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/CEST. We are also happy to share the recordings of the seminar on YouTube.

How to take part

Our Virtual Seminar Series is hosted entirely online. To join, please subscribe to the mailing list below; we will send you a single recurring link that provides access to every lecture in 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: Signal, Bounds, and Baselines: Principles for Evaluating Virtual Cell Perturbation Models

2 September 2026 5:30 p.m. - 6:30 p.m. CET/CEST

Speaker: Michael Vollenweider, Peter Bühlmann group, Seminar for Statistics, ETH Zurich

Abstract:

Foundation models and deep learning systems are increasingly proposed as core components of "virtual cells" that forecast transcriptomic responses to unseen perturbations. But evaluating such predictions in high-dimensional gene expression space is harder than it looks, and reported performance often does not survive closer inspection. This talk introduces the SBB principles (Signal, Bounds, and Baselines) for evaluating biological perturbation prediction. The Signal pillar provides diagnostic meta-metrics that assess how sensitive an evaluation metric is to biological signal masked by high-dimensional noise, and how much signal a dataset contains in the first place. The Bounds pillar uses technical duplicates and uninformative controls as empirical reference points, showing how far a model is from the achievable optimum on a given perturbation, and whether that gap is meaningful. The Baselines pillar establishes a hierarchy of interpretable linear models as performance floors. Across seven single- and double-perturbation datasets, deep learning methods often fail to meaningfully surpass simple linear baselines, and substantial room for improvement remains where they do.

Upcoming speakers

Previous speakers

The scientific committee