Estimation of relative protein abundance and statistical analysis of proteomic data from multiple itraq experiments

Transcription

1 Estimation of relative protein abundance and statistical analysis of proteomic data from multiple itraq experiments Ingo Ruczinski Department of Biostatistics Johns Hopkins Bloomberg School of Public Health April 17, 2012 Acknowledgments Shelley Herbrich Keith West, Bob Cole, Kerry Schulze, Parul Christian, Peter Scholl, Alain Labrique, Jim Yager, John Groopman. Funding from the Bill and Melinda Gates Foundation (#521).

2 What is Hidden Hunger? Hidden hunger is unlike the hunger that comes from a lack of food. It is a chronic lack of vitamins and minerals that often has no visible warning signs, so that people who suffer from it may not even be aware of it. Its consequences are nevertheless disastrous: hidden hunger can lead to mental impairment, poor health and productivity, or even death. One in three people in the world suffer from hidden hunger. Women and children from the lower income groups in developing countries are often the most affected. Prevalence of low micronutrient status Jiang et al, J Nutr (2005).

3 Strategies for preventing micronutrient deficiencies However Lacking is an ability to routinely and quantitatively assess population status with respect to multiple essential and potentially deficient micronutrients to Identify highest risk populations Estimate prevalence and severity Monitor population trends over time Inform, target and guide program development Evaluate and improve programs

4 Tandem mass spectrometry itraq Isobaric Tag for Relative and Absolute Quantitation

5 itraq Isobaric Tag for Relative and Absolute Quantitation Masterpools

6 Spectra files Variability in ratios SD ( ŶS Ŷ R ) Y S Y R (σŷs Y S ) 2 2ρ {ŶS,Ŷ R } ( )( σŷs σŷr Y S Y R ) ( ) 2 σŷr + Y R = ( ) 1 Y R σ 2 2ρ Ŷ S { Ŷ S,ŶR} σ YS σŷr Ŷ S Y R ( ) 2 + σ 2 YS Ŷ R Y R

7 Spectra files Reporter ion intensities gi gi gi5065 gi Run with 8 technical replicates of a master pool

8 Log reporter ion intensities (median polish) gi gi gi gi Run with 8 technical replicates of a master pool Log reporter ion intensities (median polish, by peptide) gi gi gi gi Run with 8 technical replicates of a master pool

9 Estimated relative protein abundance gi gi gi gi Run with 8 technical replicates of a master pool Missing data

10 Missing data Statistical models [KERR ET AL J COMP BIO 7: 2000 ]

11 Statistical models [HILL ET AL J PROT RES 7: 2008 ] Estimable parameters The true protein abundance for subject k {1,...,8} in a particular experiment is a k = µ + + δ k. However, for each spectrum s {1,...,S} of log 2 reporter ion intensities we only observe Note that (since k δ k = 0) Y sk = s + δ k + ɛ sk E[Ȳs] = s. Thus, for the de-meaned log 2 reporter ion intensities we have E[Z sk ]=E[Y sk Ȳs] =E[Y sk ] E[Ȳs] =δ k. Also note that the errors for the de-meaned reporter ion intensities are not independent.

12 Estimable parameters The question arises how information from multiple itraq runs can be combined. Housekeeping proteins can not be leveraged for the data normalization across itraq experiments. Assume that the true protein abundance for subject k in experiment r is a rk = µ + r + δ rk. r is not estimable from the proteomic data, so augmenting estimates of δ r across experiments as a surrogate for absolute abundance fails to take the variance component r into account. Randomization of subjects to experiments helps to avoid systematic biases, but does not eliminate the random mean shift r expected across experiments. Estimable parameters The r can not be estimated and eliminated using the proteomic data alone. Not an issue in (well designed) case-control studies. r can also be accounted for in other types of association studies. For example, assume that in truth we have E[N rk ]=β 0 + β 1 a rk. Substituting µ + r + δ rk for a rk, we get E[N rk ] = β 0 + β 1 (µ + r + δ rk ) = {β 0 + β 1 µ} + β 1 r + β 1 δ rk = γ 0 + B r + β 1 δ rk

13 Linear mixed model fit Linear mixed model fit by REML Formula: y x + (1 id) Data: dat AIC BIC loglik deviance REMLdev Random effects: Groups Name Variance Std.Dev. id (Intercept) Residual Number of obs: 27, groups: id, 9 Fixed effects: Estimate Std. Error t value (Intercept) x Association

14 ROC ELISA comparison

15 More results Vitamin A

16 Vitamin E Estimable parameters In case there is more than one protein with log 2 abundances linearly related to the nutrient concentration: E[N rk ] = β 0 + j = β 0 + j β j ( µ j + jr + δ jrk ) β j µ j + j β j jr + β j δ jrk = γ 0 + B r + j β j δ jrk Thus, even though we have multiple proteins, and therefore multiple random effects for between experiment differences, the resulting linear mixed effects model still only has one random effect that jointly summarizes the between experiment differences.

17 Multivariate models joint marginal β p β p retinol-binding protein 0.8 5e e-19 complement factor H isoform a e e-0 insulin-like growth factor-binding protein e e-0 complement C1r subcomponent -1 2e e-0 Multivariate models

18 Estimation comparison Concordance correlation coefficient to measure the agreement in relative abundance estimates between two methods: ρ(x, Y )= 2 cov(x, Y ) σ 2 X + σ2 Y +(µ X µ Y ) 2. 0% 25% 50% 75% 100% LM vs LME LM vs LM S LM vs LME S LME vs LM S LME vs LME S LM S vs LME S MSE / /5 6 2! Relative protein abundance estimate " 2 5/ 1 /5 2 5/ 1 /5 2 5/ Linear mixed model (joint estimation) Linear mixed model (separate estimation) Linear model (separate estimation)! Root median squared fold change " 1 8 /

19 Estimation method comparison 1. Linear mixed effects models 2. Linear model. Masterpool normalization (mean). Masterpool normalization (median) 5. Mean sweep (ignoring peptides) 6. Median sweep (ignoring peptides) 7. Mean sweep (using peptides) 8. Median sweep (using peptides) Estimation method comparison 1.0 Root mean squared fold change MP:1 MP:2 MP: MP: 1.12 Root median squared fold change MP:1 MP:2 MP: MP:

20 Estimation method comparison R gi gi gi gi gi gi gi gi gi gi gi Log reporter ion intensities (median polish) gi gi gi gi Run with 8 technical replicates of a master pool

21 http: //biostat.jhsph.edu/ iruczins/