1. Hierarchical Processing for LC/MS Metabolomics Data Generated in Multiple Batches
Douglas Walker 1, Karan Uppal 2, Dean Jones 2, Tianwei Yu 3,*
1 Department of Environmental Health,
2 Department of Medicine,
3 Department of Biostatistics and Bioinformatics,
Emory University
12/8/2018

2. Take “slices” in retention time, send to MS
Introduction
retention time
Take “slices” in retention time, send to MS
Liquid chromotography
retention time
Mass-to-charge ratio (m/z)
Mass-to-charge ratio (m/z)

3. Issues in LC/MS data processing:
Introduction
An ideal peak should show a Gaussian curve in intensity along the retention time axis, while keeping constant in the m/z axis
Issues in LC/MS data processing:
Background noise
Chemical noise (ridges in spectra, ion suppression,…)
Peak intensity (along retention time axis) deviate from Gaussian curve substantially (fronting, breaks, …)
Slight m/z shift across MS spectra (not severe with newer machines)
Retention time shift across LC/MS spectra
Multiple charge status of a single metabolite
Isotopes ……

4. Introduction
The analysis of LC/MS metabolomics data involves a number of steps.
Example - the XCMS workflow: 
XCMS
Mzmine
apLCMS
MetAlign
… …
These steps assume the peaks are from a similar distribution, in terms of their variation in m/z and retention time.

5. Introduction
Batches can be different from each other in terms of data characteristics
Retention time shift
Levels of variation in m/z and retention time
Different features may be missing at different rates ……
One set of parameters may not suite all batches. It can make tolerance levels unnecessarily large, and cause false +/- in alignments.
An example:
Batch 1 RT
Batch 2
Two tight clusters might make the program think there are two features: 

6. Peak detection, Time adjustment, Peak alignment, Signal recovery
Introduction
Batch 1
Batch 2
Batch 3
Common practice:
Peak detection, Time adjustment, Peak alignment, Signal recovery

7. Methods
Batch 1
Batch 2
Batch 3
Peak detection
Time adjustment
Peak alignment
Signal recovery
Peak detection
Time adjustment
Peak alignment
Signal recovery
Peak detection
Time adjustment
Peak alignment
Signal recovery

8. + Methods Treat each as a sample Time adjustment Peak alignment
Between batch
Within batch
Overall correction
Time adjustment
Peak alignment
Signal recovery

9. Results
The data:
Standard QC sample
Same sample measured repeatedly
33 batches, each contains 3 profiles from the same sample
Parameter settings:
All other parameters the same.
Batch-wise procedure:
Within-batch detection proportion threshold: 0.4, 0.5, 0.6, 0.7, 0.8
Between-batch detection proportion threshold: 0.1, 0.2, …, 0.8
Traditional procedure:
Detection threshold (number of profiles): 5, 10, 15, …., 95

10. Results
Evaluation of the results:
Total number of zeros in the final data matrix
Proportion of features with m/z matched to known metabolites (xMSAnnotator)
Coefficient of variation (CV) in the final data matrix (without considering batches)
Coefficient of variation (CV) after merging each triplet (batch)

11. Results

12. Results
The data:
Emory/Georgia Tech Center for Health Discovery and Well Being (CHDWB) data
Part of the data was used for the testing:
115 subjects
6 batches, each contains 60 profiles
Each subject measured in triplet; a few were measured 6 times
Parameter settings:
All other parameters the same.
Batchwise procedure:
Within-batch detection proportion threshold: 0.4, 0.5, 0.6, 0.7, 0.8, 0.9
Between-batch detection proportion threshold: 0.25
Traditional procedure:
Detection threshold (number of profiles): 30,60,90,120,180

13. Results
Evaluation of the results:
Before merging triplets for each subject:
Total number of zeros in the final data matrix
Proportion of features with m/z matched to known metabolites (xMSAnnotator)
Average coefficient of variation across batches
After merging triplets for each subject:
Coefficient of variation (CV) in the final data matrix (without considering batches)
Coefficient of variation (CV) in the final data matrix using only non-zero values (without considering batches)

14. Results
Red: Batchwise processing; Blue: traditional processing

15. Better performance appears to be achieved when parameters require
Conclusions
The stepwise processing procedure that consider batches improves the consistency of feature detection and quantification.
Better performance appears to be achieved when parameters require
- Higher within-batch consistency
- Lower between-batch consistency
Extra studies is necessary to validate and optimize the procedure
- larger datasets
- examination at the single feature level
Needs better integration with semi-supervised peak detection procedure

16. People who work on metabolomics data in Dr. Jones lab:
Acknowledgements
People who work on metabolomics data in Dr. Jones lab:
Dr. Shuzhao Li Dr. ViLinh Tran
Dr. Youngja Park Mr. Chunyu Ma
Biomedical collaborators who use metabolomics data:
Dr. Jessica Alvarez Dr. Jeremy Sarnat
Dr. Chandresh Ladva Dr. Donghai Liang
Biostatisticians:
Dr. Jian Kang Dr. Qingpo Cai
Dr. Nancy Jin Dr. Eugene Huang
Dr. Elizabeth Chong Dr. John Hanfelt