Statistical methods development for the multiomic systems biology

Resumen

Systems Biology research has expanded over the last years together with the development of omic technologies. The combination and simultaneous analysis of different kind of omic data allows the study of the connections and relationships between different cellular layers. Indeed, multiomic integration strategies provides a key source of knowledge about the cell as a system. The present Ph.D. thesis aims to study, develop and apply multiomic integration approaches to the field of systems biology. The still high cost of omics technologies makes it difficult for most laboratories to afford a complete multiomic study. However, the wide availability of omic data in public repositories allows the use of these already generated data. Unfortunately, the combination of omic data from different sources provokes the appearance of unwanted noise in data, known as batch effect. Batch effect impairs the correct integrative analysis of the data. Therefore, the use of so-called Batch Effect Correction Algorithms is necessary. As of today, there is a large number of such algorithms based on different statistical models and methods that correct batch effect and are part of the data pre-processing steps. However, the existing methods are not intended for multi-omics designs as they only allow the correction of the same type of omic data that must be measured across all batches. For this reason, we developed MultiBaC algorithm, which removes batch effect in multiomic designs, allowing the correction of data that are not measured across all batches. MultiBaC is based on PLS regression and ANOVA-SCA models and was validated and evaluated on different datasets. We also present MultiBaC as an R package to facilitate the use of this tool. Most existing multiomic integration approaches are multivariate methods based on latent space analysis. These methods are known as data-driven as they are based on the search for correlations to determine the relationships between the different variables. Data-driven methods require a large number of observations or samples to find robust and/or significant correlations among features. Unfortunately, in the molecular biology field, data sets with a large number of samples are not very common, again due to the high cost of generating omic data. As an alternative to data-driven methods, some multiomic integration strategies are based on model-driven approaches. These methods can be fitted with a smaller number of observations and are very useful for finding mechanistic relationships between different cellular components. However, model-driven methods require a priori information, which is usually a metabolic model of the organism under study. Currently, only transcriptomics and quantitative metabolomics have been successfully integrated using model-driven methods. Nonetheless, quantitative metabolomics is not very widespread and most laboratories generate non-quantitative or semi-quantitative metabolomics, which cannot be integrated with current methods. To address this issue, we developed MAMBA, a model-driven multiomic integration method that relies on mathematical optimization problems and is able to jointly analyze non-quantitative or semi-quantitative metabolomics with other types of gene-centric omic data, such as transcriptomics. MAMBA was compared to other existing methods in terms of metabolite prediction accuracy and was applied to a multiomic dataset generated within the PROMETEO project, in which this thesis is framed. MAMBA proved to capture the known biology of our experimental design and was useful for deriving new findings and biological hypotheses. Altogether, this thesis presents useful tools for the field of systems biology, covering both the pre-processing of multiomic datasets and their subsequent statistical integrative analysis.