Machine Learning Approaches for Building Virtual Cell Models

Machine Learning Approaches for Building Virtual Cell Models

Cells represent the fundamental building blocks of life, operating as intricate, self-regulating systems that respond dynamically to their environment. A major challenge in modern cell biology is understanding and predicting how these complex systems respond to various perturbations, a capability that would revolutionize drug discovery, disease treatment, and our fundamental understanding of biological processes.

Recent technological breakthroughs now enable researchers to generate vast amounts of single-cell interventional data, capturing cellular responses at unprecedented scale and resolution. However, the sheer complexity of cellular systems and the immense space of possible interventions, makes exhaustive experimental exploration impractical and prohibitively expensive.

This seminar explores the emerging frontier of machine learning models designed to simulate cellular behavior and predict how cells respond to novel perturbations. We will examine recent advances in this rapidly evolving field, including newly proposed generative models, biologically-informed architectures, and methods that aim to make perturbation predictions more explainable.

The seminar consists of an introductory lecture introducing the necessary background, and subsequent weekly meetings with two paper presentations and discussions. Students are expected to read into their assigned paper, the related literature, prepare a talk, as well as a final essay critically discussing the paper.

Requirements: The student has a solid understanding of Machine Learning and feels comfortable with Neural Networks (for example through lectures High Level Computer Vision, Neural Networks: Theory and Implementation, or Machine Learning). Previous biological knowledge is not required, but some interest is preferred.

Places: 12

Dates: Kickoff: 22.10.25, 10am; subsequent meetings are on Wednesdays 10am starting on 12.11.25

Location: Building E2 1 Room 001

Papers: Papers will be selected from this list: https://github.com/sgraf2002/virtual-cell-seminar-2025

Introduction for computer scientists: https://surajparmar.substack.com/p/from-molecules-to-matrices-part-1

Deliverables and Grading Scheme

1. Presentation and Discussion (30%)

   • Talk (30–35 minutes) followed by a Discussion (10–15 minutes) about your assigned paper

   • For each meeting, upload two criticial questions per talk to the CMS and argue why they are interesting (max. 3 sentences, pdf format). On the basis of this, we will evaluate your active participation in the seminar, which will be part of the presentation/discussion grade

   • We will provide a few guidelines on how a good talk should be structured, also including criteria we use for grading

   • Slides should be sent to your instructor 7 days before the talk in pdf format

   • If you would like to discuss the paper or have questions in advance, you are welcome to schedule a meeting with your instructor. Please arrange the meeting at least 7 days before your presentation.

2. Essay (40%)

   • Length: 4 pages (±½ page).

   • The essay should critically evaluate your assigned paper, discussing its strengths and weaknesses, and situate it within the broader context of related literature

   • Keep summaries brief (max. ½ page), the essay should reflect that you deeply thought about the paper and the literature context

   • Due date: 01.03.26, 11:59PM German time

3. Project (30%)

   • Task: Apply current virtual cell methodologies within a simplified target discovery pipeline for Alzheimer’s disease. (Details and exact project description will follow)

   • Grading: Based on a showcase Jupyter notebook and a 15-minute discussion of your approach with the instructors.

Paper Schedule

Established Methods

1. Theodoris, C. et al. (2023). Geneformer. Nature. https://www.nature.com/articles/s41586-023-06139-9 (12.11.25)

2. Roohani, Y. et al. (2023). GEARS. Nature Biotechnology. https://www.nature.com/articles/s41587-023-01905-6 (12.11.25)

3. Cui, H. et al. (2024). scGPT. Nature Methods. https://www.nature.com/articles/s41592-024-02201-0 (26.11.25)

4. Wu, Y. et al. (2023). Variational Causal Inference-Based Modeling. ICLR. https://arxiv.org/pdf/2210.00116 (26.11.25)

5. Baek, S. et al. (2025). GPO-VAE. Bioinformatics (Oxford). https://academic.oup.com/bioinformatics/article/41/Supplement_1/i599/8199372 (03.12.25)

Benchmarking

6. Torne, J. et al. (2025). Systema. Nature Biotechnology. https://www.nature.com/articles/s41587-025-02777-8 (10.12.25)

7. Miller, H. et al. (2025). Well-Calibrated Metric Benchmarking. bioRxiv. https://www.biorxiv.org/content/10.1101/2025.10.20.683304v1.full.pdf (10.12.25)

Recent Methods

8. Adduri, V. et al. (2025). STATE. bioRxiv. https://www.biorxiv.org/content/10.1101/2025.06.26.661135v2 (17.12.25)

9. He, Y. et al. (2025). MORPH. bioRxiv. https://www.biorxiv.org/content/10.1101/2025.06.27.661992v1.full.pdf (17.12.25)

10. Mejia, J. et al. (2025). Diversity by Design. ICML. https://arxiv.org/abs/2506.22641 (07.01.26)

11. Wu, Y. et al. (2025). PerturbQA. ICLR. https://arxiv.org/pdf/2502.21290 (07.01.26)

12. Beheler-Amass, K. et al. (2025). in-CAHOOTTS. bioRxiv. https://www.biorxiv.org/content/10.1101/2025.09.19.676870v1 (14.01.26)