/DANGO

Primary LanguagePythonMIT LicenseMIT

DANGO: Towards the prediction of higher-order genetic interactions

This is the implementation of the algorithm DANGO for the prediction of higher-order genetic interactions. This repo contains the code and data for the analysis of the yeast trigenic interactions studied in the manuscript Towards the prediction of higher-order genetic interactions.

Requirements

  • h5py
  • numpy
  • pytorch
  • scikit-learn
  • tqdm
  • xgboost (used in the baseline, if you don't plan to run the baseline, it's not required)

Data used in this repo

All the data are stored in the ./data folder.

  • Within the ./data/yeast_network folder, there are six yeast protein-protein interaction (PPI) networks obtained from STRING v9.1 database (http://string91.embl.de/)
  • The ./data/AdditionalDataS1.tsv is obtained from the paper Systematic analysis of complex genetic interactions (Kuzmin et al., Science, 2018), which contains the measured trigenic interaction scores of around 100,000 triplets

Configure the parameters

There are several parameters when running the main program python main.py

usage: main.py [-h] [--thread THREAD] [--mode MODE] [--identifier DANGORUN1] [--split SPLIT] [--predict PREDICT]

The parameters and their meanings are

  • -t, --thread, default=8, the number of parallel threads for training the ensemble of Dango models. The code is optimized for utilizing multiple GPUs. Each Dango model roughly takes about 2G GPU memory, so one 2080Ti card could train three Dango models at a time. Change this parameter based on your machine condition.
  • -m, --mode, default='train', can be eval, train, predict or combinations of them separated by ; (no space). For instance train;test.
    • When in eval mode, the main program would run cross-validation (either random split, gene split 1 or gene split 2). See details in the below sections
    • When in train mode, the main program would train a Dango model using the whole dataset
    • When in predict mode, the main program would use a trained Dango model to make predictions on a specific set of triplets. See details in the below sections
  • -i, --identifier, default=None, the identifier for this run. A corresponding folder would be created under the ./temp folder where the trained model and prediction results would be stored. It can be useful when you train one model and want to use that specific model for making predictions later. When left empty, the program would automatically create an identifier with the date, time information embedded. When there is an existing folder under ./temp with the same name as the identifier, the original one would be overwritten.
  • -s, --split, default=0, how to split the training/validation/test set during evaluation. This parameter would only be used when eval is included in the mode
    • 0 stands for random split
    • 1 stands for gene split 1
    • 2 stands for gene split 2
  • -p, --predict, default=0, which set of triplets to predict on.
    • 0 stands for predicting on the original set of triplets (set 1 in the paper)
    • 1 stands for predicting on the set of triplets where all three genes are observed in the original dataset (set 2 in the paper)
    • 2 stands for predicting on the set of triplets where two of the genes are observed and the third gene is unobserved in the original dataset

Usage

  1. cd Code
  2. Run python process.py This script would process the data in the ./data folder. Specifically, it would:
    1. Generate two dictionaries : ./data/gene2id.npy and ./data/id2gene.npy which is two dictionaries that map between the gene name and node id (starts from 1). Can be load with np.load(xx, allow_pickle=True).item()
    2. Process PPI networks.
    3. Parse the trigenic interaction scores. The processed signals are stored as : ./data/tuples.npy, ./data/y.npy, ./data/sign.npy which corresponds to the triplets, measured trigenic interaction scores and p-values respectively.
  3. Run python main.py. The parameters can be configured as described in the above section. A folder named as the --identifier parameter would be created under the ./temp/ dir where all results would be stored. The output under that folder includes:
    1. If the --mode parameter includes eval. The training process and the final performance would be printed on the screen. The performance would also be written to a file named result.txt.
    2. If the --mode parameter includes eval or train. The selected groups of the model parameters (model enesembling trick, so there will be multiple sets of parameters) is stored as the file Experiment_model_list. The parameters can be transformed into a list of Dango models by
    from main import create_dango_list
ckpt_list = list(torch.load("../Temp/%s/Experiment_model_list" % (dir_name), map_location='cpu'))
dango_list = create_dango_list(ckpt_list)
    1. If the model runs in the random 5-fold CV (--mode=eval, --split=0), the predicted results for each fold would be stored in an hdf5 file named as Experiment.hdf5. The structure of this hdf5 file is listed. The ensemble_inverse and test_y_inverse correspond to the predicted scores and measured scores. They are both inversely transformed to the original value scale of the measured trigenic interaction scores.
    ./Cross_valid_0
	./ensemble_inverse
	./test_y_inverse
./Cross_valid_1
    1. If the --mode parameter includes predict, the corresponding results would be stored as XX_tuples.npy, XX_y.npy. Only tuples with predicted signals larger than 0.05 are kept to save space. Based on the --predict parameter, XXis named as re_eval, within_seen, two_seen_one_unseen respectively.

Cite