BaSe (abbreviation for Balancing Selection ) is a supervised machine learning algorithm, implementing both regular artifical neural network (ANN) and convolutional neural network (CNN), to predict selection and distinguish it between incomplete sweep and balancing selection. It includes 3 sequantial tests:
- Test 1: Neutrality vs Selection
- Test 2: Incomplete sweep vs Balancing selection
- Test 3: Overdominance vs Negative frequency dependent selection
Each test performs a binary classification. The first test is to determine if the target allele is under selection. Here, the selection class is a mixture of both balancing and positive selection scenarios. If the target allele is under selection, the second test is to determine whether it is balancing selection or incomplete sweep. If it is a balancing selection, the third test distinguishes different types of balancing selection: overdominance and negative frequency dependent selection.
To generate training data, we performed simulations using SLiM simulation program.
To train ANN, a number of potentially informative summary statistics are calculated for each simulation
to represent output sequence, with the help of scikit-allel package. To train CNN, simulation outputs
are converted into images. Then, trained models are used to perform prediction on real data.
Please look at the Simulations/README.md for a quick tutorial on how to generate simulations.
The jupyter notebooks Tutorials/ANN_training.ipybn and Tutorials/CNN_training.ipybn
provide examples on how to use BaSe to train ANN and CNN, respectively. Then, the jupyter
notebook Tutorials/Prediction.ipybn demonstrate how to use trained models for prediction on
real data.