Overview
Single-cell transcriptomics measures the activity of thousands of genes in individual cells, producing extremely high-dimensional datasets. To interpret these data, researchers typically project them into two or three dimensions for visualization. Widely used methods such as t-SNE and UMAP can reveal local structure but may distort global relationships or depend heavily on parameter choices.
Led by Dmitry Kobak at the University of Tübingen, the project “Contrastive learning for dimensionality reduction and visualization of transcriptomic data” (2021–2025) investigated whether contrastive learning—a self-supervised machine-learning approach—could improve dimensionality reduction and visualization for single-cell and other transcriptomic datasets.
Key results
- Developed a conceptual link between t-SNE and UMAP via contrastive learning. The project showed that t-SNE can be optimised using noise-contrastive estimation, while UMAP relies on negative sampling, another contrastive method. This clarified why the two methods often produce visually different embeddings.
- Proposed and implemented a generalization of negative sampling that allows interpolation (and even extrapolation) between t-SNE-like and UMAP-like embeddings. Moving along this spectrum reveals a trade-off between local/discrete and global/continuous structure, reducing the risk of over-interpreting features of any single embedding.
- Created parametric contrastive-learning methods for 2D visualization, such as t-SimCNE and related approaches, which train neural networks to directly map high-dimensional data into two dimensions using contrastive objectives.
- Applied contrastive and manifold-learning ideas to single-cell RNA-seq and other biological datasets, improving the stability and interpretability of low-dimensional visualizations.
- Provided open-source implementations (e.g., in PyTorch and as part of the openTSNE ecosystem) to make these methods accessible to the bioinformatics and machine-learning communities.
Why it matters
Dimensionality reduction and visualization are central to analysing modern transcriptomic datasets. By grounding popular methods like t-SNE and UMAP in a unified contrastive-learning framework, the project helps researchers understand what these visualizations represent and how they depend on algorithmic choices. The resulting methods and guidelines can support more reliable interpretation of single-cell data in neuroscience, immunology, developmental biology, and related fields.
This project produced the following peer-reviewed publications:
- Kobak, D., & Berens, P. (2019). The art of using t-SNE for single-cell transcriptomics. Nature Communications, 10, 5416. doi:10.1038/s41467-019-13056-x
- Damrich, S., Böhm, J. N., Hamprecht, F. A., & Kobak, D. (2023). From t-SNE to UMAP with contrastive learning. In International Conference on Learning Representations (ICLR) 2023. arXiv:2206.01816 [cs.LG]
- Böhm, J. N., Berens, P., & Kobak, D. (2023). Unsupervised visualization of image datasets using contrastive learning. In International Conference on Learning Representations (ICLR) 2023. arXiv:2206.01816 [cs.LG] (t-SimCNE work)
- Kobak, D., & Berens, P. (2020). Heavy-tailed kernels reveal a finer cluster structure in t-SNE. Philosophical Transactions of the Royal Society A, 378(2180), 20190231. doi:10.1098/rsta.2019.0231
