A machine learning approach identifies cellular senescence on transcriptome data of human cells in vitro

Although cellular senescence has been recognized as a hallmark of aging, it is challenging to detect senescence cells (SnCs) due to their high level of heterogeneity at the molecular level. Machine learning (ML) is likely an ideal approach to address this challenge because of its ability to recognize complex patterns that cannot be characterized by one or a few features, from high-dimensional data. To test this, we evaluated the performance of four ML algorithms including support vector machines (SVM), random forest (RF), decision tree (DT), and Soft Independent Modelling of Class Analogy (SIMCA), in distinguishing SnCs from controls based on bulk RNA sequencing data. The dataset includes 162 in vitro samples, covering three human cell types: fibroblasts, melanocytes, and keratinocytes, and three senescence inducers: irradiation, bleomycin treatment, and replication. Under tenfold and leave-one-out cross-validation, as well as independent dataset validation, all methods provided ~ 80% or higher accuracy, with SVM reaching over 99%. Similar accuracy was achieved using expert-curated gene lists, e.g., SenMayo and CellAge, instead of our algorithm-prioritized gene list using minimum redundancy-maximum relevance (mRMR). However, only a few genes overlapped between the gene sets, suggesting a wide impact of senescence on the transcriptome. Overall, our study demonstrated a proof-of-concept for identifying senescence using ML.