactions
Following Biochemical Pathways
Bioinformatics
What we measure -- biochemically
Metabolites – small molecules
Pathways (eg, purine catabolites)
Interactive pathways (eg, amino acid metabolism)
Compound classes (eg, lipids)
Conceptually linked systems
eg antioxidants, redox damage products
What we measure -- conceptually
Biochemical constituents
Excretion products
Precursor – product
Balances (eg, redox systems)
“collection depots”
Flux
Snapshot view of biochemistry
Integrated signal from genome and environment
Short and long term status
Temporal image
Sub-threshold changes (eg (toxicology, nutrition)
Metabolomics – Some Advantages
Sensitivity
“silent phenotypes”/sub-threshold effects
Discovery
Knowledge base (ie, metabolic pathways)
Limited repertoire – simplifies possibilities
(2500 non-lipid endogenous metabolites??)
Metabolome integrates signal
Nature and Nurture -- genome and environment
Measurement of system status/defects
Metabolome has the fastest response time
Metabolomics – Some Disadvantages
Too Sensitive?
cohort effects, site effects, time effects
sample handling
individual metabolites responsive to multiple factors
genes, environment, health status, location
experiment design must account for all factors
controlled or fuzzy, multiple sources
Practical
Set-up costs
Possible need for multiple platforms (NMR, MS, HPLC)
early industry dominance – lots of propriety data
incompatible data standards
Metabolomics Technology
=
Metabolomics Platform
Biology
Analytical
Chemistry
Data
Analysis
Sample Analysis
Sample Collection
Database Curation
Response (µA)
0.80
0.60
Analytical
0.40
0.20
0.00
0.0
20.0
40.0
60.0
80.0
Retention time (minutes)
1
100.0
Objectively Defining
Class Identity
3 SD
2 SD
Actual
Mechanistic Insight
Drug Development
Toxicology
Classification
Prediction
Functional genomics
Sub-threshold studies
Others
AL8
AL7
AL5
AL1
AL4
AL3
AL2
AL6
DR8
DR6
DR5
DR7
DR1
DR4
DR2
DR3
1.0
Computational Modeling
of Metabolic Serotypes
Observed Values vs.
Predicted Values
0.8
0.6
0.4
0.2
0.0
2 SD
Predicted
Modeling
Metabolic Interactions
Following Biochemical Pathways
Bioinformatics
Sample Analysis
Sample Collection
Database Curation
Response (µA)
0.80
0.60
0.40
0.20
0.00
0.0
20.0
40.0
60.0
80.0
Retention time (minutes)
1
Data
Analysis
100.0
Objectively Defining
Class Identity
3 SD
2 SD
Actual
Mechanistic Insight
Drug Development
Toxicology
Classification
Prediction
Functional genomics
Sub-threshold studies
Others
AL8
AL7
AL5
AL1
AL4
AL3
AL2
AL6
DR8
DR6
DR5
DR7
DR1
DR4
DR2
DR3
1.0
Computational Modeling
of Metabolic Serotypes
Observed Values vs.
Predicted Values
0.8
0.6
0.4
0.2
0.0
2 SD
Predicted
Modeling
Metabolic Interactions
Following Biochemical Pathways
Bioinformatics
Sample Analysis
Sample Collection
Database Curation
Response (µA)
0.80
0.60
0.40
0.20
0.00
20.0
40.0
60.0
80.0
Retention time (minutes)
Biology
Mechanistic Insight
Drug Development
Toxicology
Classification
Prediction
Functional genomics
Sub-threshold studies
Others
1
100.0
Objectively Defining
Class Identity
AL8
AL7
AL5
AL1
AL4
AL3
AL2
AL6
DR8
DR6
DR5
DR7
DR1
DR4
DR2
DR3
1.0
Computational Modeling
of Metabolic Serotypes
Observed Values vs.
Predicted Values
3 SD
2 SD
Actual
0.0
0.8
0.6
0.4
0.2
0.0
2 SD
Predicted
Modeling
Metabolic Interactions
Following Biochemical Pathways
Bioinformatics
Caloric intake
as a case study
High points only
Ignoring details, other studies,
etc
Survival Data, 1987 Cohort, Casein Diet
Survival Rate
100
80
AL
DR
60
40
20
0
0
200
400
600
800
Days
1000
1200
1400
Hypothesis:
Long-term, low-calorie diets
induce changes in metabolism
that persist throughout the
lifespan
Predictions
• CR alters the sera “metabolome”
• There exists a “CR Serotype”
• …Part of “CR serotype” reflects beneficial
physiological status --- ie, serotype defines
health without reference to disease…
Goals
1)
2)
Insights into the mechanism of CR
Recognize CR in other organisms
(e.g., non-human primates)
3)
Biochemically determine the effective, longterm caloric intake of an individual
(e.g., for epidemiological studies)
4)
Identify predictive markers of disease
(e.g., to intervene/prevent/focus resources;
focus on diseases where intervention is possible)
Experimental Design
Model: F344 x BN F1 Rat
Overall Design:
AL/CR, male/female, 5 different ages
Different extents and duration of diets
Total experiment ~36 groups, 82 cohorts.
Approach:
HPLC separations with coulometric array detection
(LC/LC-MS for plasma proteomics)
Multilayer statistical and data analysis
Analytical Stability
Biologic Variability
Analytical vs Biological Variation
0.75
r ² = 0.75
0.36
0.24
0.12
0.00
0.00
AvA
0.45
0.45
0.30
0.30
0.15
0.15
0.12
0.24
0.36
0.48
0.00
0.00
0.60
0.15
0.30
0.45
0.60
0.00
0.00
0.75
0.15
0.30
r ² = 0.31
2.00
0.45
0.60
AL vs DR
ANALCV2
ANALCV1
1.50
r ² = 0.36
0.60
0.60
ANALCVT
ANALCVT
0.48
0.75
r ² = 0.55
ANALCV2
0.60
r ² = 0.25
1.50
r ² = 0.31
1.75
0.75
ANALCV1
2.00
1.50
0.50
1.25
FALCV
0.75
1.00
0.75
FDRCV
1.00
1.00
FDRCV
MDRCV
1.50
0.75
1.00
0.75
0.50
0.25
0.25
0.25
0.25
0.50
0.75
1.00
1.25
0.00
0.00
1.50
0.25
0.25
0.50
0.75
1.00
1.25
1.50
1.75
0.00
0.00
2.00
0.25
0.50
0.75
FALCV
MALCV
1.00
1.25
0.00
0.00
1.50
0.25
0.50
0.75
1.00
1.25
1.50
1.25
r ² = 0.01
0.75
0.50
2.00
2.00
r ² = 0.00
r ² = 0.01
1.00
1.50
ANALCV1
ANALCV1
1.00
1.75
1.75
1.25
1.00
0.75
0.50
ANALCV1
r ² = 0.01
1.25
1.50
MDRCV
MALCV
Biological vs Analytical
1.50
ANALCV1
1.25
0.50
0.50
0.00
0.00
r ² = 0.09
1.75
1.25
1.25
0.75
0.50
1.25
1.00
0.75
0.50
0.25
0.25
0.25
0.25
0.00
0.00
0.25
0.50
0.75
MALCV
1.00
1.25
1.50
0.00
0.00
0.25
0.50
0.75
1.00
MDRCV
1.25
1.50
0.00
0.00
0.25
0.50
0.75
FALCV
1.00
1.25
0.00
0.00
0.25
0.50
0.75
1.00
1.25
1.50
1.75
2.00
FDRCV
In Rats:
Biological variability 5 fold greater than analytical variability
Analytical variability does not influence biological variability
Primary Data Analysis
• Multivariate analyses are relatively noise-resistant
• Minimize loss of informative metabolites
• Reduce false negatives (Type II errors)
• Increase false positives (Type I errors)
Does Serotype Encode
Sufficient Information to
Identify Diet Group?
Data Exploration and
Classification Analysis
• Hierarchical Cluster Analysis (HCA)
– Identifies natural groups in data
• Principal Component Analysis (PCA)
– Finds linear combinations of original variables
that account for maximal variation
Model
Feature Selection
Biological Model
0.80
AL
DR
Response (µA)
Survival Rate
80
60
40
0.60
0.40
0.20
0.00
20
0.0
0
T-tests, p<0.2 ?!
1075 analytically detectable peaks
sensitivity ~300 pA = ~10 fmole/125 l sera
100
20.0
40.0
60.0
80.0
100.0
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
350
300
250
200
150
100
50
Retention time (minutes)
0
200
400
600
800
1000
1200
0
1400
Days
HCA
Proof of Principle
HCA Distinguishes Female AL and DR Rats
Single
Autoscale
Range
Scale
Linkage Method
Dependent
Categorization
Accuracy
AL8
AL7
AL5
AL1
AL4
AL3
AL2
AL6
DR8
DR6
DR5
DR7
DR1
DR4
DR2
DR3
AL8
AL5
AL7
AL1
AL4
AL3
AL6
AL2
DR8
DR7
DR1
DR4
DR2
DR3
DR6
DR5
Complete
AL
DR
AL
DR
1.0
1.0
0.8
0.8
0.6
0.4
0.6
0.4
100%
AL8
AL7
AL1
AL5
AL6
AL4
AL3
AL2
0.2
0.0
DR8
DR7
DR1
DR6
DR5
DR4
DR2
DR3
AL8
AL5
AL7
AL1
AL6
AL4
AL3
AL2
0.2
0.0
DR8
DR7
DR1
DR6
DR5
DR4
DR2
DR3
AL
DR
AL
DR
1.0
0.8
1.0
0.8
0.6
0.6
100%
0.4
0.4
AL8
AL7
AL1
AL5
AL6
AL4
AL3
AL2
DR8
DR7
DR1
DR6
DR5
DR4
DR2
DR3
0.2
AL8
AL5
AL7
AL1
AL6
AL4
AL3
AL2
0.2
0.0
DR8
DR7
DR1
DR6
DR5
DR4
DR2
DR3
AL
DR
AL
DR
1.0
1.0
0.8
0.8
0.6
0.6
100%
63 variables (confirmed), NonNon-independent samples
PCA Distinguishes AL and DR Female Rats
Preprocessing
Dependent
Categorization
Accuracy
Centroid
0.4
0.4
PCA
Autoscale
Range Scale


100%






Factor1
Factor1
0.2
Factor2
0.0
DR
100%







0.2




Factor3
AL
DR
Factor2


 
Factor3




0.0
63 variables (confirmed), NonNon-independent samples
AL
HCA
Validation
PCA Distinguishes AL and DR Female Rats
Complete
1.0
0.8
0.6
0.4
0.2
PCA
0.0
Factor3
AL11
DR16
DR15
DR13
DR14
DR12
DR11
DR9
DR10
AL16
AL15
AL12
AL10
AL13
DR
AL
DR
Factor2
AL
Factor1
AL9
AL14
63 variables (confirmed), independent female Cohort #2
94% Accuracy
HCA
Simplify Model PCA
PCA Distinguishes AL/DR Using Either
Dataset
Removed variables can contribute
to AL-DR separations
63 Variables
Autoscale, complete
63 Variables
1.0
0.8
0.6
0.4
36 Variables
0.2
0.0
1.0
DR21
DR21
AL21
AL21
AL17
AL18
AL20
AL
AL19
AL22
AL19
AL22
AL18
DR19
AL20
DR18
DR
DR17
91% Accuracy
0.6
0.4
AL
Factor2
0.2
0.0
Factor3
AL
AL
Factor1
Factor1
Factor2
DR19
AL17
DR20
0.8
36 Variables
DR20
DR18
DR17
DR
63% Accuracy
63/(37/36) variables (confirmed), Independent female Cohort #3
DR
DR
Factor3
37 variables (confirmed), Independent female Cohort #3
Status
Proof of principle accuracy:
HCA (100%)
PCA (100%)
Validation Accuracy:
HCA (94%)
PCA (100%) - subjective rotation
Simplification –
HCA (Fails)
PCA (100% Accuracy)
Use larger models?
Test components vs distance
“Expert Systems/Supervised
Analysis”
KNN
–
–
–
–
k-nearest neighbor analysis
Supervised HCA (HCA is KNN with K=1)
Distance-based metric
Strength is with small (training) datasets
SIMCA
–
–
–
–
Soft Independent Modeling of Class Analogy
Supervised PCA
Component-based metric
Strength is modeling flexibility (eg, group-specific interactions)
In our DR sera metabolomics
data – components greatly
outperform distance-based
algorithms
In OUR DR SERA
METABOLOMICS data –
components greatly
outperform distance-based
algorithms
Profiles are
cohort specific
Cohort Separations
male samples modeled
with male/female data set
4
2
0
-2
-4
-6
6
4
2
0
-2
8 6
-4
4 2
0 -2 -8-6
-4
AMAL
AMDR
BMAL
BMDR
CMAL
CMDR
Cohort Effects
2
Cohort Separations
4
10
8
6
4
2
0
-2
-4
-6
-8
6
4
t[1
]
4
-2
-4
-6
-8
8
-2
4
-6
-8
-10
-6
2
0
-2
t[2]
-10
-4
winsorize (3SD)
-4
6
-8
0
log
winsorize (2SD)
winsorize (3SD)
2
0
-2
10
-6
2
log
winsorize (2SD)
6
0
t[1
]
t[3]
t[3]
AMAL
AMDR
BMAL
BMDR
CMAL
CMDR
-4
log + win (2SD)
-8
-6
-8
-12
log + win (2SD)
-10
-10
log + win (3SD)
log + win (3SD)
0.0
(d) log + win (3SD)
(c) winsorize (3SD)
0.2
0.4
0.6
0.8
1.0
0.0
(c) Pareto scaling (females)
0.2
0.4
0.6
0.8
1.0
(d) Pareto scaling (males)
none
none
log
8
log
6
6
4
8
2
0
-2
-4
-6
-8
8
t[3]
4
-2
0
-2
8
-4
6
4
6
0
2
0
-8
-2
-4
-6
-8
-4
-6
-8
-8
-10
-12
log + win (3SD)
-10
log + win (3SD)
0.0
(f) log + win (2SD)
t[1
(e) winsorize (2SD)
log + win (2SD)
-6
-2
t[2]
-10
winsorize (3SD)
log + win (2SD)
-4
4
-6
2
t[2]
winsorize (3SD)
2
-4
-6
-8
t[1
]
-2
10
winsorize (2SD)
6
0
2
0
winsorize (2SD)
8
2
6
4
t[1
]
4
t[3]
t[3]
6
4
2
0
-2
8 6
-4
4 2
-6
0 -2 -8
-4
t[2]
(b) no scaling (males)
none
8
2
8
6
4
2
0
-2
-4
8
t[2]
0
-2
-4
-6
(a) no scaling (females)
6
10
2
PLS-DA
RX
2
RY
2
Q (cum)
(b) log
none
male samples modeled
with male/female data set
4
AMAL
AMDR
BMAL
BMDR
CMAL
CMDR
(a) none
0.2
0.4
0.6
0.8
0.0
1.0
]
(e) UV scaling (females)
none
6
8
0.4
0.6
0.8
1.0
0.8
1.0
none
log
log
winsorize (2SD)
winsorize (2SD)
winsorize (3SD)
winsorize (3SD)
log + win (2SD)
log + win (2SD)
4
6
t[3]
4
2
-4
6
4
2
t[1
]
-2
8
-2
-4
-8
-6
-8
-10
4
-2
-6
0
t[2]
6
0
-4
-6
-8
10
0
10
8
2
6
2
0
-2
-4
-6
-8
2
0
-2
8
6
4
t[1
]
4
t[3]
0.2
(f) UV scaling (males)
-4
2
0
-6
-2
t[2]
-4
-6
-8
-8
-10
-10
-12
log + win (3SD)
log + win (3SD)
-10
0.0
0.2
0.4
0.6
0.8
0.0
1.0
0.2
0.4
0.6
PLS-DA
8
t[2]
0 .4 0
-6
-12 -10 -8
- 0 .2 0
Comp[3]
0
-4
0 .0 0
Comp[2]
2
-2
0 .2 0
Comp[1]
54.808
0 .3 0
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
-6
-4
-2
0 2
t[1]
4
6
8
AL
0 .2 0
0.80
7.642
0.60
0.40
0 .0 0
-0 .1 0
-0 .2 0
0.20
CR
10 12
-0 . 1 0
0.00
0 .0 0
0 .1 0
w*c[1]
Predicted AL probability
46.142
45.417
20.167.642
40.825
34.783
34.992
23.5
13.525
55.658
75.483
30.242
61.092
70.017
34.9
9.817 52.858
31.617
45.308
9.44216.933
8.183
21.117
13.467
47.442
41.642 75.525
62.942
68.575
68.6
11.45
30.133 54.058
56.117
48.808
14.975
24.483
6.408
49.6
68.7
40.4
82.225
27.892 92.792
22.192 17.475
62.492
77.917
40.292
28.608
12.608
65.408 28.858
63.358
15.475
48.533
76.2
62.65
100.242
9.4593.892
46.725
46.575
43.033
29.317
84.808
22.467
5.067 77.442
25.492
44.325
16.925
26.392
54.283
5.042
40.367
35.5
27.142
69.05
37.358
44.225
39.417
47.9675.758
72.342
52.108
58.167
47.192
11.458
14.033
73.075
0 .1 0
w*c[2]
4
0 .6 0
p<0.001
1.00
6
0 .8 0
Actual AL probability
1 .0 0
0 .9 0
0 .8 0
0 .7 0
0 .6 0
0 .5 0
0 .4 0
0 .3 0
0 .2 0
0 .1 0
0 .0 0
0 .2 0
110
100
N = 90
Each group p<0.05
vs others
77.917
L
90
K
80
70
60
r ² = 0.877;
r ² = 0.994 (m eans)
50
40
40
50
60
70
80
90
100
110
16.92527.142
120
41.642
22.192
INTAKE
120
w*c[2]
Predicted Degree Restriction
Markers “Predict” Caloric Intake
with High Quantitative Accuracy
-- Proof of Concept --
34.783
70.017
21.117
8.183
77.867 68.7
84.808
44.325 13.525
6.408
26.392
56.117
54.808 52.85815.475
68.575
68.6
75.483
82.225
28.608 48.533
81.7
17.475
47.967 30.133 54.058
65.408
28.858
45.417
62.942
35.5
92.792
77.442
29.317
16.933
44.225
40.292
62.492
30.108
100.242
30.242 7.642
58.167
5.067
46.142
13.467
55.658
37.358
34.483
67.642
72.933
40.4 43.033
61.092
62.65
5.758
9.45
62.442
47.442
47.192
23.5
54.283
34.992
9.442 14.975
52.108 63.358
34.9
45.308
27.892
73.075
49.6
76.2
24.483
5.042
11.308
75.525
48.808
11.45
11.458
39.417
Actual Degree of Restriction
w*c[1]
25.492
O-PLS models built for better testing
female basic model.M1 (OPLS)
t[Comp. 1]/t[Comp. 2]
Colored according to value in variable female basic model(Group Y new)
Series (Variable Group Y new)
1 - 1.5
1.5 - 2
7
6
5
4
3
2
t[2]O
1
0
-1
-2
-3
-4
-5
-6
-7
-8
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
t[1]P
R2X[1] = 0.105822
R2X[2] = 0.162921
Ellipse: Hotelling T2 (0.95)
SIMCA-P 11 - 3/22/2006 2:34:11 PM
8
Iteratively improve
models – focus on
analytical robustness
Then test one model…
O-PLS Prediction Score (A.U.)
Validation: Across Lifespan
6
8
4
6
2
4
2
0
0
-2
-2
-4
-4
-6
6 12 18 24 30
AL
6 12 18 24 30
[Months of Age]
AGE
CR
-6
0
2
Validation: Duration
8
6
6
4
4
r²=0
2
2
0
0
-2
-2
-4
-4
4 30
CR
-6
-6
0
2
4
6
8
10 12 14 16 18
Weeks on CR Diet
0
Validation: Extent
6
r ² = 0.645
4
2
0
-2
-4
-6
16 18
0
10
20
30
Percent Restriction
40
OPLS S
0
-2
Validation: Extent
-4
0
10
20
30
40
50
Percent Restriction
6
4
Only including 8 week CR
No AL group
8
r ² = 0.338
6
Includes 12 and 18 AL and CR
r ² = 0.636
r ²0.636
OPLS Score AU
OPLS Score AU
4
2
0
2
0
-2
-4
-2
-6
-8
-4
0
10
20
30
Percent Restriction
8
6
Includes 12 and 18 AL and CR
r ² = 0.636
r ²0.636
40
50
0
10
20
Percent Restriction
30
40
Experimental Design
Analytical
Issues
AL vs DR
Biological
Issues
Analytical vs Biological Variation
0.75
r ² = 0.75
0.36
0.24
0.12
0.00
0.00
AvA
0.45
0.45
0.30
0.30
0.15
0.15
0.12
0.24
0.36
0.48
0.00
0.00
0.60
0.15
0.30
0.45
0.60
0.00
0.00
0.75
0.15
0.30
r ² = 0.31
2.00
0.45
0.60
AL vs DR
ANALCV2
ANALCV1
1.50
r ² = 0.36
0.60
0.60
ANALCVT
ANALCVT
0.48
0.75
r ² = 0.55
ANALCV2
0.60
r ² = 0.25
1.50
r ² = 0.31
1.75
0.75
ANALCV1
2.00
1.50
0.50
1.25
FALCV
0.75
1.00
0.75
FDRCV
1.00
1.00
FDRCV
MDRCV
1.50
0.75
1.00
0.75
0.50
0.25
0.25
0.25
0.25
0.50
0.75
1.00
1.25
0.00
0.00
1.50
0.25
0.25
0.50
0.75
1.00
1.25
1.50
1.75
0.00
0.00
2.00
0.25
0.50
0.75
FALCV
MALCV
1.00
1.25
0.00
0.00
1.50
0.25
0.50
0.75
1.00
1.25
1.50
1.25
r ² = 0.01
0.75
0.50
2.00
2.00
r ² = 0.00
r ² = 0.01
1.00
1.50
ANALCV1
ANALCV1
1.00
1.75
1.75
1.25
1.00
0.75
0.50
ANALCV1
r ² = 0.01
1.25
1.50
MDRCV
MALCV
Biological vs Analytical
1.50
ANALCV1
1.25
0.50
0.50
0.00
0.00
r ² = 0.09
1.75
1.25
1.25
0.75
0.50
1.25
1.00
0.75
0.50
0.25
0.25
0.25
0.25
0.00
0.00
0.25
0.50
0.75
MALCV
1.00
1.25
1.50
0.00
0.00
0.25
0.50
0.75
1.00
MDRCV
1.25
1.50
0.00
0.00
0.25
0.50
0.75
FALCV
1.00
1.25
0.00
0.00
0.25
0.50
0.75
1.00
1.25
1.50
1.75
2.00
FDRCV
In Rats:
Biological variability 5 fold greater than analytical variability
Analytical variability does not influence biological variability
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
Profile = …
• Recent Food (ie, Fast)
–Effect size ~ 0.2*StDev
• BMI
• Food intake
• Physiology
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
BMI
Profile Score Does Not Reflect BMI
45
45
40
40
35
35
30
30
25
25
20
20
15
15
-8
-6
-4
-2
0
2
4
Profile Score (Individual)
6
8
-8
-6
-4
-2
0
2
4
Profile Score (3 point mean)
6
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
FFQ vs score (individuals)
y = 0.001x - 1.9244
R² = 0.033
8
6
4
2
0
0
-2
-4
-6
-8
500
1000
1500
2000
2500
3000
Extreme FFQ Quintiles I
• No obvious association with individual
FFQs, but…
• FFQs are known to be weakly
predictive of individual caloric intake
• But extreme quintiles on the FFQ
should differ – or we would likely never
see anything in epidemiology
Extreme FFQ Quintiles II
0.60
0.40
0.20
y = 0.0011x - 2.0509
R² = 0.834
0.00
-0.20 0
-0.40
1000
2000
3000
Individual values in
extreme quintiles
differ, p<0.005
-0.60
-0.80
FFQ Quintiles “predict” score
Extreme FFQ Quintiles II
0.60
0.40
0.20
y = 772.21x + 1881.5
R² = 0.8336
y = 0.0011x - 2.0509
R² = 0.834
2500
2000
0.00
-0.20 0
3000
1500
1000
2000
3000
1000
-0.40
500
-0.60
0
-0.80
-1.00
-0.50
FFQ Quintiles “predict” score
Score “predicts” FFQ quintile
0.00
0.50
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Testing Physiology
• Hypothesis: Profile score will differ/predict future risk
of diseases reduced by CR, eg, breast cancer
• Proposed 1000 cases/paired controls for breast cancer
– Nested within Nurses’ Health Study
– Samples taken 2-10 years before onset
• Initial funding, 2 years, 750 case control pairs
• Today, 210 case:control pairs (first blind break)
Case-Controls - Design
Using the raw RAT models, no corrections, no
optimization except for those that addess analytical
issues in humans
All observations that served as both cases and
controls (due to conversion) were excluded for this
analysis
210 total case/control pairs scored
The original model currently the best, so I'll present
those numbers...others are close
Case-Controls - Data
These slides not cleared for posting at cohort study level,
contact me at
bkristal@partners.org if you have questions
Human Studies:
Profile = …
•
•
•
•
Recent Food (ie, Fast)
BMI
Food intake
Physiology
Summary
Created and validated a working model of the CR serotype
in both male and female rats
Profiles distinguish diet in blinded studies
Markers pass analytical tests in human plasma
Rat profiles pass key tests in human case/control studies
Other studies ongoing, lipids, macronutrient shifts
The metabolomic markers and
profiles identified appear
analytically and biologically
suitable for studies in defined
human populations such as
national clinical trials and
epidemiological cohorts
Data suggests that the protective phenotype
exemplified in the CR rat is broadly conserved
across species
Data suggests that we have the “raw material”
needed to build marker profiles that can be tailored
to provide a continuous ruler for objective
measurement of factors such as caloric intake in
epidemiological and clinical studies
Data suggests that we have the “raw material” to
begin to address personalized disease risk from the
nutritional/environmental side
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Opportunities
Genetics
Demographics
Questionnaires
Clinical data
Metabolomics, proteomics, lipidomics = GxE
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Complications
Genetics…low effect size, low proportion explained, low
penetrance, multi-gene effects, sequence vs SNP,
Epigenome vs sequence, tissue specificity
Demographics…definition and measurement problems
Questionnaires…people lie, people are biased liars,
questions are *hard* and often population specific
Clinical data… limited, costly, by the time you have it
may be too late, thresholds, different data
structure/distributions
Metabolomics… proteomics, lipidomics = GxE, very
complex interactions, time effects, diet effects
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Complications
Are we limited to decision level analysis?
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Complications
Are we limited to decision level analysis?
No… Biology-based fusion works
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Step 1
What is Success?
What is Failure?
How do we balance Success or Failure
What does it mean to optimize?
How do we assess?
Long-Range Disease Risk Prediction – Algorithmic
Information Fusion in a Life and Death Environment
Step 1
What is Success?
What is Failure?
How do we balance Success or Failure
What does it mean to optimize?
How do we assess?
Long-Range Disease Risk Prediction
Algorithmic Information Fusion
in a Life and Death Environment
Long-Range Disease Risk Prediction
Algorithmic Information Fusion
in a Life and Death Environment
Long-Range Disease Risk Prediction
Algorithmic Information Fusion
in a Life and Death Environment
Long-Range Disease Risk Prediction
Algorithmic Information Fusion
in a Life and Death Environment
Acknowledgements
• Brigham and Women’s Hospital and Burke/Cornell
– Yevgeniya Shurubor, Honglian Shi, Diane Sheldon,
Sophie Guo, Susan (Schiavo) Bird, Rose Gathungu
– Ugo Paolucci, Ruiwen Zhou, Vasant Marur, Neil Russell,
Matt Sniatynski
• Outside Collaborators
– Walter Willett, Sue Hankinson, Frank Hu, Paul Vouros
– Wayne Matson, Karen Vigneau-Callahan, Paul Milbury
– Tom Vogl, Frank Hsu, Christina Schweikert
• Funding
– NIH (NIA; NCI; NIEHS)
– Winifred Masterson Burke Relief Foundation
– Brigham and Women’s Hospital, and Dept of
Neurosurgery
Acknowledgments
Bioinformatics
Analytical Work
Animal Studies and Mitochondria Research