Causality for Computational Biology
Citation: Aoki, Raquel, and Martin Ester. "ParKCa: Causal Inference with Partially Known Causes." Pac Symp Biocomput. 2021 (link)
TODO: Update train.py
To use a particular method, check this tutorial.
a. Real-world dataset (TODO - Missing TCGA)
b. Simulated dataset (link)
from CompBioAndSimulated_Datasets.simulated_data_multicause import *
sdata_gwas = gwas_simulated_data(prop_tc=0.05, pca_path='/content/CompBioAndSimulated_Datasets/data/tgp_pca2.txt')
X, y, y01, treatement_columns, treatment_effects, group = sdata_gwas.generate_samples()The following causal inference methods were implemented:
- BART (very time-consuming for experiments with large number of treatments)
- CEVAE
- Deconfounder Algorithm
colnamesX = ['Col'+str(i) for i in range(X.shape[1])]
level1data = learners(['CEVAE','DA','BART'],pd.DataFrame(X),y01, TreatCols = None, colnamesX=colnamesX)- from train.py, meta_learner(data, meta-learners, prob): run the meta-learners in the level 1 dataset + known causes. If working on the real-world datset, set prob = 1.
level1data['y'] = 0
level1data['y'] = [1 if i in known_causes else 0 for i in level1data['causes'].values]
roc, output, y_full_prob = parkca.meta_learner(level1data.fillna(0), ['lr','nn','upu','rf'],'y')
output
metalearners pr_test re_test auc_test f1_test f1_full pr_full re_full
0 rf 0.112211 0.202381 0.510435 0.252295 0.144374 0.173531 0.461988
1 lr 0.102883 0.467262 0.502693 0.126300 0.168636 0.072530 0.488304
2 random 0.136508 0.127976 0.518104 0.063189 0.132104 0.063516 0.062865
3 upu 0.103728 0.571429 0.505856 0.127821 0.175583 0.071643 0.592105
4 adapter 0.101818 1.000000 0.500000 0.128042 0.184818 0.068400 1.000000
5 nn 0.103462 0.773810 0.506844 0.125644 0.182520 0.068434 0.766082
6 ensemble 0.103504 0.571429 0.505181 0.125747 0.175262 0.070447 0.584795 - eval.py: ROC plots, ROC curve for learners evaluation, simulation evaluation (pehe, roc curve to find new causes, diversity, and others)
- eval.R + pickle_reader.py: code to generate the plots in the paper
qav, q_ = eval.diversity(['cevae' ],['coef'], data) #qav: average value, q_: array with the pairwise diversity
qav
-0.0482939868370808