Deep learning the regulatory grammar of Hox transcription factors

A fully funded PhD position in the group of Dr. Alexander Sasse (Center for Synthetic Genomics, ZMBH & IWR), in close collaboration with the group of Prof. Ingrid Lohmann (COS Heidelberg).

Three dose regimes, three occupancy patterns and functional outcomes on the same genome. Which sites are occupied — and so which programs run — follows from the combination of doses and the cis-regulatory syntax of each region.

Hox genes carve the animal body plan along the head-to-tail axis with startling precision — yet in vitro, all eight Drosophila Hox proteins bind more or less the same short AT-rich motif. Hox specificity is not written in motifs, it is written in syntax, and reading that syntax is a sequence-modelling problem. The prize is general: what grammar makes cis-regulatory sequence specify function, and can a model learn it well enough to read it back out?

In the first part of the project, you will use the remarkable public resources of the Drosophila community to train a multi-modal sequence-to-function model that predicts chromatin accessibility, factor occupancy, and transcriptional output directly from DNA — think of it as an AlphaGenome for flies. In the second part, you will use the rich, dose-resolved perturbation datasets of the Lohmann lab to ask how the regulatory grammar changes with dosage, cofactor identity, and Hox paralog identity.

This is a project about reading powerful models, not just training them: designing and training multi-task, multi-modal deep networks; developing attribution and interpretability methods that resolve predictions down to individual nucleotides and recover motif syntax; running in silico perturbations to turn a trained model into testable regulatory rules; and feeding experimental results back into the model. The project runs alongside a second, experimentally focused PhD student in the Lohmann lab, with shared lab meetings and regular joint discussions.

Who we're looking for:

• A Master's degree (or equivalent) in bioinformatics, computational biology, computer science, physics, statistics, or a related quantitative field.
• Strong Python and hands-on experience with a deep learning framework (PyTorch preferred). Comfort with GPU training is a plus.
• Genuine interest in what models learn, not only how well they score — interpretability is at the centre of this project, not an afterthought.
• Experience with genomic data (ATAC-seq, RNA-seq, ChIP-seq) is welcome but not required — we are glad to teach the biology to someone with strong quantitative foundations.
• Good written and spoken English, and excellent communication skills: this project runs alongside an experimentally focused PhD student in the Lohmann lab, and close, regular communication between computation and the wet lab is central to it.

What we offer:

• An exciting project: Build a foundation model, then use it to answer a question the field has argued about for forty years.
• Experimental data designed around the modelling rather than inherited from it, including dose-resolved perturbations.
• A highly collaborative lab working on diverse problems in sequence-to-function modelling, method development, and interpretation.
• In-house GPU infrastructure plus Heidelberg's HPC resources (bwForCluster Helix), in one of Europe's densest life-science environments — ZMBH, Center SynGen, IWR, and COS, alongside EMBL and DKFZ.
• Funding according to TV-L E13 for three years, with the prospect of extension.

How to apply: Send a single PDF to a.sasse@zmbh.uni-heidelberg.de with:

• a short motivation letter (what you want to work on, and why)
• transcripts
• contact details for two referees
• a CV — please add a few bullet points on relevant projects you have worked on, with links to any GitHub repositories you contributed to, and, where the account isn't your own, a short note on your contribution

Applications are open until 15 September 2026; the start date is flexible. Informal enquiries before applying are welcome, to the same address.

Heidelberg University stands for equal opportunities and diversity. Qualified female candidates are especially invited to apply. Disabled persons will be given preference if they are equally qualified. Information on the application process and the collection of personal data is available at www.uni-heidelberg.de/stellenmarkt.

Toward a virtual cell: learning the post-transcriptional regulatory code

Current efforts to build "virtual cells", deep learning models of how cells respond to perturbation, focus almost entirely on transcription: which genes are turned on. They largely ignore the post-transcriptional layer, where RNA-binding proteins (RBPs) decide whether each transcript is spliced, transported, stabilized, or translated. This layer matters: RBP dysregulation underlies neurodegenerative diseases (ALS, FTD, Fragile X) and contributes to several cancers, and the first generation of RNA-targeting therapeutics, splice-modulating drugs, mRNA medicines, is already in clinical use.

From an ML perspective, the problem is well-defined. Hundreds of RBPs have been profiled by CLIP-seq, producing millions of binding sites in public datasets such as ENCODE, yet predicting RBP binding from RNA sequence alone remains an open and challenging problem, even with recent genomic foundation models like AlphaGenome. Building on our group's current modeling work, you will develop deep learning models, convolutional, transformer-based, or extensions of foundation architectures, to learn the sequence grammar of post-transcriptional regulation. These models should be tested on downstream directions: identifying genetic variants that disrupt this layer and contribute to disease, and designing sequences for therapeutic applications such as mRNA vaccines. The specific path is flexible and shaped to your interests: nucleotide-resolution models across many proteins at once, cross-cell-type or cross-species generalization, methodological extensions toward more mechanistic and interpretable models, or integration of additional datasets.

Requirements: strong Python; comfortable with a deep learning frameworks (e.g., PyTorch, TensorFlow). For suitably qualified candidates, a HiWi contract (40 h/month, 6 months) is available.

Search pubmed for publications by Alexander Sasse