Kipoi - Seminar

The monthly virtual seminar series is designed as a platform for interested Kipoi users and developers and will host talks on the applications of deep learning on biological data. The seminar is held on every first Wednesday of the month at 5:30 p.m. - 6:30 p.m. CET. We are also happy to share the recordings of the seminar on YouTube.

How to take part

The Virtual Seminar Series takes place via Zoom. To take part in the seminar, you can register for the online Zoom conference. Your personal join link will be valid for all upcoming lectures of the series.

Register

How to apply as a speaker

The seminar is a great opportunity to present your recent work to a large international audience. If you want to apply as a speaker, please use the contact in the registration confirmation email.

Next seminar

Title: Advanced training strategies for genomic sequence-to-function models

5 March 2025 5:30 p.m. - 6:30 p.m. Central European Time

Speaker: Alexander Sasse, Heidelberg University

Abstract:

Multi-task Convolutional Neural Networks (CNNs) have emerged as powerful tools for deciphering how DNA sequence influences gene regulatory features such as chromatin accessibility and transcript abundance. These models can learn sequence patterns recognized by regulatory factors that control gene expression, theoretically enabling prediction of individual genomic variant effects across trained cell types. However, our recent studies revealed that despite strong performance on various variant effect prediction benchmarks (Avsec et al. 2021), these models fail to correctly determine how variants affect gene expression direction across individuals (Sasse et al. 2023), an essential capability for associating variants with phenotypes or diseases. To address these limitations and enhance model learning from available data, I present two strategies. First, training with sequence variation: we developed a modeling approach that directly contrasts sequence differences to predict allele-specific and personalized functional measurements from RNA-seq, ATAC-seq, and ChIP-seq (Spiro and Tu et al. 2025). We applied this approach to F1 mouse data and personal genomes with varying degrees of success. Second, training at higher resolution: we created models that analyze ATAC-seq at base-pair resolution, capturing both overall chromatin accessibility and the distribution of Tn5 transposase cuts (Chandra et al. 2025). Our results demonstrate that incorporating ATACprofile information consistently improves differential chromatin accessibility predictions. Furthermore, simultaneous learning across related cell types through multi-task modeling outperforms single-task approaches. Systematic analysis of the models sequence attributions suggests that base-pair resolution training enables the model to learn a more sensitive representation of the regulatory syntax driving differences between immunocytes, potentially improving variant effect predictions.

Upcoming speakers

Previous speakers

The scientific committee