Nsforest

Documentation: https://nsforest.readthedocs.io/en/latest/

BMC Methods Link: https://bmcmethods.biomedcentral.com/articles/10.1186/s44330-024-00015-2

Download and installation

In terminal:

git clone https://github.com/JCVenterInstitute/NSForest.git

cd NSForest

conda env create -f nsforest.yml

conda activate nsforest

pip install .


Tutorial

Follow the on readthedocs: https://nsforest.readthedocs.io/en/latest/tutorial.html

Pipeline

<img src="pipeline.PNG">

NS-Forest is an algorithm designed to identify minimum combinations of necessary and sufficient marker genes for a cell type cluster identified in a single cell or single nucleus RNA sequencing experiment that optimizes classification accuracy. NS-Forest proceeds through the following steps (default setting):

1. Data input: An AnnData object (e.g., .h5ad file) with cell type cluster labels.

2. Binary score calculation: Each gene is assigned a binary score for every cluster. Binary score is a measurement of the binary expression pattern of a gene. A higher binary score means a gene is expressed in one cluster and not others. A lower binary score means a gene is expressed in many clusters and would not be an ideal candidate for a cell type-specific marker gene.

3. Binary scoring criterion: NS-Forest then filters for genes with high binary scores. Candidate genes are selected if their binary scores are 2 standard deviations above the mean of all genes expressed in the cluster.

4. Random forest: The top 15 binary score genes are used as input into a random forest classifier, which ranks the genes by Gini Impurity, while producing a classification model for each cluster.

5. Decision tree evaluation: The top 6 ranked random forest genes are used as input into decision trees where all combinations of input genes are evaluated and the combination with the highest F-beta score is selected.

6. Output: The NS-Forest algorithm outputs 1-6 marker genes per cluster along with the classification metrics (F-beta, PPV (precision), recall) and the On-Target Fraction expression metric.

NS-Forest Marker Gene Evaluation

The final module in the NS-Forest algorithm can also be used to assess the performance of any collection of marker gene combinations identified using any approach. The marker gene evaluation module includes the following steps (default setting):

1. Data input: 1) An AnnData object (e.g., .h5ad file) with cell type cluster labels. 2) A list of marker genes for every cluster to be evaluated.

2. Decision tree creation: One-vs-all decision trees are created for each gene in the cluster combination and evaluated for classification accuracy.

3. Decision tree evaluation: Each gene in the cluster combination is evaluated using these decision trees to determine if the gene gives the correct classification. If even one gene in the cluster combination gives a misclassification, then the prediction is considered incorrect. Note: This strict criteria may lead to PPV = 0 when no true positives (TP) classification are obtained.

4. Output: The NS-Forest marker gene evaluation outputs the classification metrics (F-beta, PPV (precision), recall) and On-Target Fraction for every cluster combination, which can be used to compare against other marker gene lists.


Prerequisites
* This is a python script written and tested in python 3.11, scanpy 1.9.6.
* Other required libraries: numpy, pandas, sklearn, plotly, time, tqdm.

Versions and citations

Earlier versions are managed in [Releases](https://github.com/JCVenterInstitute/NSForest/releases).

Follow the [tutorial](https://jcventerinstitute.github.io/celligrate/tutorials/NS-Forest_tutorial.html) to get started.

Download 'NSForest_v4dot0_dev.py' and replace the version in the tutorial. Sample code below.


adata_median = preprocessing_medians(adata, cluster_header)
adata_median.varm["medians_" + cluster_header].stack().plot.hist(bins=30, title = 'cluster medians')

adata_median_binary = preprocessing_binary(adata_median, cluster_header, "medians_" + cluster_header)
adata_median_binary.varm["binary_scores_" + cluster_header].stack().plot.hist(bins=30, title='binary scores')

make a copy of prepared adata
adata_prep = adata_median_binary.copy()

NSForest(adata_prep, cluster_header=cluster_header, n_trees=1000, n_genes_eval=6,
medians_header = "medians_" + cluster_header, binary_scores_header = "binary_scores_" + cluster_header,
gene_selection = "BinaryFirst_high", outputfilename="BinaryFirst_high")


**Full Changelog**: https://github.com/JCVenterInstitute/NSForest/compare/v3.9...v4.0_dev

*[Release Note:] Pre-release of NS-Forest v4.0.*

*[Release Note:] Major code optimizations based on algorithm v3.0. No algorithmic change to v3.0.*

NS_Forest(adata, clusterLabelcolumnHeader = "louvain", rfTrees = 1000, Median_Expression_Level = 0, Genes_to_testing = 6, betaValue = 0.5)
* adata = scanpy object
* rfTrees = Number of trees
* clusterLabelcolumnHeader = column header in adata.obs['header_here!'] where cluster assignments reside. Typically 'louvain' if louvain clustering was used.
* Median_Expression_Level = median expression level for removing negative markers
* Genes_to_testing = How many ranked genes by binary score will be evaluated in permutations by fbeta-score
* betaValue = Set values for fbeta weighting. 1 is default f-measure. close to zero is Precision, greater than 1 weights toward Recall


Description

Necessary and Sufficient Forest is a method that takes cluster results from single cell/nuclei RNAseq experiments
and generates lists of minimal markers needed to define each “cell type cluster”.

The method begins by re-encoding the cluster labels into binary classifications, and Random Forest models are generated comparing each
cluster versus all. The top fifteen genes are then reranked using a score measuring how binary they are, e.g., a gene with expression in
the target cluster but no expression in the other clusters would have a high binary score. Expression cutoffs for the top six genes ranked
by binary score are then determined by generating individual decision trees and extracting the decision path information. Then all combinations
of the top six most binary genes are evaluated using f-beta score as an objective function (the beta value default set at 0.5, which weights the
f-measure score more toward precision as opposed to recall).


See code for detailed comments.


Versioning

This is version 3.0 The earlier releases were described in the below publications.

Version 2

Aevermann BD, Zhang Y, Novotny M, Keshk M, Bakken TE, Miller JA, Hodge RD, Lelieveldt B, Lein ES, Scheuermann RH. A machine learning method for the discovery of minimum marker gene combinations for cell-type identification from single-cell RNA sequencing. Genome Res. 2021 Jun 4:gr.275569.121. doi: 10.1101/gr.275569.121. Epub ahead of print. PMID: 34088715.

Necessary and Sufficient Forest (NS-Forest) for Cell Type Marker Determination from cell type clusters

Getting Started

Install Jupyter notebook and python 2.7

Prerequisites

* The script is a Jupyter notebook in python 2.7. Required libraries: Numpy, Pandas, Sklearn, graphviz, numexpr
* The input data is a tab delimited expression Cell x Gene matrix with one column containing the cluster assignments
* The cluster-label column must be named "Clusters" and the labels must be non-numeric (if currently numbers, please add "Cl" or any text would work).
* The gene identifiers used must avoid special characters such as ./-/ or beginning with numbers (I prefix identifiers beginning with numbers and substitute all special characters with "_")


Description

Necessary and Sufficient Forest is a method that takes cluster results from single cell/nuclei RNAseq experiments
and generates lists of minimal markers needed to define each “cell type cluster”.

The method begins by re-encoding the cluster labels into binary classifications, and Random Forest models are generated comparing each
cluster versus all. The top fifteen genes are then reranked using a score measuring how binary they are, e.g., a gene with expression in
the target cluster but no expression in the other clusters would have a high binary score. Expression cutoffs for the top six genes ranked
by binary score are then determined by generating individual decision trees and extracting the decision path information. Then all permutations
of the top six most binary genes are evaluated using f-beta score as an objective function (the beta value default set at 0.5, which weights the
f-measure score more toward precision as opposed to recall).



See code for detailed comments.


Versioning

This is version 2.0 The initial release was version 1.3. Version 1.0 was described in:

Aevermann BD, Novotny M, Bakken T, Miller JA, Diehl AD, Osumi-Sutherland D, Lasken RS, Lein ES, Scheuermann RH.
Cell type discovery using single-cell transcriptomics: implications for ontological representation.
Hum Mol Genet. 2018 May 1;27(R1):R40-R47. doi: 10.1093/hmg/ddy100.


Authors

* Brian Aevermann baevermajcvi.org and Richard Scheuermann RScheuermannjcvi.org


License

This project is licensed under the MIT License - see the [LICENSE.md](LICENSE.md) file for details

Acknowledgments

* Allen Institute of Brain Science
* Chan Zuckerberg Initiative
* California Institute for Regenerative Medicine