24.07.2025

Abschlussarbeiten, Bachelor- und Masterarbeiten:
Master Thesis: Modeling SNP-Conditioned Cellular Phenotypes via Flow Matching in the labs of Prof. Fabian Theis and Dr. Matthias Heinig   

Context

Advances in high-throughput single-cell profiling and large-scale phenotypic screens have unlocked unprecedented insights into cellular heterogeneity and the effects of experimental perturbations. Tools like CellFlow (Klein et al. 2025), a generative framework based on flow matching, have demonstrated powerful capabilities in predicting cellular responses to perturbations such as drug treatments, gene knockouts, or cytokine stimulations.

In parallel, population-scale genomics projects such as OneK1K (Yazar et al. 2022), which combine genetic variation (e.g., SNPs) with single-cell RNA-seq data, provide a rich substrate for exploring how genetic variation modulates cellular phenotypes. This thesis will integrate these two directions, leveraging CellFlow to model how SNP-defined genetic backgrounds influence cellular responses to perturbations, with a particular focus on QTL-associated variants and epistatic interactions.

Thesis Goals and Research Questions

The primary objective of this thesis is to extend CellFlow to incorporate individual-specific genetic variation, particularly common SNPs, as conditioning variables. This enables predictive modeling of single-cell phenotypes under both genetic and experimental perturbation contexts, with applications ranging from fundamental questions in gene regulation, genetic risk prediction and genotype-informed response prediction. A longer-term goal is to position this framework as a tool for modeling cellular phenotypes relevant to disease predisposition, therapeutic response, and inter-individual variability. Specific questions include:

• Can we model epistatic interactions between SNPs in terms of their effect on phenotypic response trajectories?
• Can we identify causal variants and study gene-environment interactions?
• How well can CellFlow predict phenotypes for unseen SNP combinations or rare alleles, particularly under perturbation conditions?
• Can SNP-conditioned models reveal genotype-dependent variation in cellular responses that align with known or hypothesized disease mechanisms?
• To what extent can SNP-conditioned models be generalized or transferred across genetically diverse populations or cell types?

Methodology and Scope

• Adapt the CellFlow architecture to ingest SNP-based sample covariates, including QTLs and epistatic variant pairs.
• Use genotype-informed embeddings (e.g., one-hot encoding, gene embeddings inferred from ESM3 (Hayes et al. 2025), learned SNP embeddings inferred from DNA language models (Nguyen et al. 2023; Hingerl et al. 2025), or integration with known functional annotations (Zheng et al. 2024)).
• Evaluate on datasets such as OneK1K, eQTL Consortium, or sc-eQTLGen.
• Compare against baseline models, such as:
  • Conditional VAEs
  • Epistasis prediction models

Candidate Profile

• Strong background in machine learning, computational biology, or bioinformatics
• Familiarity with deep generative models
• Experience with single-cell RNA-seq data and basic statistical genetics
• Programming skills in Python, PyTorch or JAX, and data science tools

How to Apply

Send an email to matthias.heinig@tum.de, dominik.klein@helmholtz-munich.de, and lucas.arnoldt@tum.de with the following information:

• Your CV
• A brief introduction outlining your background and motivation
• Your preferred start date
• Academic transcripts

We look forward to receiving your email!

References

• Klein et al. (2025). CellFlow enables generative single-cell phenotype modeling with flow matching. bioRxiv; DOI: 10.1101/2025.04.11.648220
• Yazar et al. (2022). Single-cell eQTL mapping identifies cell type–specific genetic control of autoimmune disease. Science, 376 (6589); DOI: 10.1126/science.abf3041
• Hayes et al. (2025). Simulating 500 million years of evolution with a language model. Science, 387 (6736); DOI: 10.1126/science.ads0018
• Nguyen et al. (2023). HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution. arXiv; DOI: 10.48550/arXiv.2306.15794
• Hingerl et al. (2025). scooby: Modeling multi-modal genomic profiles from DNA sequence at single-cell resolution. bioRxiv; DOI: 10.1101/2024.09.19.613754
• Zheng et al. (2024). Leveraging functional genomic annotations and genome coverage to improve polygenic prediction of complex traits within and between ancestries. Nature Genetics, 56; DOI: 10.1038/s41588-024-01704-y