Generating Labeled Low-Heterogeneity Transcriptomes Using CRISPRa and CRISPRi Can Improve Phenotype Prediction by Deep Learning

The development of deep learning (DL) methods in biology holds great promise for fundamental knowledge, biomedicine, and agriculture. The most global challenge in this area is the development of models that use omics data to predict the complex phenotypic traits, in particular cancer resistance to radiotherapy and chemotherapy. Due to the insufficient volume and high heterogeneity, the available datasets are not suitable enough for DL-based phenotype prediction without dimensionality reduction. This paper briefly discusses the major deficiencies of existing omics datasets for phenotype prediction and proposes a framework to create the phenotypically labeled low heterogeneity datasets to supplement the existing ones. The authors formulated specific experimental considerations that are critical for obtaining phenotypically labeled cell lines with small incremental differences in transcriptomes using CRISPRa and CRISPRi technologies. These cell lines and corresponding datasets will serve as the uniform genetic information sources for gain/loss of function in DL-based phenotype prediction models.