Sample-specific Bayesian Networks

A framework for estimating the structures and parameters of Bayesian networks (DAGs) at per-sample or per-patient resolution, formally dubbed NOTMAD (NO-TEARS Mixtures of Archetypal DAGs).

Based on the manuscript NOTMAD: Estimating Bayesian Networks with Sample-Specific Structures and Parameters.

Implemented by Dr. Ben Lengerich (MIT) and Caleb Ellington (CMU)

Install and Use NOTMAD

pip install git+https://github.com/cnellington/SampleSpecificDAGs.git

Then in your Python notebook/script use

from notmad.notmad import NOTMAD

Load your Context data C and Target data X, specify hyperparameters, and train the model

C, X = your_data_loader()

sample_specific_loss_params = {'l1': 0., 'alpha': 2e1, 'rho': 1e0}
archetype_loss_params = {'l1': 0., 'alpha': 1e-1, 'rho': 1e-2}

model = NOTMAD(C.shape, X.shape, k_archetypes, 
                sample_specific_loss_params, archetype_loss_params)
model.fit(C, X, batch_size=1, epochs=50)

Then use it to estmate samples-specific networks! Simple as that.

ss_networks = model.predict_w(C_unseen)

Reproduce Experiments

First, clone the repo where you wish to run the experiments (We recommend somewhere that has a GPU and can run Jupyter notebooks)

Simulations

Run one round of simulations with run_experiments.sh to compare population-based, cluster-based, and sample-specific network inference. Results will be under experiments/simulations/results/ by default.

Real Data

To estimate single-cell regulatory networks using SNARE-seq and NOTMAD and reproduce our figures, run the notebook at experiments/SNAREseq_demo/single_cell_networks.ipynb