Introduction

A collection of machine learning models that make predictions of cancer molecular subtypes. Users can predict cancer subtypes on their molecular data. Data platforms supported are gene expression, DNA methylation, miRNA, copy number, and/or mutation calls.

These tools were created from the GDAN-TMP group where minimal molecular markers were used to accurately predict 26 different cancer cohorts and 106 subtypes. Of the 1000's of models ran, the models with high predictive accuracy are made available to the public here.

• For the publication of this work, visit: (TBA until after manuscript publication)

• For the publication page, visit: (TBA until after manuscript publication)

Model data downloads from publication page are required to run certain models

Subtype predictions can be made for the following TCGA cohorts:

ACC, BLCA, BRCA, CESC, COADREAD, ESCC, GEA, HNSC, KIRCKICH, KIRP, LGGGBM, LIHCCHOL, LUAD, LUSC, MESO, OV, PAAD, PCPG, PRAD, SARC, SKCM, TGCT, THCA, THYM, UCEC, UVM

Model Library Available

There are five methods (SK Grid, AKLIMATE, CloudForst, JADBio, and subSCOPE) and each ran tens to thousands of models. The top performing models of each method, for each of the 26 cancer cohorts have been made available, and include:

1. Best OVERALL model - highest performing model
2. Best GEXP only model - highest performing model using only gene expression features
3. Best CNVR only model - highest performing model using only copy number features
4. Best MUTA only model - highest performing model using only mutation features
5. Best METH only model - highest performing model using only DNA methylation features
6. Best MIR only model - highest performing model using only miRNA features

There are a few exceptions to models provided by certain methods, see "Additional Info: Model Selection and Input Specifications" section below.

Docker images for each model are pulled automatically in workflow shown in the "Analyze" section below. Docker images are stored in CCG_TMP_Public Synapse Space.

Which Machine Learning Models

We have made publicly available the top models (above section) and any new data can get subtype predictions from these models. Explore these well-performing models by seeing the algorithm name, parameters, and required feature list. The feature lists will be returned in TMP nomenclature.

Example of model information. For more details see: Explore_models.md

{'model': 'sklearn.ensemble.RandomForestClassifier',
 'model_params': {'criterion': 'entropy', 'n_estimators': 200},
 'fts': ['N:GEXP::CENPA:1058:',
  'N:GEXP::FOXC1:2296:',
  'N:GEXP::ESR1:2099:',
  'N:GEXP::MBOAT1:154141:',
  'N:GEXP::MIA:8190:',
  'N:GEXP::ANXA3:306:',
  'N:GEXP::WDR67:93594:',
  'N:GEXP::NAT1:9:',
  'N:GEXP::EXO1:9156:']}

Requirements

The following are required:

• Python 3+ https://www.python.org/downloads/
• Docker or Docker Desktop https://www.docker.com/
• Synapse client https://help.synapse.org/docs/Installing-Synapse-API-Clients.1985249668.html and create an account
• AWS https://docs.aws.amazon.com/cli/latest/userguide/getting-started-install.html

Setup

1. Create an environment and install dependencies. This will install cwlref-runner version 1.0, this version is required to run analysis.

python3 -m venv venv
. venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

2. Synapse Sign In

synapse login --remember-me

3. Docker Sign In (if not already)

docker login

4. Synapse Docker Registry Sign In - using Syanpse username and password

docker login -u <synapse-username> docker.synapse.org

1. Download Data (Method Models and Feature Renaming Reference Files)

1A. Reference files for transform (project matrices)

Download and decompress the reference files that are used as the target data space for data transformations (ex. quantile rescaling).

The TMP_20230209.tar.gz file can be downloaded from the Publication Page and then placed in tools/

cd tools
tar -xzf TMP_20230209.tar.gz
cd ..

1B. Method Models

Certain methods require large or source files to run models. These files are available for download from the Publication page.

Required step: download associated model data for certain methods

CloudForest download of model data: download from the publication page models_cf.tar.gz into the directory cloudforest/data/ and decompress.

JADBio download of model data: download from the publication page models_jadbio.tar.gz into the directory jadbio/data/ and decompress.

SK Grid download of model data: copy over this file from tools

cp -r tools/TMP_20230209 skgrid/data/src/training_data/

AKLIMATE and subSCOPE do not need manual model data download.

1C. Feature Renaming Reference Files

Download and decompress the reference files - renaming any user data feature to nomenclature that machine learning models will recognize (TMP nomenclature).

The ft_name_convert.tar.gz file can be downloaded from the Publication Page and then placed in tools/

cd tools
tar -xzf ft_name_convert.tar.gz
cd ..

Data Requirements

User input data must be in tab separated format. Where original user data has rows labeled with samples and columns labeled with features (ex. genes).

2. Pre-processing User Data

Input data must have proper feature labeling and rescaling prior to running machine learning models for subtype predictions.

2A. Feature Relabeling and Transposing

First, machine learning models need to be able to match genes to GDAN-TMP specific gene IDs (ex. convert gene TP53 to feature N:GEXP::TP53:7157:). Then reformatting to a sample x feature matrix.

python tools/convert.py \
	--data <original-user-data> \
	--out <relabeled-user-data> \
	--cancer <cancer>

An optional argument of --delete_i_col can be included. An optional argument to inform which column to remove (0 based indexing). Use if a meta-data column is in data. If not specified, then will run with no column deletions.

2B. Quantile Rescaling

Second, relabeled data must be transformed with a quantile rescale prior to running machine learning algorithms. The rescaled output file will always be located in user-transformed-data/transformed-data.tsv.

# Transform
bash tools/run_transform.sh \
  <relabeled-user-data> \
	<cancer>

# Handle 10 quantile differences
python tools/zero_floor.py \
  -in user-transformed-data/transformed-data.tsv \
  -out user-transformed-data/transformed-data.tsv

3. Run Machine Learning Models to Predict Cancer Subtypes

Simple command to call one of the five methods. This will predict the molecular subtype for each sample bash RUN_MODEL.sh <arguments>

Available methods are skgrid, aklimate, cloudforest, jadbio, and subscope.

Specifically:

bash RUN_model.sh <cancer> <platform> <method> <your-data>

Examples for BRCA cancer cohorts are:

• bash RUN_model.sh BRCA GEXP cloudforest user-transformed-data/transformed-data.tsv
• bash RUN_MODEL.sh BRCA GEXP aklimate user-transformed-data/transformed-data.tsv
• bash RUN_MODEL.sh BRCA GEXP skgrid user-transformed-data/transformed-data.tsv
• bash RUN_MODEL.sh BRCA GEXP jadbio user-transformed-data/transformed-data.tsv
• bash RUN_MODEL.sh BRCA GEXP subscope user-transformed-data/transformed-data.tsv

Tutorial

An example of how to run the prediction workflow is shown here using SK Grid best performing gene expression model on a breast cancer cBioPortal dataset.

Alternative Model Download (Optional)

Docker images for methods are automatically pulled and built by CWL workflows and tools from the public Synapse repository.

Alternatively:

1. Docker images can be manually downloaded by going to the publication page for each method image file.

# ImageFiles
sk_grid.tar.gz
aklimate.tar.gz
cloudforest.tar.gz
jadbio.tar.gz
subscope.tar.gz

2. Build each method's docker image.

docker load -i <imagefile.tar.gz>

Check these images have been successfully loaded with docker images.

Note that some methods have an additional model data file to run. These can be found at the publication page (see section Download Method Model Data)

Additional Info: Model Selection and Input Specifications

Each the file in user-job-ymls/ is associated with the method and is automatically generated from RUN_MODEL.sh.

Available methods are skgrid, aklimate, cloudforest, jadbio, and subscope.

Each method file is slightly different, but all will require selection of at least the input dataset, cancer model, and data platform. Available values are noted in tools/options.yml.

Model platform name differs for each method, see below:

SK Grid options

• Model options: OVERALL, CNVR, GEXP, METH, MIR, or MUTA
• GEXP best model that uses only mRNA expression (gene expression) for predictions
• CNVR best model that uses only copy number variation for predictions
• METH best model that uses only DNA methylation for predictions
• MIR best model that uses only miRNA for predictions
• MUTA best model that uses only somatic mutations for predictions
• OVERALL best model regardless of platform (can be single data platform type or a combination of platforms).
• Additional method details found in SK Grid README

AKLIMATE options

• Model options: TOP, GEXP, CNVR, METH, or MULTI
• GEXP best model that uses only mRNA expression (gene expression) for predictions
• CNVR best model that uses only copy number variation for predictions
• METH best model that uses only DNA methylation for predictions
• MIR best model that uses only miRNA for predictions
• MUTA best model that uses only somatic mutations for predictions
• TOP best model regardless of platform (can be single data platform type or a combination of platforms). (recommended over MULTI)
• MULTI best model can use all available data types (CNVR, GEXP, METH, MIR, and MUTA)
• Additional method details found in AKLIMATE README

CloudForest options

• Model options: OVERALL, All, CNVR, GEXP, METH, MIR, or MUTA
• GEXP best model that uses only mRNA expression (gene expression) for predictions
• CNVR best model that uses only copy number variation for predictions
• METH best model that uses only DNA methylation for predictions
• MIR best model that uses only miRNA for predictions
• MUTA best model that uses only somatic mutations for predictions
• OVERALL best model regardless of platform (this can be a model that uses a single platform, several platforms, or all platforms; it simply returns the model that had the highest performance). (recommended over MULTI)
• All best model must use all available data types (CNVR, GEXP, METH, MIR, and MUTA)
• Additional method details found in CloudForest README

subSCOPE options

• Model options: allcohorts, CNVR, GEXP, METH, MIR, or MUTA
• GEXP best model that uses only mRNA expression (gene expression) for predictions
• CNVR best model that uses only copy number variation for predictions
• METH best model that uses only DNA methylation for predictions
• MIR best model that uses only miRNA for predictions
• MUTA best model that uses only somatic mutations for predictions
• allcohorts best model that can predict on a mixed cancer cohort (ex. pancreatic tumors can be mixed with breast tumors). Will return the best subtype predictions regardless of cohort
• Additional method details found in subSCOPE README

JADBio options

• Model options: MULTI, CNVR, GEXP, METH, MIR, or MUTA
• GEXP best model that uses only mRNA expression (gene expression) for predictions
• CNVR best model that uses only copy number variation for predictions
• METH best model that uses only DNA methylation for predictions
• MIR best model that uses only miRNA for predictions
• MUTA best model that uses only somatic mutations for predictions
• MULTI best model that must use all available data types
• Additional method details found in JADBio README

Acknowledgment and Funding

We would like to thank the National Cancer Institute for the support.

Troubleshooting

See How to Fix Common Issues for common error messages.

Maintainers

Current maintainers:

• Jordan A. Lee (GitHub jordan2lee)
• Kyle Ellrott (GitHub kellrott)
• Brian Karlberg (GitHub briankarlberg)