annalisabarla

The analysis of biological data is characterized by a large amount of features representing each given example (high-dimensionality) and a relatively small number of samples. Traditional statistical tools were developed and studied to adapt in the opposite scenario, where the samples outnumber many times the variables, so different approaches have to be explored. In particular, there is a broad class of methods for feature selection that has been at the centre of researchers attention in the recent years, since this step is necessary to select the most informative variables at hand.

My current interest is on sparsity-enforcing regularization methods, exploiting different penalties to achieve sparse models depending only on few and relevant variables, with the main goal of modelling complex biological systems. Such methods should possibly allow for the integration of different data types and be flexible enough to incorporate all the available biological prior knowledge.

• My PhD project focused on kernel methods for images. In this context, I designed novel and effective kernel functions tailored on the specific representation of the data at hand that allowed for a content-based classification of images. I also proved their effectiveness on different real dataset, publishing the results of my research in journals [Odone et al. IEEE Trans. on Image Processing, 2005] and conferences (ECCV 2002; ICPR 2002; ICIAP 2003; ICIP2002).
• The time at Dr. C. Furlanello's lab in Trento represented the possibility to apply my expertise in machine learning to a very interesting and yet tricky field: the analysis of molecular high-throughput data. I learned how to deal with DNA microarray data, I participated to the design of selection-bias aware frameworks [G. Jurman et al. Bioinformatics, 2008] and I started a new research line on mass-spectrometry proteomic data [Cannataro et al. IEEE Trans. on Nanobiosc. 2007; Barla et al. Brief. In Bioinf. 2008].
• With the Health-e-Child project, I had the opportunity to form and lead a unit of four researchers within Prof. A. Verri's group, with the aim of studying machine learning methods for computational biology. Together with my coworkers, I carried on successful collaborations with biologists and MDs, gaining enough interdisciplinary knowledge to understand their fundamental biological questions and to cast such issues into a valid and robust statistical framework. Our results have been published on relevant journals in the field [Fardin et al. BMC Bioinformatics 2009; Fardin et al. . J. of Biomed. and Biotech. 2010a; Fardin et al. 2010b Molecular Cancer]. Moreover, in the Pediatric Glioma study (Health-e-Child project) our results have also undergone a biological validation phase that confirmed their validity.