Actionable Gene Expression Cell Profiling with Foundation ModelsTrnovec, Lena (Avtor) Zupan, Blaž (Mentor) Shaulsky, Gad (Komentor) foundation modelscell embeddingssingle-cell analysistransfer learningModern techniques for analyzing cellular data enable a deeper understanding of diseases and the development of more effective therapies. Recently, foundation models — advanced neural networks trained on vast amounts of data — have been developed, representing an important advancement in the analysis of cellular processes. However, since these models have mostly been trained on human cells, an open question remains: can their knowledge be successfully transferred to evolutionarily distant organisms? In this master’s thesis, we address this very question. As a test case, we chose the social amoeba Dictyostelium discoideum, which diverged from humans more than a billion years ago but has a well-studied biology. We systematically compared different approaches to cellular data analysis, ranging from traditional methods to the latest foundation models, including Geneformer, scGPT, and Universal Cell Embedding (UCE). Our results show that foundation models can effectively analyze even evolutionarily distant organisms, with UCE emerging as the most successful approach. This model is based on analyzing protein sequences rather than gene names, which allows it to recognize functional similarities regardless of evolutionary distance. UCE successfully identified different cell types, key signaling pathways, and developmental transitions in the amoeba, suggesting the existence of universal principles of cellular functioning. Our study thus paves the way for applying foundation models to the study of a wide range of organisms without the need for additional model training.20252025-09-10 13:25:00Magistrsko delo/naloga172663sl