Our seminar series is free and available for anyone to attend. Unless otherwise stated, seminars take place on Wednesday afternoons at 2pm in the Kilburn Building during teaching season.

If you wish to propose a seminar speaker please contact Antoniu Pop.


Applications and possibilities for signal processing, machine learning and computational statistics in clinical proteomics and metabolomics

  • Speaker:   Dr.  Andrew Dowsey  (Inst.of Human Development, UoM)
  • Host:   Andrew Brass
  • 19th November 2014 at 14:00 in KB L.T. 1.4
Liquid Chromatography interfaced to Mass Spectrometry (LC-MS) has emerged as the central technique for discovering patterns of protein or metabolite changes between groups of biological or clinical samples in systems biology and translational medicine. Various statistical models have been introduced that facilitate the analysis of large-scale experiments including appropriate treatment of often substantial sources of variation e.g. batch effects, technical vs biological variation, combining multiple peptide measurements per protein. These models are dependent on sensitive and robust feature extraction from a preceding raw data processing pipeline. However, current pipelines are ad-hoc, involving a set of self-contained error-propagating steps that reduce the data down to a symbolic feature-based representation at the earliest stage. In this talk we will overview some unique properties of raw mass spectrometry signals, highlight the challenges faced by processing this big data (up to 40Gb per sample) and discuss the potential role of signal processing, machine learning and computational statistics in the field. We will then describe our progress with these approaches in developing a statistically rigorous raw data analysis workflow (Liao et al, Proc IEEE ISBI, 2014). In this workflow, the LC-MS output for each sample is first transformed into an image; these are then registered to account for chromatography inconsistency; and finally, a functional mixed-effects model is applied. We will also describe progress towards complementary domain-specialised modelling: Using non-negative matrix factorisation with a minimum volume constraint, we have learnt a low-dimensional representation that approximates all possible peptide signals. This dictionary is incorporated into an L1-regularised sparse regression approach with Poisson noise model, resulting in a unique approach that separates and accurately quantifies overlapping peptide signals.

Bio Andrew is currently Lecturer in Medical Bioinformatics at the Institute of Human Development. In January he will be taking up the post of Reader in the School of EEE and Computer Science, University of Liverpool. As Bioinformatics Research Lead for the Centre for Advanced Discovery and Experimental Therapeutics (CADET), he spearheads the application of computational approaches to CADET?s basis bioscience, drug discovery and drug development programmes. Andrew is currently an MRC New Investigator and Associate Editor for the IEEE Journal of Biomedical and Health Informatics. During PhD and postdoctoral training at the Royal Society/Wolfson Medical Image Computing Laboratory, Imperial College London, he was awarded the MICCAI Young Investigator Award and held an EPSRC Overseas Postdoctoral Fellowship at the Life Sciences Interface.
▲ Up to the top