Living cells are fundamental building blocks of Life, playing both structural and functional roles in all types of organisms. First described in the XVIIth century by Robert Hooke and despite the myriad of breakthroughs that were achieved in Cell Biology since then, many aspects of their biology are still unknown today. Eukaryotic cells store their genetic information within DNA molecules enclosed within their nucleus, that is transcribed into messenger RNA molecules which serve as blueprints for synthesizing proteins, which is a diverse family of molecules responsible for various functional and structural roles within cells. Technological advances that took place during the last decade such as next generation sequencing (NGS) and single-cell assays opened the door to incredibly rich datasets able to quantitatively describe cell populations with extreme precision: in one experiment, one can today approximate the gene expression of tens of thousands of cells over tens of thousands of genes. In parallel, the machine learning field also witnessed a surge of new and revisited approaches (deep neural networks theory, optimal transport, kernels...), made possible by mathematical and hardware developments. One of the main goals of computational biology today is to link these two fields by applying machine learning approaches to complex biological datasets in order to answer challenging biological questions.One key question referred to as data integration is to conceive algorithms able to yield a joint representation of several datasets coming from different sources or measured along different biological modalities, so that similar cells end up close to one another independently from their dataset of origin. This problem is highly challenging in the general case, and solving it has very sought-after applications such as creating comprehensive cell atlases for a disease by aggregating data from many patients, or inferring models including factors from different biological modalities. Many approaches have been proposed to tackle data integration over the last ten years, so much that despite regular benchmark studies it is puzzling to know what to use for a given application. To tackle this issue we developed a new data integration framework named transmorph, that provides many machine learning algorithms as building blocks that can be assembled into complex data integration pipelines. We show that transmorph can be used to build data integration pipelines that work on par with state-of-the-art approaches, while also proving to be useful to determine which algorithmic subunit is more adapted to a given situation. Transmorph is today distributed as an open-source python framework, and embarks an ecosystem of benchmarking datasets, quality assessment metrics, plotting tools as well as a comprehensive user API to build end-to-end data integration models.Another hope for these highly resolute single-cell assays is that they could improve our understanding of cancer, as tumors are highly heterogeneous cell formations embedded in a complex microenvironment. In particular, factor analysis approaches can be applied to discover multidimensional signals in the gene expression space that can then be enriched using databases, and related to interpretable biological processes such as cell proliferation, metabolic activity or metastasis. Using Ewing sarcoma inducible cell lines where the oncogene presence can be precisely monitored, we highlighted a dozen of such processes and studied their dependence on the oncogene activity. We paid a particular attention to proliferation-related signals, and we could highlight in many datasets a multidimensional trajectory taking place in the gene expression space corresponding to the cell cycle process. We were able to derive from these observations a segment-wise cell cycle model, able to approximate the state of individual cells within the cycle as well as other features such as cell doubling time.