Serena Chan, Léa Kaufmann et Carl Munoz présentent leurs travaux de recherche lors de Futurs Numériques IVADO 2025

Serena Chan et Léa Kaufmann ont chacune présenté une affiche de leurs travaux de recherche, et Carl Munoz fut sélectionné pour une présentation orale.

La 5e édition de Futurs Numériques IVADO fut une belle occasion pour nos étudiant.e.s de mettre en valeur leurs recherches au près de l’ensemble de la communauté en intelligente (industrielle et académique). Toutes les présentations devaient porter une attention particulière à la vulgarisation sicentifique ainsi qu’aux applications concrètes des travaux de recherche.

AI for RNA-Seq Embeddings in Transcriptomic Profiling

Affiche présentée par Serena Chan [EN]

Gene embeddings capture co-expression patterns from RNA-seq data, enabling deep neural networks to model gene functions and regulation. This project compares the use of raw expression values with ranking-based embeddings, inspired by natural language processing, across tasks including cell line identification and masked token pretraining. Preliminary results show similar performance for identification, while expression profiles outperform ranked embeddings in masked pretraining, with largest errors at mid-ranked genes. Ongoing work will evaluate downstream tasks and explore hybrid architectures to integrate both representations, aiming to capture complementary transcriptomic features and improve prediction accuracy across tasks throughout precision medicine and drug discovery.

AI for predicting the biological activity of chemical compounds in cancer cells

Affiche présentée par Léa Kaufmann [EN] Predicting in silico the effect of existing compounds on cancer cells can enable efficient virtual screening of candidate therapeutics. We develop a deep neural network predicting the expression profile of a treated target cell line given four modalities in input: (1) an embedding of the compound’s structure, (2) an embedding of a reference cell line, (3) the expression profile of the reference line treated with the same compound, and (4) an embedding of the target cell line. The embeddings will be derived from open-source foundation models. The prediction validation will be facilitated, as no de novo synthesis will be necessary.

AI for denoising of low-depth RNA-seq data

Présentation orale par Carl Munoz [EN] RNA sequencing is a powerful but expensive technology. Reducing the sequencing depth decreases the price, but it also the data quality. As such, we are developing machine-learning-based denoising models to increase data quality while maintaining lower costs. Currently, we developed a neural network that denoises low-depth RNA-seq and recovers most information lost. We are also developing a Bayesian normalizing-flow-based framework as an additional method that better captures uncertainty caused by depth reduction. This research has the potential to change RNA-seq standards by reducing costs for all applications, including data acquisition or costs for personalized medicine.