Periodic Reporting for period 1 - SingleCellAI (Deep-learning models of CRISPR-engineered cells define a rulebook of cellular transdifferentiation)
Berichtszeitraum: 2019-07-01 bis 2021-06-30
Over-expression of just a few transcription factors (TFs) can transform a fully differentiated cell into a state of induced pluripotency (iPS). Even more astoundingly, a similar strategy can perform cellular transdifferentiation, i.e. direct conversion of a cell from one type to another. However, the selection of TFs to overexpress or repress to induce transdifferentiation is typically found unsystematically through a combination of biological intuition and extensive trial-and-error required for each combination of source and target cell types. This approach does not scale to transdifferentiation of any cell type into any other cell type, which would be desirable not only for regenerative medicine but also to answer fundamental questions about the biology of cellular identity.
A premise for an unbiased, quantitative approach to designing TF transdifferentiation cocktails is the knowledge of the impact of each of the TFs on gene expression with a single-cell resolution. Generating such a resource is now possible thanks to the advances in single-cell sequencing and pooled CRISPR gene editing, which enable coupling gene editing with transcriptome readout. At the same time, the analysis and data-driven prediction for single-cell transcriptomic datasets is currently being revolutionized by deep learning. This is due to the ability to scale the learning process to very large datasets and to extract informative data representations directly from the data.
Considering the above, we formulated the following objectives: Objective 1: Development of deep-learning models for CRISPR-engineered cells; Objective 2: Computational modeling of transdifferentiation using the deep-learning models; Objective 3: Experimental validation of fibroblast transdifferentiation into hematopoietic cells.
A premise for an unbiased, quantitative approach to designing TF transdifferentiation cocktails is the knowledge of the impact of each of the TFs on gene expression with a single-cell resolution. Generating such a resource is now possible thanks to the advances in single-cell sequencing and pooled CRISPR gene editing, which enable coupling gene editing with transcriptome readout. At the same time, the analysis and data-driven prediction for single-cell transcriptomic datasets is currently being revolutionized by deep learning. This is due to the ability to scale the learning process to very large datasets and to extract informative data representations directly from the data.
Considering the above, we formulated the following objectives: Objective 1: Development of deep-learning models for CRISPR-engineered cells; Objective 2: Computational modeling of transdifferentiation using the deep-learning models; Objective 3: Experimental validation of fibroblast transdifferentiation into hematopoietic cells.
Towards Objective 1, a pipeline for pre-processing, quality control, and analysis of single-cell CRISPR screens was developed; and a package for automatic machine learning that can efficiently operate in a distributed computing environment to select most reproducible model of many candidates was implemented. Towards Objective 2, an approach for biologically-informed deep learning was implemented. Several gradient-based interpretability approaches were implemented and their usefulness evaluated. This interpretable approach will be used to identify the most promising candidates for TF transdifferentiation cocktails, which will be experimentally validated in Objective 3.
Spanning the duration of the project covered by the final report, we developed a framework for reproducible and interpretable deep learning that can be used to analyze single-cell CRISPR screens. We are determined to continue this project beyond the reporting period by exploiting the interpretable nature of the developed model to identify likely TF transdifferentiation cocktails and perform an experimental validation. After successful completion, this approach could allow for novel strategies for cell transdifferentiation and regenerative medicine.