A.-Large-scale probabilistic models for bioimaging research



Biomiaging scales

[Fig. 1 . Figure from article: Subramaniam, S. "Bridging the imaging gap: visualizing subcellular architecture with electron tomography". Curr Opin Microbiol, 8(3):316–322, November 2006.]

Life expands all ranges of shapes, sizes and locations and it is important to study its components at different scales, from molecules to organs, in order to be able to explain complex biological structures, functions and behaviors. In most of these scales, different imaging technologies play a vital role to understand different phenomena (Fig. 1). As the topic says, "an image is worth a thousand words", and life science is no exception. Moreover, through improved technology methods, the capabilities of different imaging modalities and the amount of acquired data is increasing very rapidly.


One of the current challenges is to develop new algorithms to efficiently transform all this data into knowledge. Computational approaches can achieve this goal in multiple ways, such as, increasing data throughput to study larger population samples and increase the likelihood of new discoveries, or uncovering patterns that otherwise would be to extract by hand.  My research interest and approach to these challenges is based on large-scale probabilistic models. Probabilistic models have many advantages: first, they are very flexible and can model very complex phenomena like the ones present in life sciences. Moreover, in biological analyses, it is common to have a rich set of prior information from different sources that we should incorporate into the model. Second, since the output are probability distributions, we have uncertainty estimations of our results, which can be interpreted as confidence intervals or guide semi-automatic in order to improve and accelerate conclusions through a tight feedback loop between scientists and computational tools. Finally, we can benefit from a great deal of research that has been conducted in the field of probabilistic models from different fields such as statistics, computer science and engineering, mathematics or physics.


Unfortunately, the great flexibility of probabilistic models is a double-edged sword and it is still a challenge (and an art) to extend inference techniques to the models and the amount of data present nowadays in many bioimaging problems. My main research goal is to design new algorithms to process all this data efficiently and with minimum human interaction, in order to transform images into specific knowledge and to expand the set of scientific questions that can be answered. In particular, two main directions to achieve this goal is to exploit the modularity of highly structured biological systems and to find the right features that can be extracted from different imaging modalities and scales.