Séminaire AROBAS – Evaluating transfer learning strategies for bulk gene expression data
Kevin DRARJAT, doctorant équipe AROBAS.
Abstract
RNA sequencing (RNA-seq) has become an essential tool in precision oncology, providing comprehensive transcriptome profiles enabling biomarker discovery, patient stratification, and treatment prediction. However, deep learning models for transcriptomic data are limited by small cohort sizes and strong dimensionality constraints. Transfer learning, widely used in computer vision and natural language processing, offers a promising strategy to mitigate these limitations, but its systematic evaluation in transcriptomics remains incomplete. We investigate the application of transfer learning to bulk gene expression data across various tasks. We train multilayer perceptrons (MLPs) and graph neural networks (GNNs) across six classification tasks, including tumour detection, tissue type, stage, survival, and treatment response. Three transfer learning paradigms are explored: pan-cancer to single-cancer transfer, inter-task transfer, and multi-task learning. We further evaluate transferability metrics (LogME, NLEEP, GBC, TransRate) as predictors of transfer effectiveness. Our findings enable to identify situations where transfer learning is most beneficial and provide insights into task relatedness and representation sharing in transcriptomic data.
- Date: 02/04/2026, 13h30-14h
- Lieu: IBISC, site IBGBI, 2ème étage, salle de réunion
- Organisation: Salma MAKBOUL (MCF Univ. Évry, IBISC équipe AROBAS)
