C18 - Machine learning-assisted analyses for identification and prediction of genomic abnormalities in human pluripotent stem cell populations

Introduction: : Human pluripotent stem cells (hPSCs) have generated much interest in the scientific community based on their potential applications in regenerative medicine. However, scalability and genomic stability challenges need to be addressed before use in clinical applications. Work from our laboratory and others has demonstrated that amplification of chromosomal regions is correlated with increased gene expression. Before use in biomanufacturing processes for regenerative cell-based therapies, accurate classification of cells to detect abnormalities is critical. Such classifications will ensure that only healthy, viable cells are utilized in therapeutic applications, preventing issues that could limit effectiveness or introduce complications in clinical practice.

One such method of classification is via machine learning, which is a transformative tool that allows researchers to process and interpret vast, complex datasets, including stem cell gene expression profiles. By leveraging machine learning, we can uncover subtle patterns within these profiles that might otherwise go undetected, which offers a deeper understanding of cell quality. Machine learning models can analyze thousands of genes simultaneously, allowing them to identify key biomarkers and expression patterns that distinguish normal from abnormal stem cells. This capability is valuable for this classification task and the potential for future predictive modeling. Furthermore, machine learning allows for high-throughput analysis, making it possible to evaluate large numbers of stem cells quickly and with less bias and greater precision. This not only accelerates the research process but also supports scalable, reproducible insights into stem cell health, ultimately enhancing regenerative medicine approaches and the safe application of stem cell therapies.

Materials and

Methods: : The methodology followed a systematic workflow, beginning with the acquisition and preparation of the gene expression data. This was followed by the implementation and training of four classification models: Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting. Finally, the performance of each model was evaluated using a range of metrics, providing a comprehensive assessment of its effectiveness in classifying stem cell abnormalities. The gene expression data used for this study were derived from a public database called StemMapper, which is a well-known and well-respected source that contains manually curated gene expression data for different lineages of human and mouse stem cells. Our focus was on hESCs and hiPSCs, which included a total of 68 samples, each encompassing 20,273 evaluated genes, resulting in a high-dimensional dataset with over 1.5 million data points. To determine biologically accurate class labels, a manual thresholding approach was used to define what constituted an “abnormal” sample based on the number of highly expressed target genes above the median. The goal was to identify a cutoff that produced a realistic and interpretable division of samples into “normal” and “abnormal” groups. Based on the sample labeling outputs, it was determined that 13 genes would be used as the threshold for this study, as this indicated that 59.1% of the target genes expressed above their median value and provided a significant population for both normal and abnormal samples. The selected threshold achieved a balance between statistical practicality and biological plausibility, ultimately enabling more meaningful classification and analysis.

Results, Conclusions, and Discussions:: The results of the four machine learning models offer critical insights into their relative strengths and limitations in the context of classifying stem cell abnormalities. Among the models evaluated, Random Forest emerged as the most effective for the primary goal of maximizing recall, correctly identifying the highest proportion of abnormal stem cell samples while maintaining strong overall performance across other metrics. Although Random Forest tended to produce more false positives, this skew reflects a conservative approach, favoring the identification of abnormal stem cells, even at the risk of overcalling. In the context of stem cell therapy, this trade-off is desirable: a false negative could allow a harmful cell to slip through, while a false positive simply errs on the side of caution. Random Forest’s ensemble-based architecture enables it to capture complex, non-linear patterns within high-dimensional gene expression data without strong parametric assumptions, making it particularly well-suited for this application. Compared to simpler models like Logistic Regression, Random Forest offers greater flexibility and robustness while achieving comparable recall, ultimately supporting more confident identification of stem cell abnormalities.

Culturing stem cell lines and monitoring them for signs of abnormality is resource-intensive and often relies on periodic genetic testing or subjective morphological assessment. A robust machine learning classifier trained to detect subtle patterns in gene expression could help flag potentially unstable or abnormal lines early in the workflow, allowing researchers to focus validation efforts on only the most promising cultures. Early identification of abnormal expression profiles is particularly important given the tendency of hPSCs to acquire genetic aberrations over time, even in well-controlled environments.

In the context of stem cell manufacturing, the need for consistent quality control is especially critical. Regulatory frameworks increasingly demand rigorous characterization of stem cell products. Integrating a Random Forest-based screening system into workflows could help ensure that only genomically stable lines are stored or distributed. Automated abnormality detection would offer reproducible, unbiased assessments across batches, enhancing the safety and consistency of products available for research or therapeutic development.

Acknowledgements and/or References (Optional)::