Metient

Metient (METastasis + gradiENT) is a tool for inferring the metastatic migrations of a patient's cancer. You can find our preprint on bioRxiv.

Installation

Metient is available as a python library, installable via pip. It has been tested on Linux and Apple M1 Pro.

# mamba used for speed, can use conda instead if mamba is not installed
mamba create -n "met" python=3.8.8 ipython
mamba activate met
pip install metient 

Tip

If pip install metient fails due to fatal error: graphviz/cgraph.h: No such file or directory, you need to set environment variables to point to your graphviz header files. Locate the path to your mamba environment (for e.g. using mamba env list), and run the following:

export CFLAGS="-I/path/to/mamba/env/include"
export LDFLAGS="-L/path/to/mamba/env/lib"
pip install pygraphviz --user
pip install metient

Tutorial

To run the tutorial notebooks, clone this repo:

git clone [email protected]:morrislab/metient.git
cd metient/tutorial/

There are different Jupyter Notebook tutorials based on your use case:

1. I have a cohort of patients (~5 or more patients) with the same cancer type. (Metient-calibrate)
   • I want Metient to estimate which mutations/mutation clusters are present in which anatomical sites. Tutorial 1
   • I know which mutations/mutation clusters are present in which anatomical sites. Tutorial 2
2. I have a small number of patients, or I want to enforce my own parsimony metric weights. (Metient-evaluate)
   • I want Metient to estimate which mutations/mutation clusters are present in which anatomical sites. Tutorial 3
   • I know which mutations/mutation clusters are present in which anatomical sites. Tutorial 4

Tip

If your jupyter notebook does not automatically recognize your conda environment, run the following:

pip install ipykernel
python -m ipykernel install --user --name myenv --display-name "met"

Then in the jupyter notebook, select Kernel > Change kernel > met.

Inputs

There are two required inputs, a tsv file with information for each sample and mutation/mutation cluster, and a txt file specifying the edges of the clone tree.

1. Tsv file

There are two types of tsvs that are accepted, depending on if you'd like Metient to estimate the presence of cancer clones in each tumor site (1a), or if you'd like to input this yourself (1b).

1a. If you would like Metient to estimate the prevalance of each cancer clone in each tumor site, use the following input tsv format.

1a example tsv

Each row in this tsv should correspond to the reference and variant read counts at a single locus in a single tumor sample:

Column name	Description
anatomical_site_index	Zero-based index for anatomical_site_label column. Rows with the same anatomical site index and cluster_index will get pooled together.
anatomical_site_label	Name of the anatomical site
character_index	Zero-based index for character_label column
character_label	Name of the mutation. This is used in visualizations, so it should be short. NOTE: due to graphing dependencies, this string cannot contain colons.
cluster_index	If using a clustering method, the cluster index that this mutation belongs to. NOTE: this must correspond to the indices used in the tree txt file. Rows with the same anatomical site index and cluster_index will get pooled together.
ref	The number of reads that map to the reference allele for this mutation or mutation cluster in this anatomical site.
var	The number of reads that map to the variant allele for this mutation or mutation cluster in this anatomical site.
site_category	Must be one of primary or metastasis. If multiple primaries are specified, such that the primary label is used for multiple different anatomical site indices (i.e., the true primary is not known), we will run Metient multiple times with each primary used as the true primary. Output files are saved with the suffix _{anatomical_site_label} to indicate which primary was used in that run.
var_read_prob	This gives Metient the ability to correct for the effect copy number alterations (CNAs) have on the relationship between variant allele frequency (VAF, i.e., the proportion of alleles bearing the mutation) and subclonal frequency (i.e., the proportion of cells bearing the mutation). Let j = character_index. var_read_prob is the probabilty of observing a read from the variant allele for mutation at j in a cell bearing the mutation. Thus, if mutation at j occurred at a diploid locus with no CNAs, this should be 0.5. In a haploid cell (e.g., male sex chromosome) with no CNAs, this should be 1.0. If a CNA duplicated the reference allele in the lineage bearing mutation j prior to j occurring, there will be two reference alleles and a single variant allele in all cells bearing j, such that var_read_prob = 0.3333. If using a CN caller that reports major and minor CN: var_read_prob = (p*maj)/(p*(maj+min)+(1-p)*2), where p is tumor purity, maj is major CN, min is minor CN, and we're assuming the variant allele has major CN. For more information, see S2.2 of PairTree's supplementary info for more details.

1b. If you would like to input the prevalence of each cancer clone in each tumor site, use the following input tsv format.

1b example tsv

Each row in this tsv should correspond to a single mutation/mutation cluster in a single tumor sample:

Column name	Description
anatomical_site_index	Zero-based index for anatomical_site_label column. Rows with the same anatomical site index and cluster_index will get pooled together.
anatomical_site_label	Name of the anatomical site
cluster_index	If using a clustering method, the cluster index that this mutation belongs to. NOTE: this must correspond to the indices used in the tree txt file. Rows with the same anatomical site index and cluster_index will get pooled together.
cluster_label	Name of the mutation or cluster of mutations. This is used in visualizations, so it should be short. NOTE: due to graphing dependencies, this string cannot contain colons.
present	Must be one of 0 or 1. 1 indicates that this mutation/mutation cluster is present in this anatomical site, and 0 indicates that it is not.
site_category	Must be one of primary or metastasis. If multiple primaries are specified, such that the primary label is used for multiple different anatomical site indices (i.e., the true primary is not known), we will run Metient multiple times with each primary used as the true primary. Output files are saved with the suffix _{anatomical_site_label} to indicate which primary was used in that run.
num_mutations	The number of mutations in this cluster.

2. Tree txt file

A .txt file where each line is an edge from the first index to the second index. Must correspond to the cluster_index column in the input tsv.

Example tree .txt file

Outputs

Metient will output a pickle file in the specificed output directory for each patient that is inputted.

In the pickle file you'll find the following keys:

Pkl key name	Description
ordered_anatomical_sites	a list of anatomical sites in the order used for the matrices detailed below.
node_info	list of dictionaries, in order from best to worst solution. This is solution specific because reolving polytomies can change the tree. Each dictionary maps node index (as used for the matrices detailed below) to a tuple: (label, is_leaf, is_polytomy_resolver_node) used on the tree. The reason labels can be different from what is inputted into Metient is that Metient adds leaf nodes which correspond to the inferred presence of each node in anatomical sites. Each leaf node is labeled as <parent_node_name>_<anatomical_site>.
clone_tree_labeling_matrices	list of numpy ndarrays, in order from best to worst solution. Each numpy array is a matrix (shape: len(ordered_anatomical_sites), len(node_info[x])), where x is the x best solution. Row i corresponds to the site at index i in ordered_anatomical_sites, and column j corresponds to the node with label node_info[x][j][0]. Each column is a one-hot vector representing the location inferred by Metient for that node.
full_adjacency_matrices	list of numpy ndarrays, in order from best to worst tree. Each tensor is a matrix (shape: len(node_info[x]), len(node_info[x])), where x is the x best solution. A 1 at index i,j indicates an edge from i to j.
observed_clone_proportion_matrix	numpy ndarray (shape: len(ordered_anatomical_sites), num_clusters). Row i corresponds to the site at index i in ordered_anatomical_sites, and column j corresponds to the node with label node_info[x][j][0]. A value at i,j greater than 0.05 indicates that that node is present in that antomical site. These are the nodes that get added as leaf nodes.
losses	a list of the losses, from best to worst solution.
primary_site	str, the name of the anatomical site used as the primary site.