mahdikooshkbaghi/mavebay

Bayesian Version of MAVENN1.0

Installation

One can install the package directly from the github repository or install it from the local cloned version.

The base installation "mavebay" avoids additional packages such as matplotlib, arviz, logomaker and etc.

The full installation which is suitable to run demos is also provided "mavebay[examples]".

Installing from github

1. Create the virtual environment.
2. Install the base or full mavebay package.

python -m venv test_mavebay
source test_mavebay/bin/activate
pip install git+https://github.com/mahdikooshkbaghi/mavebay "mavebay[examples]"

Installing from the cloned repository

1. Clone the repo.
2. Create the virtual environment.
3. Install the base or full mavebay package.

git clone git@github.com:mahdikooshkbaghi/mavebay.git
cd mavebay
python -m venv test_mavebay
source test_mavebay/bin/activate
pip install . "mavebay"
# OR
pip install . "mavebay[examples]"

Demos

The global epistasis (GE) measurement process example script is provided in the example folder. The following command can be used to run demos

python global_epistasis_demo.py -n [NUM_SAMPLES]        \
                                -lr [LEARNING_RATE]     \
                                -m [METHOD]             \
                                -ds [DATA]              \
                                -k [INTERACTION_ORDER]  \
                                -d [DEVICE]             \
                                -i [INIT_LOC_FN]        \
                                -p [PROGRESS_BAR]

All the arguments has some default values which are provided in the script.

• NUM_SAMPLES: number of samples for mcmc or number of steps in the svi.
• LEARNING_RATE: the learning rate for the optimizer in svi method.
• METHOD: method of inference: mcmc or svi
• DATA: dataset to use for the inference. Descriptions of the datasets are given in the MAVE-NN manuscript.
  • abeta: DMS data for Aβ (default dataset).
  • tdp43: DMS data for TDP-43.
  • mpsa: MPSA data for 5' splicing sites.
• INTERACTION_ORDER: k=1 (default) corresponds to the additive GP map, k=2 pairwise GP map and so on.
• DEVICE: cpu (default) or gpu.
• INIT_LOC_FN: initialization for the svi sampling. Default it feasible. Others can be assigned based on the numpyro documentation.
• PROGRESS_BAR: enable (default) or disable the progress bar of the inference.

TODO list for MAVEBAY

Bayesian Version of MAVENN1.0

• The SVI is working both on the abeta and TDP-43 additive inferences.
• The MCMC on the abeta is working only on small samples.
• Need to figure out the batch and plate.
• Need to modify the setup.py to have additional python requirements for the example folder. Check the numpyro github repo for hint.
• Implementing the skewed-T noise model similar to one we have in MAVENN.
• K-th order interaction GP map implementation.
  • The K-th=1: which is practically an additive model and it is working.
  • The K-th=2: which is practically pairwise works on mpsa data.
• Measurement process agnostic (MPA) implementation.
• Need to add different initialization strategy for SVI sampling with default being init_to_feasible
• Saving the model:
  • SVI
  • MCMC
• Make the ppc smooth.
  • As I suspected the number of samples from posteriors were not enough to make the phi_to_yhat smooth. Increasing that fixed the issue.
• Put the MAVENN heatmap and pairwise to utils function
• Information metrics calculation.