In Silico and Single-Cell-based Approaches for Discovering New Reprogramming Cocktails

Sammanfattning

Reprogramming of human cells into induced pluripotent stem cells (iPSCs) is traditionally done through a transcription factor cocktail known as OKSM. Concerns have however been raised about the efficiency of the method. Hence finding new more efficient reprograming cocktails is of great importance. This is experimentally done through perturbation screens, where many gene alterations or chemical compounds are tested. However large-scale perturbation screens are time consuming. Therefore, the recent advancements in the field of in silico gene perturbation (ISGP) show great promise. The main idea is to create models based on large sets of tested perturbations, where the outcome is known. The model should then leverage the learned knowledge to make inferences of the outcome of yet untested perturbations. In this project the aim is to create a method for predicting whether a yet untested perturbation has the potential to induce reprograming, to find candidates for experimental testing. This is done in two parts. The first part involves training a Biolord gene perturbation prediction models (GPPMs) on the Replogle et al. 2022 genome-wide essential gene perturbation screens in cell lines RPE1 and K562. The final model should be able to predict the cells resulting gene expression of a gene perturbation. The second part is to make a logistic regression model on the Liu et al. 2020 naïve reprograming time-course. The model is trained to differentiate reprogramed naïve cells the source cell type (fibroblasts). The models should then be combined to predict the resulting cell state of cells after an untested perturbation. The GPPMs created in the project are all found to result in high mean squared errors in the predicted value compared to the true values. The predictions all have high Pearson correlations coefficient, suggesting that the predictions are still corelated with the true value. However more parameter tuning is needed to make a model accurate enough for real life applications. The reprograming models created perform well on the test sets. However, the data used for training was not diverse enough to allow for parameter tuning. In absence of large perturbation screens in fibroblast cell line, the reprograming model was applied to the predicted gene expressions of the Replogle et al. RPE1 dataset. It was however found that since the screen was made in an immortalized cell line, that was found to already be very similar to the naïve iPSC, result gave limited information of the perturbations effect on reprograming.