NeuroVelo: interpretable learning of cellular dynamics

NeuroVelo: physics-based interpretable learning of cellular dynamics. It is implemented on Python3 and PyTorch, the model estimate velocity field and genes that drives the splicing dynamics.

The main contributions of NeuroVelo are,

• Using linear projection and embedding for spliced and unspliced RNA to keep interpretability.
• Introducing sample specific velocity estimation, and sample specific interpretation of cellular dynamics.
• Presenting a loss function based on splicing dynamics.

Installation

pip install neurovelo

or

pip install git+https://github.com/idriskb/NeuroVelo

To avoid potential conflict, it is advised to create a seperate virtual envrionment to run the method

Getting started

1. Import package

from neurovelo.train import Trainer
from neurovelo.utils import ModelAnalyzer, latent_data, evaluate, decode_gene_velocity

1. Training

model = Trainer(adata, sample_obs='sample')
mode.train()
model.save_model('/to/folder/','trained.pth')

2. Visualization You can visualize both velocity in latent space or velocity in gene space

To visualize velocity in latent space, just use latent_data function then the usual visualization pipeline from ScVelo

latent_adata = latent_data(adata, '/to/folder/trained.pth')
scv.pp.neighbors(latent_adata, use_rep='X_z', n_neighbors=20)
scv.tl.velocity_graph(latent_adata, vkey='spliced_velocity', xkey='spliced')
scv.tl.velocity_embedding(latent_adata, vkey='spliced_velocity', basis='umap')
scv.pl.velocity_embedding_stream(latent_adata,basis='umap',vkey='spliced_velocity', color='sample')

To visualize velocity in gene space, just use decode_gene_velocity function then the usual visualization pipeline from ScVelo

adata.layers['spliced_velocity'] = decode_gene_velocity(adata, '/to/folder/trained.pth', layer='spliced')
scv.tl.velocity_graph(adata, vkey='spliced_velocity', xkey='spliced')
scv.tl.velocity_embedding(adata, vkey='spliced_velocity', basis='umap')
scv.pl.velocity_embedding_stream(adata,basis='umap',vkey='spliced_velocity', color='sample')

3. Analysis

   1. Gene ranking

      analyzer = ModelAnalyzer(adata, n_vectors=20, '/path/to/trained_models/')
      results = analyzer.models_output()
      gene_ranking_order, gene_ranking_mean = analyzer.gene_ranking()

      gene_ranking_mean can be directly used with prerank gene set enrichment analysis

   2. Gene network

      genes_of_interest = gene_ranking_order['no_treatment'][0][:15].values  # List of genes you're interested in
      
      # Call the graph_maker method
      A, gi, g, mean_ptime, sorted_cell_value, cell_of_ptime = graph_maker.graph_maker(
      	genes_of_interest,
      )
      # Step 3: Create the graph
      threshold_a = 0.05  # Threshold for significant connections
      G = graph_maker.create_graph(A, gi, g, threshold_a)
      # Step 4: Plot the graph
      graph_maker.plot_custom_graph(G,edge_width=20)

      The output is a gene interaction network indeacting how a gene affect the expression of another gene, the package will be extended to make this animated networks for 2 different branches in the differentation of human bone marrow

For further and detailed instructions check the notebooks

Citation

If you find this work useful please cite:

@article {Idris2023.11.17.567500,
	author = {Kouadri Boudjelthia, Idris and Milite, Salvatore and El Kazwini, Nour and Huang, Yuanhua and Sottoriva, Andrea and Sanguinetti, Guido},
	title = {NeuroVelo: interpretable learning of cellular dynamics from single-cell transcriptomic data},
	elocation-id = {2023.11.17.567500},
	year = {2024},
	doi = {10.1101/2023.11.17.567500},
	publisher = {Cold Spring Harbor Laboratory},
	abstract = {Reconstructing temporal cellular dynamics from static single-cell transcriptomics remains a major challenge. Methods based on RNA velocity are useful, but interpreting their results to learn new biology remains difficult, and their predictive power is limited. Here we propose NeuroVelo, a method that couples learning of an optimal linear projection with non-linear Neural Ordinary Differential Equations. Unlike current methods, it uses dynamical systems theory to model biological processes over time, hence NeuroVelo can identify what genes and mechanisms drive the temporal cellular dynamics. We benchmark NeuroVelo against several state-of-the-art methods using single-cell datasets, demonstrating that NeuroVelo has high predictive power but is superior to competing methods in identifying the mechanisms that drive cellular dynamics over time. We also show how we can use this method to infer gene regulatory networks that drive cell fate directly from the data. NeuroVelo is available at: https://github.com/idriskb/NeuroVelo},
	URL = {https://www.biorxiv.org/content/early/2024/06/10/2023.11.17.567500},
	eprint = {https://www.biorxiv.org/content/early/2024/06/10/2023.11.17.567500.full.pdf},
	journal = {bioRxiv}
}