alma-classifier 0.1.0

A Python package for applying pre-trained epigenomic classification models

ALMA Classifier

A Python package for applying pre-trained epigenomic classification models to methylation data. This package provides three main predictive models:

1. ALMA Subtype: Predicts 28 subtypes/classes (27 WHO 2022 subtypes of acute leukemia + otherwise-normal control).
2. AML Epigenomic Risk: Predicts the probability of death within 5 years for AML patients using ALMA.
3. 38CpG AML Signature: Predicts same as above, but using a targeted panel of 38 CpGs.

Installation

# first make sure you are running python3.8
python --version

# create a virtual env
python -m venv .venv
source .venv/bin/activate

# install pacmap==0.7.0 (required)
pip install pacmap==0.7.0

# Then install alma-classifier
pip install alma-classifier

# Download model files (required)
python -m alma_classifier.download_models

Important Notes

1. This is a pre-release research tool. Initial versions will be picky and annoying to deal with. Future versions will be flexible and easy to use.
2. Our diagnostic model currently does not know about important cases, which we really need training data for: AML with Down Syndrome, juvenile myelomonocytic leukemia, transient abnormal myelopoiesis, bone marrow failures, low-risk MDS, lymphomas, and others.
3. Our prognostic models (AML Epigenomic Risk and 38CpG AML Signature) are currently only limited to AML cases.

Usage

Command Line Interface

alma-classifier --input path/to/data.pkl --output path/to/predictions.xlsx

Input Data Format

The input data should be a matrix of methylation beta values with:

• Rows representing samples
• Columns representing CpG sites
• Values between 0 and 1

Model Outputs and Prediction Behavior

The package generates predictions with the following behavior:

1. ALMA Subtype Classification:

   • Outputs the predicted subtype and its probability
   • If confidence is below threshold (default 0.5), outputs "Not confident"
   • For predictions with confidence between 0.5-0.8, also outputs the second most likely subtype and its probability

2. AML Epigenomic Risk (only for AML/MDS samples):

   • Outputs "High" or "Low" risk prediction
   • Includes P(Death) at 5y probability
   • Outputs "Not confident" if prediction confidence is below threshold
   • Non-AML/MDS samples receive "NaN" values

3. 38CpG AML Signature (only for AML/MDS samples):

   • Outputs continuous hazard score (38CpG-HazardScore)
   • Provides binary risk stratification (38CpG-AMLsignature: High/Low)
   • Non-AML/MDS samples receive "NaN" values

Citation

If you use this package in your research, please cite:

Francisco Marchi, Marieke Landwehr, Ann-Kathrin Schade et al. Long-read epigenomic diagnosis and prognosis of Acute Myeloid Leukemia, 12 December 2024, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-5450972/v1]

License

See license file.