Our paper is now officially published in Nature Genetics and is selected as the Cover Article.

Overview of SpiceMix

SpiceMix is an unsupervised tool for latent representation learning for spatial transcriptomics. SpiceMix models the observed expression of genes within a cell/spot as a mixture of latent factors. These factors are assumed to have some spatial affinity between neighboring cells. The factors and affinities are not known a priori, but are learned by SpiceMix directly from the data. In this way, SpiceMix learns a more expressive representation of the identity of cells from their spatial transcriptome data than other available methods.

SpiceMix can be applied to any type of spatial transcriptomics data, including MERFISH, seqFISH, and Visium, and other spatial omics data.

Reproducibility of our Nature Genetics paper

To reproduce the results presented in our Nature Genetics paper, please use the code in branch NatureGenetics-2022.

Major updates since our Nature Genetics paper

After the acceptance of our paper, we further improved the implementation of SpiceMix. We made the following updates:

• SpiceMix no longer depends on Gurobi.
• All backend optimization algorithms are implemented via PyTorch and can utilize GPU.
• A better initialization algorithm based on Louvain clustering is added and is the default choice.

These updates led to a better performance on the Visium dataset for identifying cortical layers and white matter. Specifically, we trained SpiceMix and did clustering on FOVs 151607-151610, following the same procedure as in our paper. The inferred spot annotations got an ARI score of 0.64 when evaluated at the 4 FOVs collectively. When evaluated on each FOV separately, the ARI scores are 0.64, 0.63, 0.66, and 0.69, respectively. Please see this notebook for details.

Running SpiceMix

Installation

Python package dependencies:

• PyTorch
• pandas
• scikit-learn
• h5py
• tqdm
• scanpy, python-igraph, louvain
• seaborn
• jupyterlab

An environment file for Anaconda can be found here. An Anaconda environment can be set up with the following command

conda env create -f SpiceMix.yml

Running SpiceMix from command line

All code is contained within the SpiceMix folder. The script main.py runs the program -- see examples for usage.

Step 1: Preparing input files

All files for one run of SpiceMix must be put into one directory. SpiceMix can be applied to multiple samples (or replicates, or fields-of-view (FOV)) simultaneously, learning shared parameters across samples (we use the term FOV for an independent sample). For each FOV with name <FOV> of N cells and G genes, the following two files stored as tab-delimited txt format are required for SpiceMix:

• expression_<FOV>_<expr_suffix>.txt, an N-by-G nonnegative-valued matrix of normalized single-cell expression profiles. In our paper, we applied the following steps of normalization to all data sets:
  • Filter out genes with low nonzero rates and/or cells that express only a few genes
  • Log transformation: Let be the read counts of gene g in cell i, and the number of counts after log transformation is
• neighborhood_<FOV>_<neigh_suffix>.txt, a neighbor graph represented as a list of cell pairs, i.e., an |E|-by-2 integer-valued matrix, where |E| is the number of edges in the graph. Cells are assigned with integer indices starting from 0 in the order that they appear in the expression profile file. We recommend the following two methods to generate the neighbor graph from cells' spatial coordinates:
  • K-nearest neighbor graph under Euclidean metric
  • Delaunay triangulation followed by discarding interactions between cells that are far away from each other

The following files for each FOV are required for downstream analysis and visualization:

• genes_<FOV>_<expr_suffix>.txt, a multi-line file containing gene IDs or symbols, one gene per line. The order of genes should match that in expression_<FOV>_<expr_suffix>.txt.
• meta_<FOV>.csv, a comma-separated table with one header line followed by N lines, one for each cell. The order of cells must be the same as the expression_<FOV>_<expr_suffix>.txt files. It is recommended to include the following columns:
  • coor X and coor Y that store the spatial coordinate of single cells.
  • cell type that stores the reference cell type assignments. Note that additional reference can be included with different column names.
• (Deprecated) coordinates_<FOV>.txt, an N-by-2 matrix of single cells' spatial coordinates in the 2D or 3D space. The current tutorial notebook works only with 2D coordinates.
• (Deprecated) celltypes_<FOV>.txt, a multi-line file containing cell types from other analysis, one cell type per line. The order of cells should match that in expression_<FOV>_<expr_suffix>.txt.

When there are multiple isolated FOVs, the order of genes in expression_<FOV>_<expr_suffix>.txt for all FOVs must be identical.

The FOV name, denoted by <FOV> here, is required in order to distinguish between cells from different FOVs, especially when they don't share the same coordinate system. FOV names can be any string, such as '1', 'cortex', 'mouse brain Oct-10-2020'.

To compare different preprocessing approaches, the two suffices, <expr_suffix> and <neigh_suffix>, can be specified for each normalization and neighbor graph, respectively. For example, neighborhood_3_10NN.txt and neighborhood_3_Delaunay.txt may denote the neighbor graph of FOV 3 generated by 10-nearest neighbor and Delaunay triangulation, respectively. It is feasible for different preprocessing approaches to use different sets of genes. For example, we may 1) use all genes and set <expr_suffix> to all, and 2) keep genes that are expressed in at least 10% of cells and set <expr_suffix> to nonzero10. The two suffices are optional, and the preceding underscore _ should be absent when the corresponding suffix is not used (an empty string).

Step 2: Organizing files

We recommend creating one directory for every data set. In the directory, denoted by <dataset>, a directory named files should be created and all aforementioned input files should be put in this directory.

For example, when we have one dataset, named simulation 1, consisting of 3 FOVs, we should arrange the files in the following manner

simulation 1
└── files
    ├── expression_1.txt
    ├── expression_2.txt
    ├── expression_3.txt
    ├── neighborhood_1.txt
    ├── neighborhood_2.txt
    ├── neighborhood_3.txt
    ├── genes_1.txt
    ├── genes_2.txt
    ├── genes_3.txt
    ├── meta_1.csv
    ├── meta_2.csv
    └── meta_3.csv

When we have another dataset, named simulation 2, consisting of 2 FOVs, and we'd like to try different preprocessing approaches - all and nonzero10 for filtering genes and 10NN and Delaunay for generating neighbor graphs. We need to arrange files to

simulation 2
├── files
|   ├── expression_1_all.txt
|   ├── expression_2_all.txt
|   ├── expression_1_nonzero10.txt
|   ├── expression_2_nonzero10.txt
|   ├── neighborhood_1_10NN.txt
|   ├── neighborhood_2_10NN.txt
|   ├── neighborhood_1_Delaunay.txt
|   ├── neighborhood_2_Delaunay.txt
|   ├── genes_1_all.txt
|   ├── genes_2_all.txt
|   ├── genes_1_nonzero10.txt
|   ├── genes_2_nonzero10.txt
|   ├── meta_1.txt
|   └── meta_2.txt
└── some other directories // SpiceMix discards any other directories

Step 3: Inferring latent states, metagenes, and pairwise affinity matrix

SpiceMix requires a few arguments to specify the input files and hyperparameters. Below is the table of the description of all arguments:

Input related parameters

params	type	description	example
--path2dataset	str	path to the dataset	"../data/simulation 1" or "../data/synthetic_cortex"
--neighbor_suffix	str	suffix of the name of the file that contains interacting cell pairs	"", "KNN", or "Delaunay"
--expression_suffix	str	suffix of the name of the file that contains expressions	"", "allgenes", or "top100genes"
--repli_list	list of strings (Python expression)	A Python expression of a list of FOV names	"[0,1,2]", "[A,B]", or "range(3)"
--use_spatial	list of booleans (Python expression)	A Python expression of a list of boolean variables controlling whether to use the neighbor graph for each FOV	"[True,True,True]", "[False]*5", or "[True,False,True]"

Hyperparameters

params	type	description	example
-K	int	number of metagenes; equivalently, dimension of latent space	15
--initialization_method	str	Supports louvain, kmeans, svd, and precomputed clusters	"louvain"
--initialization_kwargs	dict (Python expression)	A dictionary specifying arguments for the initialization	"dict()"
--lambda_SigmaXInv	float	regularization on Sigma_x^{-1}	1e-6
--max_iter	int	maximum number of coordinate optimization iterations	200
--init_NMF_iter	int	number of NMF iterations for initialization of SpiceMix	10

Reproducibility related parameters

params	type	description	example
--random_seed	int	the initial random seed fed to k-means	0

Output control

params	type	description	example
--result_filename	str	the name of the hdf5 file that stores inferred parameters	"../data/simulation 1/results/SpiceMix.h5"

Examples

This notebook shows how to run SpiceMinx via command line (part 1.1) and through a notebook (part 1.2).

The following command runs SpiceMix on FOVs 1 and 3 from dataset simulation 1 with K=20 metagenes and the spatial information of both FOVs are used:

python main.py -K=20 --path2dataset="path/to/simulation 1" --repli_list="[1,3]" --use_spatial="[True]*2" --result_filename="path/to/simulation 1/results/SpiceMix_K20_FOV13.h5"

To run NMF on FOVs 1 and 3 on the same dataset with identical settings, we can use the following command:

python main.py -K=20 --path2dataset="path/to/simulation 1" --repli_list="[1,3]" --use_spatial="[False]*2" --result_filename="path/to/simulation 1/results/NMF_K20_FOV13.h5"

To specify the files generated by a particular preprocessng approach for dataset simulation 2, we can run

python main.py -K=20 --path2dataset="path/to/simulation 2" --repli_list="[1,2]" --use_spatial="[True]*2" --neighbor_suffix=10NN --expression_suffix=nonzero10 --result_filename="path/to/simulation 2/results/SpiceMix_K20_FOV12_10NN_nonzero10.h5"

and

python main.py -K=20 --path2dataset="path/to/simulation 2" --repli_list="[1,2]" --use_spatial="[False]*2" --neighbor_suffix=10NN --expression_suffix=nonzero10 --result_filename="path/to/simulation 2/results/NMF_K20_FOV12_10NN_nonzero10.h5"

Step 4: Locating results

The output of one SpiceMix run is saved in an HDF5 file in the results directory and its name is specified via the argument to --result_filename. In an HDF5 file, there are four groups and the content is organized in the following structure:

• hyperparameters: Hyperparameters specified for this run. For example,
  • hyperparameters/K: the number of metagenes;
  • hyperparameters/lambda_SigmaXInv: the value of the regularization coefficent on Sigma_x^{-1};
  • hyperparameters/repli_list: the list of replicate names.
• metrics: Criterion of convergence. Currently, only one indicator of convergence is implemented:
  • metrics/Q/{i}: the Q-value after the ith iteration, which is the negative logarithm of the joint probability.
• latent_states: Latent states, currently only including X for each FOV.
  • latent_states/XT/{repli_name}/{i}: an N-by-K matrix containing the latent states in replicate <repli_name> after iteration i.
• parameters: Model parameters:
  • parameters/M/{i}: a G-by-K matrix containing the metagenes after iteration i;
  • parameters/Sigma_x_inv/{i}: a K-by-K matrix containing the affinity matrix Sigma_x^{-1} after iteration i;

Empirical runtime

• On a synthetic dataset, SpiceMix takes 6~30 minutes for 200 iterations, depending on the difficulty.
• On the Visium dataset (Maynard 2021), SpiceMix takes 100 minutes for 200 iterations.

Cite

Cite our paper by

@article{chidester2023spicemix,
  title={SpiceMix enables integrative single-cell spatial modeling of cell identity},
  author={Chidester, Benjamin and Zhou, Tianming and Alam, Shahul and Ma, Jian},
  journal={Nature Genetics},
  pages={1--11},
  year={2023},
  publisher={Nature Publishing Group},
  doi={https://doi.org/10.1038/s41588-022-01256-z},
}

Contact

Please contact tianming@andrew.cmu.edu or raise an issue in the github repo with any questions about installation or usage.