thrifty-experiments-1

SHM experiments with the netam package as described in the manuscript Thrifty wide-context models of B cell receptor somatic hypermutation by Sung, et al.

Start by creating an environment and installing netam into it as described in the netam README.

To do the analysis described here, you will need to install the shmex package contained in this repository. To do so, clone this repository and install the package with from the root of this repository.

pip install -e .

In an environment into which you have installed epam and netam.

Local configuration and data

First download data from Dryad, which is available at https://doi.org/10.5061/dryad.np5hqc044.

Then edit shmex/local.py to reflect your directory structure.

To run a small trial analysis to see if things are working, enter the train directory and execute

snakemake -cN --configfile config_test.yml

Running the primary experiments

To train the main models and do the validation, enter the train directory and execute

snakemake -cN

This will run the analysis on N cores (substitute your desired number of cores for N).

To run the more limited analysis on all the models, enter the train directory and execute

snakemake -cN --configfile config_human_all.yml

Notebook-based experiments

Other associated experiments are in the following notebooks.

• cnnpe.ipynb: Adding a positional encoding to the CNN
• cnnxformer.ipynb: Adding a transformer to the CNN makes it worse
• crepe_of_shmoof.ipynb: Fitting the original Spisak et al. model weights into the framework used here
• data-description.ipynb: Exploration of SHMoof data sets
• model_summaries.ipynb: Summarizing model shapes
• multihit_*: Multihit analysis to be described in a future manuscript
• neutral_codon.ipynb: Also part of the multihit analysis
• noof.ipynb: A transformer on the kmer embeddings is not a good model
• performance.ipynb: Main model comparison notebook
• persite_wrapper.ipynb: Developing the PersiteWrapper that adds a per-site component to a model and showing that regularizing it doesn't help
• reshmoof.ipynb: Re-fitting the SHMoof model, playing with regularization, showing that per-site mutability tracks per-site motif mutability
• shm_oe.ipynb: Oberved/expected plotting