Smoother: A Unified and Modular Framework to Incorporate Structural Dependency in Spatial Omics Data

Description

Smoother is a Python package built for modeling spatial dependency and enforcing spatial coherence in spatial omics data analysis. Implemented in Pytorch, Smoother is modular and ultra-efficient, often capable of analyzing samples tens of thousands of spots in seconds. The key innovation of Smoother is the decoupling of the prior belief on spatial structure (i.e., neighboring spots tend to be more similar) from the likelihood of a non-spatial data-generating model. This flexibility allows the same prior to be used in different models, and the same model to accommodate data with varying or even zero spatial structures. In other words, Smoother can be seamlessly integrated into existing non-spatial models and pipelines (e.g. single-cell analyses) and make them spatially aware. In particular, Smoother provides the following functionalities:

1. Spatial loss: A quadratic loss equivalent to a multivariate Gaussian (MVN) prior reflecting the spatial structure of the data. It can be used to regularize any spatial random variable of interest.
2. Data imputation: Mitigates technical noise by borrowing information from the neighboring spots. It can also be applied to enhance the resolution of the data to an arbitrary level in seconds.
3. Cell-type deconvolution: Infers the spatially coherent cell-type composition of each spot using reference cell-type expression profiles. Smoother is one of the few deconvolution methods that actually enforce spatial coherence by design.
4. Dimension reduction: Find the spatially aware latent representations of spatial omics data in a model-agnostic manner, such that single-cell data without spatial structure can be jointly analyzed using the same pipeline.

See the documentation page for basic usages, tutorials and examples. For mathematical details, check the Smoother paper (Su Jiayu, et al. 2022) and the Supplementary Notes.

Installation

If you only want to use the core functionalities, namely SpatialWeightMatrix and SpatialLoss, Smoother can be directly installed using pip

pip install git+https://github.com/JiayuSuPKU/Smoother.git#egg=smoother

The dimensionality reduction module (SpatialAE, SpatialVAE) is built upon scvi-tools. Here we refer to the original repository for installation instructions on different systems.

pip install scvi-tools

• Note that scvi-tools doesn't officially support Apple's M chips yet. To run SCVI and the corresponding SpatialVAE on Macs with Apple silicon, a temporary solution is to compile both Pytorch and PyG on M1 chips using compatible wheel files.

To solve data imputation and deconvolution models using convex optimization, you need to also install the 'cvxpy' package.

conda install -c conda-forge cvxpy

To run other functions, e.g. the simulation scripts, we recommend using the conda environment provided in the repo. You can create a new conda environment called 'smoother' and install the package in it using the following commands:

# download the repo from github
git clone git@github.com:JiayuSuPKU/Smoother.git

# cd into the repo and create a new conda environment called 'smoother'
conda env create --file environment.yml
conda activate smoother

# add the new conda enviroment to Jupyter
python -m ipykernel install --user --name=smoother

# install the package
pip install -e .

Basic usage

Spatial loss construction

# import spatial losses and models
import torch
from smoother import SpatialWeightMatrix, SpatialLoss, ContrastiveSpatialLoss
from smoother.models.deconv import NNLS
from smoother.models.reduction import SpatialPCA, SpatialVAE

# load data
x = torch.tensor(...) # n_gene x n_celltype, the reference signature matrix
y = torch.tensor(...) # n_gene x n_spot, the spatial count matrix
coords = pd.read_csv(...) # n_spot x 2, tspatial coordinates

# build spatial weight matrix
weights = SpatialWeightMatrix()
weights.calc_weights_knn(coords)

# scale weights by transcriptomics similarity
weights.scale_by_expr(y)

# transform it into spatial loss
spatial_loss = SpatialLoss('icar', weights, rho=0.99)

# or contrastive loss
spatial_loss = ContrastiveSpatialLoss(
    spatial_weights=weights, num_perm=20, neg2pos_ratio=0.1)

# regularize any spatial random variable of interest
variable_of_interest = torch.tensor(...) # n_vars x n_spot
loss = spatial_loss(variable_of_interest)

Downstream tasks

# choose model and solve the problem
# deconvolution
model = NNLS()
model.deconv(x, y, spatial_loss=spatial_loss, lambda_spatial_loss=1, ...)

# dimension reduction de novo from spatial data
SpatialVAE.setup_anndata(adata, layer="raw")
model = SpatialVAE(st_adata=adata, spatial_loss=spatial_loss)
model.train(max_epochs = 400, lr = 0.01, accelerator='cpu')

# dimension reduction from single-cell models
baseline = SpatialPCA(rna_adata, layer='scaled', n_latent=30)
baseline.reduce(...)
model_sp = SpatialPCA.from_rna_model(
    rna_model=baseline, st_adata=sp_data, layer='scaled',
    spatial_loss=spatial_loss, lambda_spatial_loss=0.1
)

model_sp = SpatialVAE.from_rna_model(
    st_adata = sp_data, sc_model = rna_scvi_model, 
    spatial_loss=sp_loss, lambda_spatial_loss=0.01,
    unfrozen=True,
)

Tutorials and examples

https://smoother.readthedocs.io/en/latest/index.html

Citation

Su, Jiayu, et al. "Smoother: a unified and modular framework for incorporating structural dependency in spatial omics data." Genome Biology 24.1 (2023): 291. https://link.springer.com/article/10.1186/s13059-023-03138-x