MetaPhenotype:\nA Transferable Meta-Learning Model\nfor Single-Cell Mass Spectrometry-Based Cell Phenotype Prediction\nUsing Limited Number of Cells

Single-cell mass spectrometry (SCMS) is an emerging tool\nfor studying\ncell heterogeneity according to variation of molecular species in\nsingle cells. Although it has become increasingly common to employ\nmachine learning models in SCMS data analysis, such as the classification\nof cell phenotypes, the existing machine learning models often suffer\nfrom low adaptability and transferability. In addition, SCMS studies\nof rare cells can be restricted by limited number of cell samples.\nTo overcome these limitations, we performed SCMS analyses of melanoma\ncancer cell lines with two phenotypes (i.e., primary and metastatic\ncells). We then developed a meta-learning-based model, MetaPhenotype,\nthat can be trained using a small amount of SCMS data to accurately\nclassify cells into primary or metastatic phenotypes. Our results\nshow that compared with standard transfer learning models, MetaPhenotype\ncan rapidly predict and achieve a high accuracy of over 90% with fewer\nnew training samples. Overall, our work opens the possibility of accurate\ncell phenotype classification based on fewer SCMS samples, thus lowering\nthe demand for sample acquisition.