TIVelo: RNA velocity estimation leveraging cluster-level trajectory inference

RNA velocity inference is a valuable tool for understanding cell development, differentiation, and disease progression. However, existing RNA velocity inference methods typically rely on explicit assumptions of ordinary differential equations (ODE), which prohibits them to capture complex transcriptome expression patterns. In this study, we introduce TIVelo, a novel RNA velocity estimation approach that first determines the velocity direction at the cell cluster level based on trajectory inference, before estimating velocity for individual cells. TIVelo calculates an orientation score to infer the direction at the cluster level without an explicit ODE assumption, which effectively captures complex transcriptional patterns, avoiding potential inconsistencies in velocity estimation for genes that do not follow the simple ODE assumption. We validated the effectiveness of TIVelo by its application to 16 real datasets and the comparison with five benchmarking methods.

Demo

The instructions for running TIVelo for different kinds of datasets can be found in examples.

Reproducibility

To reproduce our results, please refer to the folder notebooks or our tutorial website.

Datasets

All datasets used in this research are openly accessible to the public. To achieve the datasets used in this research, please refer to our Figshare page. The datasets in Figshare are the version after scVelo standard preprocessing pipeline.

Installation

TIVelo requires Python 3.8 or later. We recommend using Miniconda for managing the environment. The typical time for installing our package is 3 minutes.

Step 1: Create and Activate the Conda Environment

First, create a new Conda environment with Python 3.9:

conda create -n tivelo python=3.9 -y
conda activate tivelo

Step 2: Install Dependencies

We have published the TIVelo package on PyPI. To ensure a smooth and stable installation process, we recommend installing large dependencies separately before installing TIVelo in a Conda environment.

PyTorch

Install PyTorch along with torchvision, torchaudio, and CUDA support:

conda install pytorch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 pytorch-cuda=12.1 -c pytorch -c nvidia -y

Numba

Install Numba:

To enable CUDA GPU support for Numba, install the latest NVIDIA graphics drivers for your platform (the open-source Nouveau drivers do not support CUDA). Then install the CUDA Toolkit package.

For CUDA 12, install the following:

conda install -c conda-forge cuda-nvcc cuda-nvrtc "cuda-version>=12.0" -y

For CUDA 11, install the following:

conda install -c conda-forge cudatoolkit "cuda-version>=11.2,<12.0" -y

Note: You do not need to install the CUDA SDK from NVIDIA.

Cpu version

conda install numba

Scanpy

Install Scanpy along with additional dependencies:

conda install -c conda-forge scanpy python-igraph leidenalg -y

scVelo

Install scVelo:

pip install  scvelo==0.3.1

Optional dependencies for directed PAGA and Louvain modularity:

pip install igraph louvain

Optional dependencies for fast neighbor search via hnswlib:

pip install pybind11 hnswlib

Step 3: Install TIVelo

Finally, install TIVelo:

pip install tivelo

JupyterLab

To run the tutorials in a notebook locally, please install JupyterLab:

conda install jupyterlab -y

With these steps, TIVelo and its dependencies will be installed and ready for use.

Cite this article

Ge, M., Miao, J., Qi, J. et al. TIVelo: RNA velocity estimation leveraging cluster-level trajectory inference. Nat Commun 16, 6258 (2025). https://doi.org/10.1038/s41467-025-61628-x