Transfer Learning for Phenotype Prediction from Small Gene Expression Data SetsMohorčič, Domen (Avtor) Zupan, Blaž (Mentor) gene expressionsmall data set problemtransfer learningautoencodersmulti-task modelsRecent advances in biotechnology have enabled researchers to collect huge amounts of data, such as gene expression profiles from patients, which provide a foundation for personalized medicine. Such an approach requires the use of machine learning, however, a significant limitation of many medical studies is the small sample size, typically having only a few hundred patients with tens of thousands of features. In this thesis, we addressed this issue by combining multiple small gene expression data sets into a larger one, regardless of the study type, and training deep learning models capable of producing informative gene expression encodings. We used transfer learning to predict the phenotypes on unseen data sets based on the created encodings. We experimented with two model architectures: autoencoders and multi-task models. Although training multi-task models proved challenging, they achieved higher average results on test data sets than autoencoders but never surpassed the results of logistic regression. An examination of the encodings revealed that autoencoders maintained the original data structure whereas the multi-task models mixed samples from different studies, but both proved that the gene expression profile can be reduced to a few informative markers.20242024-09-13 14:45:01Magistrsko delo/naloga161755VisID: 37055COBISS_ID: 210383875sl