Saishi Cui, Department of Epidemiology and Biostatistics, Dornsife School of Public Health, Drexel University, Philadelphia, PA, 19104 Issa Zakeri, Department of Epidemiology and Biostatistics, Dornsife School of Public Health, Drexel University, Philadelphia, PA, 19104 Sina Nassiri, F.Hoffmann-La Roche Ltd., 4070 Basel, Switzerland
Poster # 7
Single-cell RNA sequencing (scRNA-seq) brings both great opportunities and challenges for transcriptomic analysis. While scRNA-seq enables the characterization of cell heterogeneity at an unprecedented resolution, analytical issues like sparsity, noise and bias can severely compromise interpretation if not addressed properly. To extract meaningful biological signals, effective feature selection is critical. We propose Mcadet, a novel framework leveraging Multiple Correspondence Analysis (MCA) and community detection for feature selection in scRNA-seq data. Mcadet aims to accurately identify informative genes from fine-resolution and rare cell datasets where existing methods falter. The strength of Mcadet lies in its integrated five-step approach: 1) Matrix pre-processing; 2) MCA; 3) Community detection for cell clustering; 4) Computation and ranking of Euclidean distances; 5) Novel statistical testing. Specifically, MCA isolates intrinsic biological variation, while community detection enables robust cell clustering by tuning resolution parameters. Thus, Mcadet can effectively analyze datasets with high cell resolution as well as those containing rare cell populations. This flexibility enables accurate detection of fine-resolution patterns and rare cells within challenging scRNA-seq data, enhancing Mcadet's analytical capabilities. Through comparative analysis on diverse simulated and real PBMC datasets, we demonstrate Mcadet's superior performance in selecting highly informative genes. Our evaluations involved established methods like Seurat and employed unbiased metrics like Jaccard similarity index and ARI, assessing feature selection accuracy and clustering outcomes. Results show Mcadet outperforms other approaches, particularly for fine-resolution and rare cell data, by effectively isolating intrinsic variation and allowing multi-resolution community detection. Mcadet exhibits balanced gene selection without biases and recovers both high and low expression genes. The integration of correspondence analysis, graph clustering, and tailored statistical testing underpins Mcadet's capability to address pervasive challenges in scRNA-seq analysis.
Kommentarer