mt-SCITE infers trees from a matrix of mutation probabilities built from a statistical model of alternate read counts in mitochondrial sequencing reads.
clang++ src/*.cpp -o mtscite
mtscite -i <mut_probabilities> -n <n_sites> -m <n_cells> -l <n_iters> -seed <random_seed>
We also provide notebooks and scripts to perform an end-to-end tree inference analysis using mt-SCITE.
conda create -n mtSCITE python=3.10
conda activate mtSCITE
conda install -c conda-forge biopython
conda install -c anaconda pandas
conda install -c anaconda scipy
conda install -c conda-forge matplotlib
conda install -c anaconda seaborn
conda install -c anaconda graphviz
pip install --global-option=build_ext --global-option="-I~/miniconda3/envs/mtSCITE/include" --global-option="-L~/miniconda3/envs/mtSCITE/include/lib/" pygraphviz
conda install -c anaconda python-graphviz
conda install -c anaconda pydot
conda install -c anaconda networkx
conda install -c anaconda numpy=1.22
Raw sequencing data was processed by the snakemake pipeline in preprocessing_pipeline/Snakefile. This generated tsv files specifying the number of reads supporting the four different nucleotides A, C, G, T and total number of reads for each position in the mitochondrial genome for each sample.
compute_mutations_probabilities.ipynb was run to generate mutation probability matrices for a range of error rates. The mutation probability matrices were stored in /path/to/matrices/
We used a 3-fold cross validation procedure to select the error rate that we used to infer the trees using mt-SCITE. After generating probability matrices for different error rates and storing them in /path/to/matrices/, you can perform cross-validation by running this command:
python scripts/cv.py </path/to/matrices/> </path/to/mtscite>
This will create a CSV file named val_scores.csv with the 3-fold cross-validation likelihood scores.
learn_error_rate.ipynb was run to analyze val_scores.csv and to select the error rate to be used for tree building.
Run mt-SCITE with this command:
`</path/to/mtscite> -i pmat.csv -n <n_mutations> -m <n_samples> -r 1 -l 200000 -fd 0.0001 -ad 0.0001 -cc 0.0 -s -a -o </path/to/output/run_id>`
where
pmat.csv is the mutation probability matrix generated with the learned error rate, n_mutations is the number of rows in pmat.csv and n_samples is the number of columns in pmat.csv.
We advise running mt-SCITE several times (in our experimental analyses, we used 10 repetitions). Each repetition starts mt-SCITE from a different starting point, or you can directly specify different starting seeds using -seed. Then, check that the best-scoring region is recovered by all chains. If the chains land in very different log-likelihood areas, increase the chain length (we used 10^6). Because the bottleneck is the number of mutations rather than cells, the search budget should be scaled with mutation count.