brisc: a multithreaded CPU toolkit for single-cell analysis

The rapidly growing volume of single-cell RNA sequencing datasets, ranging from tens of thousands to tens of millions of cells, has posed challenges for existing analysis toolkits such as Scanpy and Seurat. These standard tools require hours to complete analyses, impeding iterative exploration. To address this issue, researchers have introduced brisc, the first comprehensive multithreaded CPU toolkit for single-cell analysis.

Brisc encompasses a range of functionalities, including dimensionality reduction, integration, label transfer, clustering, embedding, differential expression, and plotting.

What sets brisc apart from its GPU-based counterparts is its strict guarantee of reproducibility. Regardless of the number of threads utilized, brisc ensures identical results in consecutive runs. This reliability is crucial for scientific reproducibility and collaboration. The toolkit achieves this through algorithmic innovations, such as a blocked inverted-file-index-based nearest-neighbors algorithm. This approach significantly reduces errors compared to Scanpy and Seurat, with brisc demonstrating 81-98% fewer errors.

To showcase brisc's performance, experiments were conducted on a 192-CPU node and a 96-CPU node. On a 192-CPU node, brisc processed a standard workflow for 10 million cells in just 3 minutes, surpassing Scanpy's performance by a factor of 116 while using less than half of Scanpy's peak memory. Similarly, on a 96-CPU node, brisc matched the speed of rapids-singlecell, which relies on four H100 GPUs.

These results highlight brisc's efficiency and scalability, enabling near-real-time analysis of tens of millions of cells without the need for specialized hardware.