Search

Single-cell technologies have provided valuable insights into the distribution of molecular features, such as gene expression and epigenomic modifications. However, comparing these complex distributions in a controlled and powerful manner poses methodological challenges. Here we propose to benefit from the kernel-testing framework to compare the complex cell-wise distributions of molecular features in a non-linear manner based on their kernel embedding. Our framework not only allows for feature-wise analyses but also enables global comparisons of transcriptomes or epigenomes, considering their intricate dependencies. By using a classifier to discriminate cells based on the variability of their embedding, our method uncovers heterogeneities in cell populations that would otherwise go undetected. We show that kernel testing overcomes the limitations of differential analysis methods dedicated to single-cell. Kernel testing is applied to investigate the reversion process of differentiating cells, successfully identifying cells in transition between reversion and differentiation stages. Additionally, we analyze single-cell ChIP-Seq data and identify a subpopulation of untreated breast cancer cells that exhibit an epigenomic profile similar to persister cells.

The KeOps library provides a fast and memory-efficient GPU support for tensors whose entries are given by a mathematical formula, such as kernel and distance matrices. KeOps alleviates the major bottleneck of tensor-centric libraries for kernel and geometric applications: memory consumption. It also supports automatic differentiation and outperforms standard GPU baselines, including PyTorch CUDA tensors or the Halide and TVM libraries. KeOps combines optimized C++/CUDA schemes with binders for high-level languages: Python (Numpy and PyTorch), Matlab and GNU R. As a result, high-level "quadratic" codes can now scale up to large data sets with millions of samples processed in seconds. KeOps brings graphics-like performances for kernel methods and is freely available on standard repositories (PyPi, CRAN). To showcase its versatility, we provide tutorials in a wide range of settings online at \url{www.kernel-operations.io}., Comment: 6 pages