Séance 2 : L’IA pour la prédiction génomique

Vendredi 7 novembre 2025
13h30 – 15h00
En ligne | Gratuit sur inscription

Au programme de cette séance :

Apport des approches d'apprentissage profond à la prédiction génomique et à l'appui aux objectifs de sélection

Par Fatima Shokor (jeune docteure, thèse soutenue en septembre 2024, UMR GABI)

Modern animal breeding depends on predicting breeding values to guide selection and accelerate genetic gain. Current genomic prediction methods, though effective, assume additivity and linearity, limiting their ability to capture complex trait architectures. This project explored integrating deep learning with statistics to model non-linear and non-additive effects, focusing on multi-trait and crossbred prediction.

Evaluation of the Performance of Neural Network Cross-Species Predictions of Chromatin Regulation Annotations

 Par Noémien Maillard (doctorant en 3ème année, GenPhyse - labellisé par DIGIT-BIO)

A better knowledge of functional annotations of livestock species can be a lever to link genome to phenome. The genomes of most livestock species have already been sequenced. However, data describing gene regulation mechanisms and chromatin state are insufficient. In contrast, abundant human and mouse data allowed the training of powerful deep learning algorithms. Here, we propose to use 3 artificial neural networks (Deepbind, DeepSEA and Enformer), trained with human and mouse data, to predict annotations on the pig, cattle, chicken and European seabass genomes. The predictions are then compared with experimental data to evaluate the cross-species performance of the neural networks.
First, human-trained neural network predictions performed on the mouse reference genome showed varying levels of accuracy depending on the experiment, with the higher performance for H3K4me3. Second, the predictions on the pig, cattle and chicken genomes showed similar and better performances than those on the seabass genome. Third, we showed a correlation between phylogenetic distance relatively to human and auPRC. Moreover, we showed, for 2 histone marks and CTCF, that the scores are lowly influenced by the sequence conservation. Finally, the evaluation of the impact of genomic features on the predictions highlighted better performances for CpG island and 5’UTR than other features.
To conclude, we showed that neural networks can be used to predict annotations on mammalian and non-mammalian (chicken) genomes with similar performances. We also showed that we can correctly predict annotations independently of sequence conservation.

Animation de la séance : Christèle Robert Granié et Marie-Laure Martin