SWIF(r): SWeep Inference Framework (controlling for correlation)
SWIF(r) Version 1
Lauren Alpert Sugden, Ramachandran Lab, Brown University
Publication: Sugden, Lauren Alpert, Elizabeth G. Atkinson, Annie P. Fischer, Stephen Rong, Brenna M. Henn, and Sohini Ramachandran. "Localization of adaptive variants in human genomes using averaged one-dependence estimation." Nature communications 9, no. 1 (2018): 703.
Questions: Contact firstname.lastname@example.org or post a git issue
This directory contains the following:
- SWIFr_train.py - source code for training SWIFr on a set of statistics calculated on simulated data
- SWIFr.py - source code for applying SWIF(r), once trained, to a file with testing data
- example_2classes - example input and output for SWIF(r) trained on two classes ("neutral" and "sweep")
- example_3classes - example input and output for SWIF(r) trained on three classes ("neutral", "sweep_recent", and "sweep_old")
- calibration.py - source code for calibrating probabilities such as those calculated by SWIF(r)
- calibration/example - example input and output for calibration.py
To train SWIF(r), run SWIFr_train.py with the --path flag pointing to the directory containing the input files.
>python SWIFr_train.py --path example_2classes/
Required Input Files:
Note: the following files and directories must be in a single directory, the path to which will be passed to SWIFr_train.py using the flag --path. The directory must contain classes.txt, component_stats.txt, and a directory called simulations. The simulations directory must have a subdirectory with the same name as each entry in classes.txt. Each subdirectory should have a set of files to be used for training that class. Example directories can be found in example_2classes/ and example_3classes/.
classes.txt component_stats.txt simulations/ neutral/ simfile1.txt simfile2.txt ... sweep/ simfile1.txt simfile2.txt ...
classes.txt - this file has one line for each class you wish to train.
component_stats.txt - this file has one line for each summary statistic. The same statistic names must be in the header of the training files below. The order of statistics does not matter.
Fst XP-EHH iHS DDAF
simfile.txt - tab-delimited file with one line per simulated SNP. File can have any number of columns for identification, but must have one column for each of the statistics in component_stats.txt. Header line should use the names of the statistics with the same spelling and capitalization as in component_stats.txt. The statistics need not be in the same order. Please note: the classifier will be trained on every line in the file, so make sure that files only contain SNPs that you wish to train with. For example, if you only want to train a sweep class on the actual swept site, the file must only contain that site. Also note: the value -998 is used to denote missing values for SNPs where the statistic of interest is undefined (or otherwise missing).
SNP position Fst XP-EHH iHS DDAF SNP1 1 3.4 2.2 -3.4 1.3 SNP2 2 0.1 5.0 -998 -3.2 ...
Command Line Options:
--path <string>: relative path to all input files
Required (default: '')
e.g. --path example_2classes/
--retrain <bool>: use to re-train classifier after manually altering joint_component_nums or marginal_component_nums
SWIFr_train.py creates three directories: AODE_params/ which contains the parameter files that SWIFr.py will use, component_statistic_distributions/ which contains illustrations of the learned marginal (1D) and joint (2D) distributions, and BIC_plots/ which contains Bayesian Information Criterion curves (https://en.wikipedia.org/wiki/Bayesian_information_criterion) that SWIF(r) uses to learn the number of components for the gaussian mixtures for each of the 1- and 2-dimensional component statistic distributions. The BIC balances model complexity with fit, so that the optimal number of components is the one that minimizes the BIC. SWIF(r) automatically chooses the number of components between 1 and 10 that defines this minimum. The number chosen is highlighted in pink in the pdf files in BIC_plots/.
If you wish to change the number of Gaussian mixture components based on the BIC curves (filenames in BIC_plots/ contain the statistic or pair of statistics, and the class), you can edit the file joint_component_nums and/or marginal_component_nums in the AODE_params/ directory, then re-run SWIFr_train.py with the flag --retrain. This will regenerate the files in AODE_params/, BIC_plots/, and component_statistic_distributions/.
Running SWIF(r) on testing data:
Command Line Options:
--path2trained <string>: relative path to the directory containing the input and output files from SWIFr_train.py (same as --path argument for SWIFr_train.py)
Required (default: '')
e.g. --path2trained example_2classes/
--pi <pi_1 pi_2 ...>: prior probabilities on the classes. Number of arguments must match the number of classes, must be in the same order as classes.txt, and must add to 1.
Default: 0.99999 0.00001
--interactive <bool>: use to calculate posterior probabilities for a single site. You will be prompted for values for each of the component statistics, and the output will be a posterior probability for each class.
--file <string>: use to calculate SWIF(r) on multiple sites, where argument is the filename (including relative path). File must be in the same tab-delimited format as simfile.txt above; any number of columns are allowed, as long as there is a column for each component statistic, and a header line that uses the names of the statistics exactly as they appear in component_stats.txt. Again, use -998 to denote any missing values. SWIFr.py will output a file with all of the columns from the original, plus n columns that record the prior value given upon input for each class, and n columns for the posterior probabilities for each class.
--stats2use <stat1 stat2 ...>: use if you want to run SWIF(r) with a subset of the statistics that were used to train. Make sure that statistic names listed match those used for training. Can be combined with either --interactive or --file.
--outfile <string>: specify file for output (including path). By default, output will be written to the same directory as the original file, with name <filename>_classified. Only use with --file.
Note: either --interactive or --file must be specified.
python SWIFr.py --path2trained example_2classes/ --interactive
Value for Fst: 2.3
Value for XP-EHH: 3.3
Value for iHS: -2.5
Value for DDAF: 3.6
Probability of neutral: 0.860299223365
Probability of sweep: 0.139700776635
application_example/ in this repository has classified files generated from the following commands:
> python SWIFr.py --path2trained example_3classes/ --pi 0.99998 0.00001 0.00001 --file application_example/test_block_3classes
>python SWIFr.py --path2trained example_2classes/ --pi 0.99999 0.00001 --file application_example/test_block_2classes
Please note: training examples in example_2classes/ and example_3classes/ are based on a toy dataset. The classifiers learned in these examples should not be used for real applications. Instead, use SWIFr_train.py to train the classifier on a complete set of user-supplied training examples.
Command Line Options:
--frac1 <float>: for calibration purposes, fraction of data points that come from class 1 (sweep).
Required (default: 0.1)
e.g. --frac1 0.001
--train <bool>: use to learn calibration function from training data
--input_train <string>: file containing training data for calibration. Two tab-delimited columns, no header, each line is a datapoint. First column contains labels (e.g. 0 for neutral, 1 for sweep) and second column contains raw (un-calibrated) probabilities.
Required for --train
e.g. see calibration/example/test_for_calibration.txt
--apply <bool>: use to apply calibration learned with --train to a new set of probabilities.
--input_apply <string>: file containing probabilities to be calibrated. Can have any number of tab-delimited columns, with a header line. One column name must be "uncalibrated" and that column must contain the probabilities to be calibrated. Other columns may have rsids, chromosome numbers, etc...
Required for --apply
e.g. see calibration/example/test_for_application
Note: Either --train or --apply must be specified. Currently, calibration functionality only supports binary classification probabilities.
>python calibration.py --frac1 0.1 --train --input_train calibration/example/test_for_calibration.txt
>python calibration.py --frac1 0.1 --apply --input_apply calibration/example/test_for_application
calibration.py will create a directory called calibration/ if one does not already exist. With the --train flag, calibration.py will draw two reliability plots (before and after calibration) and will save them in pdf format in the calibration/ directory. It will also save .p files that contain the values necessary for applying smoothed isotonic regression calibration to a new set of datapoints. With the --apply flag, calibration.py will create a new file in the calibration/ directory with the same name as the file provided with --input_apply, appended with "_calibrated". This file will contain all of the columns in the original file, plus a final column with column name "calibrated" that provides the new calibrated probabilities for each raw probability. For examples of all of these, see files in calibration/example/.