TL;DR
In this talk, Dominik Klein presents “Mapping cells through time and space with moscot” at the Foundation Models for Biology Seminar Series by GenBio AI. He introduces moscot, a scalable framework that uses optimal transport to align single-cell data across time, space, and different molecular layers. By connecting scattered snapshots into a cohesive picture, moscot helps researchers uncover how tissues grow, change, and function at the cellular level.
Introduction
Modern single-cell experiments generate snapshots by capturing cells at a single time, in one modality, or from one location. But biology is continuous and complex. To truly understand processes like development, disease, or regeneration, we need tools that can stitch together these snapshots into a living, evolving picture.
That’s where moscot comes in. Developed by Dominik Klein and his team, moscot is a computational framework that uses optimal transport to align single-cell datasets across time, space, and modality, all while scaling to millions of cells. It’s fast, interpretable, and multimodal by design.
This blog summarizes Dominik’s recent talk at the Foundation Models for Biology Seminar Series hosted by GenBio AI, where he shared how moscot works, the challenges it solves, and the future of scalable alignment in biology.
Full Talk
Watch the full talk below:
The Challenge: From Static Snapshots to Dynamic Maps
Single-cell genomics gives us a remarkable view into cell state, but often as disconnected frames. Most methods struggle with:
- Integrating multiple time points
- Mapping cells across spatial coordinates
- Handling multimodal inputs like RNA + ATAC
- Scaling to datasets with millions of cells
Optimal transport is a promising framework to address these problems. But traditional OT methods are too slow and memory-intensive for real-world datasets.
moscot addresses these limitations head-on by making optimal transport scalable and multimodal.
What Is moscot and How Does It Work?
moscot stands for Multi-Omics Single-Cell Optimal Transport. It uses mathematical principles of OT to align distributions of cells across different conditions such as time, space, or modality. Built on top of the OTT JAX library, moscot achieves linear-time and linear-memory scaling using efficient implementations, low-rank approximations and GPU acceleration.
Key features:
- Multimodal support: Aligns RNA, ATAC, and protein data
- Fast performance: Scales to millions of cells
- Unified API: Accessible to both computational and experimental biologists
- Interpretable outputs: Includes transition maps, growth rates, and cell alignments
moscot also supports multiple OT formulations: standard Wasserstein OT, Gromov–Wasserstein (for structural alignment), and fused OT (ideal for multimodal integration).
Use Case 1: Embryonic Development at Scale
In one of its flagship applications, moscot was used to reconstruct mouse embryogenesis from 1.7 million cells across 20 time points. Compared to WOT and TOME:
- moscot handled much larger datasets
- Predicted more realistic growth rates
- Better captured known lineage trajectories using marker genes
moscot accurately mapped differentiation processes, such as endoderm branching and hepatocyte emergence, with higher fidelity than existing tools.
Use Case 2: Multimodal Pancreatic Development
In a paired RNA + ATAC dataset of pancreas development, moscot modeled rare endocrine trajectories, including epsilon and delta cell lineages.
One standout finding: NEUROD2 was identified as a likely regulator of the epsilon lineage, a hypothesis validated in human iPSC-derived organoids. This highlights how moscot supports not just alignment, but biological discovery.
Use Case 3: Spatial and Rotational Alignment
moscot’s ability to bridge spatial and dissociated data opens new doors in tissue analysis. It was used to:
- Align CITE-seq data to liver spatial transcriptomics
- Rotate and register coronal brain sections, revealing consistent anatomical patterns despite orientation changes.
This capability is especially powerful when analyzing tissue regions with imperfect sectioning or orientation inconsistencies.
Use Case 4: Spatiotemporal Modeling
moscot’s spatiotemporal module jointly models how tissues evolve over both space and time. For example:
- In the mouse heart, moscot improved region annotations across developmental stages
- In oral cavity development, it improved transition accuracy and structural understanding
This module allows researchers to move beyond trajectory inference and begin building comprehensive, 4D biological maps.
Final Thoughts
moscot is a major step forward in making optimal transport practical for single-cell biology. With its ability to handle large, multimodal datasets and deliver biologically meaningful alignments, it opens the door to a more dynamic and interconnected understanding of life at the cellular level.
Beyond aligning cells across time, space, and modality, emerging methods such as neural optimal transport extend the possibilities even further. When no perfect match exists, these methods allow us to simulate likely counterparts under hypothetical conditions. Just like imagining a version of Harry Potter without glasses, neural OT helps generate previously unseen cell states.
Dominik explores this approach in two recent examples Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation and GENOT: Entropic (Gromov) Wasserstein Flow Matching with Applications to Single-Cell Genomics, which demonstrate how relaxing classic OT constraints can better preserve biological features across domains. This opens exciting new opportunities in counterfactual modeling, disease progression analysis, and in silico perturbation experiments, ultimately accelerating discoveries in drug development and personalized medicine.
Whether you are mapping tissues, modeling development, or integrating multi-omics data, moscot makes these tasks accessible, scalable, and interpretable, fitting seamlessly into both computational and experimental biology workflows.
For further details, check out the full paper on moscot.
Join us in our mission to push the frontiers of AI-driven biology and strive to make a lasting impact on medicine, biotechnology, and human health. We are hiring across teams. Visit our Careers page to learn more and apply. Follow us on X, YouTube, and LinkedIn.
