MBE: Model-based enrichment estimation and prediction for differential sequencing data

This repo provides a template for implementing the model-based enrichment (MBE) approach from the paper

A. Busia and J. Listgarten. MBE: Model-based enrichment estimation and prediction for differential sequencing data. Genome Biology, 2023.

using standard Python machine learning (ML) libraries.

Note that, because MBE is classifier-based, it can be implemented using any standard ML package that implements probabilistic classifiers. This package provides a standard MBEModel compatible with models implemented by several standard Python libraries, and its usage is demonstrated with scikit-learn in sklearn_example.py which run a simple MBE estimation analysis given a CSV of sequences and sequencing counts. This package also provides an adaptation of the base MBEModel compatible with any ML classifier implemented using TensorFlow, which (1) allows MBE to make use of modern-day neural network models (demonstrated in tf_example.py) and (2) provides a template for adapting MBEModel to a specific ML library of your choice. For specific examples of some TensorFlow classifier architectures, see the source code from the paper: https://github.com/apbusia/selection_dre

Training a model for MBE

MBE can be implemented using any probabilistic classifier, classifier_model, of choice. The base MBEModel can be used to train any underlying classifier_model with fit and predict_log_proba methods, such as scikit-learn models:

    from model_based_enrichment import model_based_enrichment as mbe
    
    mbe_model = mbe.MBEModel(classifier_model)
    # sequences is a list of the N unique observed read sequences in the sequencing dataset.
    # counts is a (n_unique_sequences, n_conditions) matrix of observed read counts.
    mbe_model.fit(sequences, counts)

This workflow can easily be adapted to run MBE with different classifier_model models using MBEModel as a base class; see e.g. MBETensorflowModel.

Predicting / estimating log-enrichment

Once the underlying probabilistic classifier has been trained, MBE can be used to predict the log-enrichment of unobserved sequences and estimate the log-enrichment of observed sequences:

    predictions = mbe_model.log_enrichment(sequences) # sequences is a list of sequences to evaluate.