Supporting data for "Reliable interpretability of biology-inspired deep neural networks"

Contents

data.tgz contains all data necessary for reproducing the analysis in the manuscript. After cloning the GitHub repository, extract the contents of this file into folder data. The archive contains the following subfolders:

dtox
DTox results, one subfolder per seed

module_relevance.tsv: contains node importance scores, with the following columns:

(first, unnamed): compound identifier

remaining columns: node identifiers (UniProt and Reactome IDs)

test_labels.csv: predictions for the test set, with two columns:

truth: true label (0 or 1)

predicted: predicted label (decimal number between 0 and 1)

mskimpact_[cancer type]_[experiment]
P-NET results using the MSK-IMPACT 2017 dataset, one subfolder per seed
[cancer type] is one of bc (breast cancer), cc (colorectal cancer), nsclc (non-small cell lung cancer), or pc (prostate cancer)
[experiment] is one of original (original setup) and shuffled (shuffled labels)

pnet_[experiment]
P-NET results using the original (prostate cancer) dataset, one subfolder per seed
[experiment] is one of deterministic (deterministic input data), original (original setup), and shuffled (shuffled labels)

node_importance.csv: contains node importance scores, with the following columns:

(first, unnamed): node name

coef: original node importance scores

coef_graph: indegree plus outdegree of node

coef_combined: adjusted node importance score (= coef / coef_graph if coef_graph > mean(coef_graph) + 5 sd(coef_graph) in the respective layer)

coef_combined_zscore: scaled coef_combined

coef_combined2: z(z(coef_graph) - z(coef))

layer: layer of the node

predictions_test.csv: predictions for the test set, with the following columns:

(first, unnamed): sample name

pred: predicted class (unfortunately, encoded by a double 1.0 or 0.0)

pred_scores: probability of the predicted class

y: true class (encoded as integer 1 or 0)

predictions_train.csv: predictions for the training set (same columns as above)

link_weights_[layer].csv: only in subfolder 234_20080808; matrices with edge weights

Changelog

v1.1.0  – 2023-06-28

added DTox results

added results of P-NET experiments with MSK-IMPACT 2017 dataset

v1.0.0 – 2023-03-22

initial release