In lwjohnst86/seminar2016: Slides and Data Analysis for PhD Seminar Fall 2016

Physiology of serum lipid fractions

devtools::load_all()

knitr::include_graphics('../img/lipid-fractions.pdf')

Glucose and fatty acid metabolism

knitr::include_graphics('../img/glucose-metabolism.pdf')

Various fatty acid length and desaturation

• Range in length and number of double bonds
• Fatty acids either from diet or de novo lipogenesis (DNL)
• Physiological role dependent on molecule
• Eg: higher palmitic acid (16:0) lipotoxic to beta-cells in vivo and in vitro[^lit]

[^lit]: @Giacca2011a; @Xiao2009a

Few large cohorts on fatty acid composition, fraction, and diabetes

• One study (METSIM) had three fractions: TAG, PL, CE^key
  • Multiple flaws
• Mainly cohorts report on PL and CE: CHS, EPIC, ARIC[^cohorts]
  • 16:0 and 18:0 higher risk for DM
  • 18:1n-7, 18:1n-9, 18:3n-3 lower risk for DM

[^cohorts]: @Wang2003a; @Forouhi2014a; @Kroger2011a; @Ma2015a; @Djousse2011a

Objectives:

Explore associations of fatty acid composition of serum lipid fractions on diabetes pathogenesis:

. . .

• NEFA: Higher total NEFA, not individual fatty acids, contribute to lower beta-cell function
• PL: Higher palmitic acid associates with declines in beta-cell function over time. Higher cis-vaccenic acid associated with higher insulin sensitivity and beta-cell function.
• CE: No strong associates with diabetes pathogenesis
• TAG: ...

Data source: The PROMISE cohort

\vspace{-3cm} \raggedleft\includegraphics[width=0.11\textwidth]{../img/promise.jpg}

PROspective Metabolism and ISlet cell Evaluation cohort.

• Recruited from London and Toronto centers
• At-risk for diabetes
• Followed every ~3 years (3 time points completed)
• Demographics, lifestyle, anthropometrics, and blood
• n=477 participants with fatty acids measured

Variables of interest

Metabolic outcomes

Calculated from OGTT:

• Insulin sensitivity: 1/HOMA-IR, ISI
• Beta-cell function: IGI/HOMA-IR, ISSI-2

. . .

Median declines of r calc_outcome_changes()$chg

. . .

TAG fatty acids

Thin layer chromatography to split the lipid fractions, gas chromatography for the fatty acids:

• 22 TAG fatty acids, as concentration (nmol/mL) and percent of total (mol%)

TAG fatty acid composition within PROMISE

plot_tagfa()

Statistical analysis

. . .

R code for these results:

https://github.com/lwjohnst86/seminar2016

knitr::include_graphics('../img/share-code.png')

Statistical analysis: Generalized estimating equations (GEE)

Variables GEE model:

Visit number, waist size, baseline age, ethnicity, sex, ALT (marker of liver fat), physical activity (MET), and total NEFA.

Time-independent: TAG, NEFA, baseline age, ethnicity, sex

. . .

• Concern: multiple models will be computed
• P-values: generally unreliable, especially with more tests[^1]
• Adjust using BH False Discovery Rate (FDR) correction

[^1]: See the American Statistical Association statement on it

fit <- analyze_gee()
num_sig <- paste0(nrow(dplyr::filter(fit, unadj.p.value < 0.05)),
                  " non-FDR vs ",
                  nrow(dplyr::filter(fit, p.value < 0.05)),
                  " FDR")

As conc, strong negative association with IS (r num_sig of r nrow(fit) models)

. . .

fit %>% 
    dplyr::filter(unit %in% c('Totals', 'nmol/mL')) %>% 
    plot_gee_main()

As mol%, very different story --- different FA have positive or negative roles

fit %>% 
    dplyr::filter(unit %in% c('Totals', 'mol%')) %>% 
    plot_gee_main()

But... GEE modeling is limited

• TAG fatty acid composition in inherently multivariate

. . .

grViz_Rmd("
digraph {
    node [shape = none]
    subgraph {
    rank = min; '14:0'; '16:0'; '18:0'; '20:0'
    }
    subgraph {
    rank = max; '14:1n7'; '16:1n7'; '18:1n9'
    }
    '14:0' -> {'16:0', '14:1n7'}; '16:0' -> {'18:0', '16:1n7'}
    '18:0' -> {'18:1n9', '20:0'}; '14:1n7' -> '16:1n7'
}
")

Correlation between TAG fatty acids

analyze_corr() %>% 
    plot_heatmap()

Partial Least Squares (PLS) allows for multivariate data

Takes:

$$ISI = 140 + 141n7 + ... + 225n3$$

Converts to:

$$ISI = Comp1 + Comp2$$

. . .

• PLS: No p-value, no p-value problem
• Cross-validation (CV) determines predictability
• CV randomly splits data into training and test sets
• Limitation: Can only use one time point (cross-sectional) and no covariates

Four long chain fatty acids (14:0, 14:1n-7, 16:0, 16:1n-7) cluster and strongly explain the variance in metabolic function

fit_is <- analyze_pls(y = 'lISI', ncomp = 2)
pred_is <- paste0(
    calc_pred_corr(fit_is, fit_is$test_data, ncomp = 1)$r,
    '-',
    calc_pred_corr(fit_is, fit_is$test_data, ncomp = 2)$r
)
fit_bcf <- analyze_pls(y = 'lISSI2', ncomp = 2)
pred_bcf <- paste0(
    calc_pred_corr(fit_bcf, fit_bcf$test_data, ncomp = 1)$r,
    '-',
    calc_pred_corr(fit_bcf, fit_bcf$test_data, ncomp = 2)$r
)
gridExtra::grid.arrange(
    plot_pls(fit_is) + ggplot2::ggtitle(paste0('A: ISI\nPredictive: r = ', pred_is)), 
    plot_pls(fit_bcf) + ggplot2::ggtitle(paste0('B: ISSI-2\nPredictive: r = ', pred_bcf)), 
    ncol = 2)

FA involved in DNL from higher carb intake associate with lower metabolic functioning

• Upregulated DNL, increased 14 and 16 chain fatty acids[^dnl]
  • 16:1n-7 shown to be highly related to directly measured DNL
  • Shown to be lipotoxic

. . .

• Two other cohort studies[^prospect] had similar findings for diabetes and HOMA-IR.

[^prospect]: @Rhee2011a; @Lankinen2015a [^dnl]: @Lee2015a; @Wilke2009a

Overall conclusions of PhD research

• Each lipid fraction behaves slightly differently on metabolic functioning
• Fatty acids from DNL may contribute to metabolic dysfunction
• DNL fatty acids may be useful biomarker for clinical use

. . .

• ...Make use of statistical and analytical advances

Acknowledgements

• Supervisor: Dr. Anthony Hanley
• Co-Supervisor: Dr. Richard Bazinet
• Committee Member: Dr. Adria Giacca
• Hanley Lab: Ingrid Santaren, Zhila Semnani-Azad, Windy Wang
• Research Nurses: Jan Neuman, Paula Van Nostrand, Stella Kink, Annette Barnie, Sheila Porter, Mauricio Marin
• Funding: CDA, CIHR, BBDC

Code: https://github.com/lwjohnst86/seminar2016

\image[width=0.2\linewidth]{../img/CDA-logo2010} \hspace{1cm} \image[width=0.2\linewidth]{../img/CIHRlogos2013-Small} \hspace{1cm} \image[width=0.3\linewidth]{../img/BBDC-logo}

References

\tiny

lwjohnst86/seminar2016 documentation built on May 21, 2019, 9:15 a.m.