Advanced analytical approaches for single cell genomics research

National Library Of Medicine

Investigators

Abstract

Cells are the fundamental unit of life. Different cells play different functional and physiological roles across tissues and organs in the human body and other multi-cellular organisms. Single cell genomics technologies are revolutionizing our understanding of the diversity of cell types and their selective genetic profiles. Traditionally, cell types are classified by their morphological and histological properties, resulted in limited resolution of cell types. Advances in single cell/nucleus RNA sequencing (sc/snRNA-seq) and spatial transcriptomics have enabled unbiased transcriptomic profiling of individual cells in tissue samples, revealing unprecedented number of distinct cell types as well as their unique 3D spatial organization. The advancement in single cell technologies also poses great computational challenges to the research community. The high-throughput, large-scale datasets exhibit highly complex data distributions and structures, calling for advanced data science approaches and new analytical methods for data processing and statistically-rigorous analysis. Artificial intelligence and machine learning have emerged as powerful tools for single cell data analysis. However, off-the-shelf machine learning techniques do not address the unique challenges from biological data, leading to sub-optimal performance and lack of interpretability. This project has a long-term goal of advancing analytical approaches for the analysis of single cell genomics data using explainable artificial intelligence and advanced biostatistics methods. The long-term goal is achieved through three Specific Aims: • Aim 1: Developing novel computational methods and refinements that show improved performance and superior biological interpretability for identifying cell type-specific biomarkers and matching transcriptomically-derived cell types from sn/scRNA-seq data. • Aim 2: Designing and implementing computational strategies for collaborative studies to investigate diseases and conditions using single cell and spatial transcriptomics methodologies. • Aim 3: Contributing to the broad data ecosystem through active community engagement. Our group have developed a random forest machine learning-based algorithm, NS-Forest, for identifying the necessary and sufficient marker genes for characterizing the data-driven cell types from sc/snRNA-seq data. We have designed a novel metric, the Binary Expression Score, to quantify the desired binary expression pattern of a gene that has high expression in the target cluster and little-to-no expression in other clusters. The Binary Score and the Gini Impurity outputted from the scikit-learn random forest classifier are used to combinatorially rank the top marker genes. The minimal marker gene combination is determined by the highest F-beta score in the decision tree evaluation module for cell type classification performance. This year we focused on optimizing the binary gene selection module in the NS-Forest algorithm. We released NS-Forest v4.0 that implements a BinaryFirst strategy to pre-select gene candidates as input features to the random forest step, which largely increases the efficiency of the random forest search for the top marker genes by reducing the input feature dimensionality and avoiding local optima in the random forest algorithm. We applied NS-Forest v4.0 in the brain, kidney, and lung sc/snRNA-seq datasets, which showed superior classification performance in the F-beta score, precision, and recall metrics when benchmarked with alternative marker gene identification approaches based on differential expression analysis and manual curation. We also introduced a new metric, the On-Target Fraction, to capture the improved robustness of the gene expression pattern in the marker gene expression heatmap. Subsequently, we released NS-Forest v4.1 with enhanced modularization and reproducibility features in the Python package. As an active member in the single cell community, we have established long-term collaborations with investigators from the NIH BRAIN Initiative, the NIH Human BioMolecular Atlas Program (HuBMAP), the NIH LungMAP, and the CZ CELLxGENE consortia, which have led to fruitful outcomes in publications, workshops, and hackathons.