Single-cell analysis using Multiple Kernel Learning, scMKL, is a binary classification algorithm utilizing prior information to group features to enhance classification and aid understanding of distinguishing features in omic and multi-omic data sets.

We have demonstrated scMKL's ability to achieve high classification performance while providing the added confidence of model interpretability on single-cell RNA, ATAC, ADT, and methylation data.

Installation

Conda install

Conda is the recommended method to install scMKL and Python version must be at least 3.11 but less than 3.14:

conda create -n scmkl_env -c conda-forge -c bioconda ivango17::scmkl

Ensure bioconda and conda-forge are in your available conda channels.

Pip install

First, create a virtual environment with 'python>=3.11.1,<3.14'.

Then, install scMKL with:

# activate your new env with python>=3.11.1 and <3.14
pip install scmkl

If wheels do not build correctly, ensure gcc and g++ are installed and up to date. They can be installed with sudo apt install gcc and sudo apt install g++.

Usage

The table below shows what type of data scMKL expects based on the modality in question. In all situations, matrices should be cells x features.

Additionally, scMKL requires a feature grouping dictionary where each key, value pair is represented as 'group1': ['feature1', 'feature2', 'feature7']. Features can be genes for RNA, regions for ATAC and MET, or proteins for ADT. Any number of groups can be used and features can be overlapping between groups (e.g. 'feature1' is in five groups). For help getting feature groupings, see our notebooks in examples.

scMKL implements AnnData.anndata objects in one of two ways:

1. Taking an existing AnnData.anndata object and formatting it for scMKL to train and test on.

import scmkl
import numpy as np
import anndata as ad

# Read in your AnnData.anndata obj
adata = ad.read_h5ad('your_adata.h5ad')

# Read in feature grouping dictionary (e.g. geneset library for RNA)
group_dict = np.load('your_grouping.pkl', allow_pickle=True)
# or for RNA, pull a gene set from the internet with 
group_dict = scmkl.get_gene_groupings('Azimuth_2021', organism='human')

# Apply scmkl formatting where 'phenotype_obs_key' is the col name in obs 
# for labels, can also be an array of labels and set `allow_multiclass` to 
# true if there are more than two cell classes 
adata = scmkl.format_adata(
    adata, 
    cell_labels='phenotype_obs_key', 
    group_dict=group_dict,
    allow_multiclass=True
    )

2. Reading in data separately and creating an scMKL formatted AnnData.anndata object.

import scmkl
import numpy as np

# Read in feature names
var_names = np.load('your_feature_names.npy')

# Read in feature grouping dictionary (e.g. geneset library for RNA)
group_dict = np.load('your_grouping.pkl', allow_pickle=True)
# or for RNA, pull a gene set from the internet with
group_dict = scmkl.get_gene_groupings('Azimuth_2021', organism='human')

# Read in cell labels
obs = np.load('your_cell_labels.npy')

# Read in data matrix (can also be a scipy.sparse matrix)
mat = np.load('your_matrix.npy')

# Create scMKL formatted AnnData.anndata and set `allow_multiclass` to true 
# if there are more than two cell classes 
adata = scmkl.create_adata(
    data_mat, 
    feature_names=gene_names, 
    group_dict=group_dict,
    allow_multiclass=True
    )

If you are not using RNA data and do not have a grouping dictionary but want to try scMKL on your dataset, you can create a random grouping with:

# Assuming feature names and numpy are loaded
n_groups = 50
group_size = 25
grouping_dict = {f'group_{i}': np.random.choice(
                                   var_names, 
                                   size=group_size, 
                                   replace=False
                                   ) 
                    for i in range(n_groups)}

Then, depending on whether or not your labels are binary, train and test your model.

For binary:

results = scmkl.run(adata)

For multiclass:

results = scmkl.one_v_rest(
    adata, 
    names=['RNA']
    )

Both of these functions return a dictionary with evaluation metrics, group weights, etc... To learn more about accessing this data, see our GitHub Pages.

Links

Repo: https://github.com/ohsu-cedar-comp-hub/scMKL

PyPI: https://pypi.org/project/scmkl/

Anaconda: https://anaconda.org/ivango17/scmkl

API: https://ohsu-cedar-comp-hub.github.io/scMKL/

Publication

If you use scMKL in your research, please cite using:

Kupp, S., VanGordon, I., Gönen, M., Esener, S., Eksi, S., Ak, C. Interpretable and integrative analysis of single-cell multiomics with scMKL. Commun Biol 8, 1160 (2025). https://doi.org/10.1038/s42003-025-08533-7

Our Shiny for Python application for viewing data produced from this work can be found here: scMKL_analysis

Issues

Please report bugs here.