toomanycells 1.0.21

A python package for spectral clustering.

too-many-cells (à la Python)

It's Scanpy friendly!

**A python package for spectral clustering based on the powerful suite of tools named too-many-cells. In essence, you can use toomanycells to partition a data set in the form of a matrix of integers or floating point numbers into clusters. The rows represent observations and the columns are the features. Initially, toomanycells will partition your data set into two subsets, trying to maximize the differences between the two. Subsequently, it will reapply that same criterion to each subset and will continue bifurcating until the modularity of the parent becomes negative, implying that the current subset is fairly homogeneous, and consequently suggesting that further partitioning is not warranted. Thus, when the process finishes, you end up with a tree structure of your data set, where the leaves represent the clusters. As mentioned earlier, you can use this tool with any kind of data. However, a common application is to classify cells and therefore you can provide an AnnData object. You can read about this application in this Nature Methods paper. **

• Free software: BSD 3-Clause License
• Documentation: https://JRR3.github.io/toomanycells

Dependencies

Make sure you have installed the graph visualization library Graphviz. For example, if you want to use conda, then do the following.

conda install anaconda::graphviz

Or, if you are using Linux, you can do

sudo apt install libgraphviz-dev

Installation

Just type

pip install toomanycells

in your home environment. If you want to install an updated version, then use the following flag.

pip install toomanycells -U

Make sure you have the latest version. If not, run the previous command again.

Quick run

If you want to see a concrete example of how to use toomanycells, check out the jupyter notebook demo.

Usage

1. First import the module as follows

   from toomanycells import TooManyCells as tmc
   

2. If you already have an AnnData object A loaded into memory, then you can create a TooManyCells object with

   tmc_obj = tmc(A)
   

   However, if you want the output folder to be a directory that is not the current working directory, then you can specify the path as follows

   tmc_obj = tmc(A, output_directory)
   

3. If instead of providing an AnnData object you want to provide the directory where your data is located, you can use the syntax

   tmc_obj = tmc(input_directory, output_directory)
   

4. If your input directory has a file in the matrix market format, then you have to specify this information by using the following flag

   tmc_obj = tmc(input_directory, output_directory, input_is_matrix_market=True)
   

   Under this scenario, the input_directory must contain a .mtx file, a barcodes.tsv file (the observations), and a genes.tsv (the features).

5. Once your data has been loaded successfully, you can start the clustering process with the following command

   tmc_obj.run_spectral_clustering()
   

   In my desktop computer processing a data set with ~90K cells (observations) and ~30K genes (features) took a little less than 6 minutes in 1809 iterations. For a larger data set like the Tabula Sapiens with 483,152 cells and 58,870 genes (14.51 GB in zip format) the total time was about 50 minutes in the same computer.

6. At the end of the clustering process the .obs data frame of the AnnData object should have two columns named ['sp_cluster', 'sp_path'] which contain the cluster labels and the path from the root node to the leaf node, respectively.

   tmc_obj.A.obs[['sp_cluster', 'sp_path']]
   

7. To generate the outputs, just call the function

   tmc_obj.store_outputs()
   

   This call will generate a PDF of the tree and a DOT file for the graph, two CSV files that describe the clusters and the information of each node, and a JSON file that contains the tree structure. If you already have a DOT file and only want to plot the tree and store the information of each node, you can use the following call

   tmc_obj.store_outputs(load_dot_file=True)
   

8. If you want to visualize your results in a dynamic platform, I strongly recommend the tool too-many-cells-interactive. To use it, first make sure that you have Docker Compose and Docker. One simple way of getting the two is by installing Docker Desktop. If you use Nix, simply add the packages pkgs.docker and pkgs.docker-compose to your configuration or home.nix file and run

home-manager switch

9. If you installed Docker Desktop you probably don't need to follow this step. However, under some distributions the following two commands have proven to be essential.

sudo dockerd

to start the daemon service for docker containers and

sudo chmod 666 /var/run/docker.sock

to let Docker read and write to that location.

10. Now clone the repository

git clone https://github.com/schwartzlab-methods/too-many-cells-interactive.git

and store the path to the too-many-cells-interactive folder in a variable, for example path_to_tmc_interactive. Also, you will need to identify a column in your AnnData.obs data frame that has the labels for the cells. Let's assume that the column name is stored in the variable cell_annotations. Lastly, you can provide a port number to host your visualization, for instance port_id=1234. Then, you can call the function

tmc_obj.visualize_with_tmc_interactive(
         path_to_tmc_interactive,
         cell_annotations,
         port_id)

The following visualization corresponds to the data set with ~90K cells (observations).

And this is the visualization for the Tabula Sapiens data set with ~480K cells.

What is the time complexity of toomanycells (à la Python)?

To answer that question we have created the following benchmark. We tested the performance of toomanycells in 20 data sets having the following number of cells: 6360, 10479, 12751, 16363, 23973, 32735, 35442, 40784, 48410, 53046, 57621, 62941, 68885, 76019, 81449, 87833, 94543, 101234, 107809, 483152. The range goes from thousands of cells to almost half a million cells. These are the results. As you can see, the program behaves linearly with respect to the size of the input. In other words, the observations fit the model $T = k\cdot N^p$, where $T$ is the time to process the data set, $N$ is the number of cells, $k$ is a constant, and $p$ is the exponent. In our case $p\approx 1$. Nice!

Similarity functions

So far we have assumed that the similarity matrix $S$ is computed by calculating the cosine of the angle between each observation. Concretely, if the matrix of observations is $B$ ($m\times n$), the $i$-th row of $B$ is $x = B(i,:)$ and the $j$-th row of $B$ is $y=B(j,:)$, then $$S(i,j)=\frac{x\cdot y}{||x||_2\cdot ||y||_2}$$ However, this is not the only similarity function. We will list all the available similarity functions and how to call them.

Cosine (sparse)

If your matrix is sparse, i.e., the number of nonzero entries is proportional to the number of samples ($m$), and you want to use the cosine similarity, then use the following instruction.

tmc_obj.run_spectral_clustering(
   similarity_function="cosine_sparse")

By default we use the Halko-Martinsson-Tropp algorithm to compute the truncated singular value decomposition. However, the ARPACK library (written in Fortran) is also available.

tmc_obj.run_spectral_clustering(
   similarity_function="cosine_sparse",
   svd_algorithm="arpack")

If $B$ has negative entries, it is possible to get negative entries for $S$. This could in turn produce negative row sums for $S$. If that is the case, the convergence to a solution could be extremely slow. However, if you use the non-sparse version of this function, we provide a reasonable solution to this problem.

Cosine

If your matrix is dense, and you want to use the cosine similarity, then use the following instruction.

tmc_obj.run_spectral_clustering(
   similarity_function="cosine")

The same comment about negative entries applies here. However, there is a simple solution. While shifting the matrix of observations can drastically change the interpretation of the data, shifting the similarity matrix is actually a reasonable method to remove negative entries. The reason is that shifting by a constant is an order-preserving transformation, which implies that if the similarity between $x$ and $y$ is less than the similarity between $u$ and $w$, i.e., $S(x,y)<S(u,w)$, then $S(x,y)+1 < S(u,w)+1$. Samples have no natural order, but similarities do. To shift the (dense) similarity matrix by $s=1$, use the following instruction.

tmc_obj.run_spectral_clustering(
   similarity_function="cosine",
   shift_similarity_matrix=1)

Note that since the range of the cosine similarity is $[-1,1]$, the shifted range with $s=1$ becomes $[0,2]$.

Laplacian

The similarity matrix is $$S(x,y)=\exp(-\gamma\cdot ||x-y||_1)$$ This is an example:

tmc_obj.run_spectral_clustering(
   similarity_function="laplacian",
   similarity_gamma=0.01)

This function is very sensitive to $\gamma$. Hence, an inadequate choice can result in poor results or no convergence. If you obtain poor results, try using
a smaller value for $\gamma$.

Gaussian

The similarity matrix is $$S(x,y)=\exp(-\gamma\cdot ||x-y||_2^2)$$ This is an example:

tmc_obj.run_spectral_clustering(
   similarity_function="gaussian",
   similarity_gamma=0.001)

As before, this function is very sensitive to $\gamma$. Note that the norm is squared. Thus, it transforms big differences between $x$ and $y$ into very small quantities.

Divide by the sum

The similarity matrix is $$S(x,y)=1-\frac{||x-y||_p}{||x||_p+||y||_p},$$ where $p =1$ or $p=2$. The rows of the matrix are normalized (unit norm) before computing the similarity. This is an example:

tmc_obj.run_spectral_clustering(
   similarity_function="div_by_sum")

Additional features

TF-IDF

If you want to use the inverse document frequency (IDF) normalization, then use

tmc_obj.run_spectral_clustering(
   similarity_function="some_sim_function",
   use_tf_idf=True)

If you also want to normalize the frequencies to unit norm, then use

tmc_obj.run_spectral_clustering(
   similarity_function="some_sim_function",
   use_tf_idf=True,
   tf_idf_norm="l2")

to use the $2$-norm, or

tmc_obj.run_spectral_clustering(
   similarity_function="some_sim_function",
   use_tf_idf=True,
   tf_idf_norm="l1")

to use the $1$-norm.

Normalization

Sometimes normalizing your matrix of observations can improve the performance of some routines. To normalize the row, use the following instruction.

tmc_obj.run_spectral_clustering(
   similarity_function="some_sim_function",
   normalize_rows=True)

Be default, the $2$-norm is used. To use any other $p$-norm, use

tmc_obj.run_spectral_clustering(
   similarity_function="some_sim_function",
   normalize_rows=True,
   similarity_norm=p)

Acknowledgments

I would like to thank the Schwartz lab (GW) for letting me explore different directions and also Christie Lau for providing multiple test cases to improve this implementation.