Ming Li, Ph.D. Research Assistant Professor Department of Biostatistics VICC Biostatistics Shared Resources

Contact Information

• E-mail: ming.li@vanderbilt.edu
• Phone: (615) 936-1378
• Fax: (615) 936-2602
• Address: 580 Preston Research Building, Nashville, TN 37232-6848
• Curriculum Vitae

Research Interests

• Bayesian nonparametric and semiparametric methods
• Mass spectrometry data preprocessing method
• Genomic and proteomic data analysis
• Bayesian statistical designs for clinical trials

Projects

Mass spectrometery data preprocessing

Mass Spectrometry (MS) technology makes it possible to study various biological samples at their protein level. The data generated by this technology holds invaluable information leading to the disease diagnosis and treatment (Paweletz et al., 2000; Adam et al., 2003; Schaub et al., 2004). However, there are tremendous challenges when dealing with MS data since the raw mass spectrometric data reflects not only the protein information but also "junk" information (often referred as noise). The noise is due to multi-sources variations, for instances, there are substantial variations in the signal intensity even for those replicate spectra. The small amount of shifts at the mass/charge locations for the peaks representing the same protein indicate another source of variations. In addition, there exists quite different decreasing baseline for each individual spectrum. MS data preprocessing aims to take control of these variations and obtain the true signal information for further statistical analysis. To approach the same goal of taking care of the variations but from different perspectives, the existing preprocessing methods could be summarized into two major ones: functional data analysis approach (Morris and Carroll 2004, Billheimer 2004) and the wavelet-based feature extraction approach (Baggerly et al., 2003, Qu, et al. 2003, Chen, Hong and Shyr 2004; Coombes et al. 2004, Jeffrey et al. 2004). The focus of my current research is on the feature extraction approach.

By feature extraction approach, it assumes that all the useful information from the MS data intensity plot are in the peaks and those identified peaks ideally correspond to individual protein. The major goal of this approach is to identify, quantify and match the peaks across spectra. Therefore, we might view the feature extraction approach as the peak detection process, which takes general steps like following: (1)data calibration; (2) denoising (smoothing); (3) baseline correction; (4) normalization; and (5) peak detection and alignments (binning). Each step is crucial for providing an accurate peak list containing important features of the MS data for further statistical analysis. Current reseach focus is to improve the existing clibration methods as well as the peak alignment methods.

Bayesian design for phase II clinical trials

The nature of the research goal on phase II clinical trial makes Bayesian methodology appealing since it seems to fit the setting of decision theory naturally. However, the requirement for the utility functions and relatively complicate computation procedures may discourage many from adopting this approach. Recently, a very user friendly Bayesian two-stage design, the STD (single threshold design) has been proposed by Tan and Machin (2002). Based on their design framework, we proposed one modification to reduce the sample size for stage II given the very promising outcome observed from stage I. To keep it consistent to STD and as user friendly as possible, this design is a slight extension of STD yet keep its original structure to the greatest extent.