Iterative programming of blood cells (ML2Cell)

Our body is made up of a plethora of cells with different characteristics, shapes, and functions. How these diverse cell types develop (or differentiate) from a single founder cell (a zygote) is the focus of ongoing developmental and molecular biology research. Regenerative medicine aims to actively direct the differentiation of stem cells to replace damaged tissues, for example, to generate skin for burn victims or platelets for patients under chemotherapy. Moreover, it may be desirable to change the identity of already differentiated cells, for example, to reprogram cancerous cells toward less malignant cell states. However, finding the right cocktails and sequences of molecules to achieve a specific differentiation outcome is a challenge. There are millions of possible combinations and often the success of the differentiation protocol can only be evaluated fully at the end of the process. To make it possible to refine protocols in real-time during ongoing differentiation experiments, we devised a combined experimental/computational approach (called ML2Cell) that borrows algorithmic principles from machine learning (ML) and integrates them directly in the design of biological experiments. The one key challenge to solve in this prototype project will be to implement an assessment regimen that can inform decision making on-the-fly between different steps of the protocol (that is, at most 24 hours). To this end, we will pair rapid genomics assays with hyper-parallelized bioinformatics analysis. We will benchmark ML2Cell by generating two blood cell types from undifferentiated blood progenitors (hematopoietic stem cells): red blood cells and B cells. These are two highly relevant proof-of-principle examples and there is urgent demand for methods to replace many other types of tissues (apart from blood, especially for skin, cartilage, and bone, but also for internal organs, e.g., liver). If successful, future applications of our approach may also include personalizing the engineering of immunotherapies. On a more abstract level, ML2Cell serves as a proof-of-concept for implementing methods from computer science in biological experiments. In a way, this turns around a long -running trend in which computer science algorithms take inspiration from biology or physics (easily visible in the names of popular algorithms, e.g., neural networks, simulated annealing, genetic algorithms, ant colony optimization). We envision that other concepts and approaches from computer science may find use in experimental study design, for instance, for search and sorting.

Iterative programming of blood cells (ML2Cell) Our body is made up of a plethora of cells with different characteristics, shapes, and functions. How these diverse cell types develop (or "differentiate") from a single founder cell (a "zygote") is the focus of ongoing developmental and molecular biology research. Regenerative medicine aims to actively direct the differentiation of stem cells to replace damaged tissues, for example, to generate skin for burn victims or platelets for patients under chemotherapy. Moreover, it may be desirable to change the identity of already differentiated cells, for example, to reprogram cancerous cells toward less malignant cell states. However, finding the right cocktails and sequences of molecules to achieve a specific differentiation outcome is a challenge. There are millions of possible combinations and often the "success" of the differentiation protocol can only be evaluated fully at the end of the process. To make it possible to refine laboratory protocols for differentiation experiments, we devised a combined experimental/computational approach (called "ML2Cell") that borrows algorithmic principles from machine learning (ML) and integrates them directly in the design of biological experiments. To this end, we paired genomics assays with deep bioinformatics analysis to describe the molecular changes that occurred in cells during differentiation. We benchmarked ML2Cell by generating red blood cells (erythrocytes) from blood stem cells. There is an strong demand for these cells for transfusions in emergency medicine and other care units. We compared many combinations of reagents that are being used in red blood cell engineering and assessed which combinations were most effective in doing so. We then identified molecular features of the cells that differed from real erythrocyte progenitors. These differences needed to be minimized by our laboratory-generated blood cells to improve their functionality. Now we could predict new reagents that could be used to specifically target these features to release the remaining roadblocks that prevented the cells from reaching their destination. This opens the way for improved differentiation protocols. In the future, the same approach may be used for many other types of tissues, for instance, for skin, cartilage, and bone. On a more abstract level, ML2Cell serves as a proof-of-concept for implementing methods from computer science in biological experiments. In a way, this turns around a long-running trend in which computer science algorithms take inspiration from biology or physics (easily visible in the names of popular algorithms, e.g., "neural networks", "simulated annealing", "genetic algorithms", "ant colony optimization"). We envision that other concepts and approaches from computer science may find use in experimental design, for instance, for search and sorting.