This repository holds active learning results for the PstP dataset.
Iterative results are stored as follows:
results/setting_true/: denotes the setting number or identifier for different settings (i.e. batch size, max iters, etc.)results/setting_true/training_data/cumulative_*.csv.gz: denotes training compounds for iteration *. Note that binary labelling for iter_i.csv and iter_j.csv may change due to adaptive labelling.results/setting_true/next_batch/iter_*.csv.gz: denotes selected compounds for iteration *.
Process for data communication goes as follows for iteration i:
- Next batch selected compounds will be posted as
results/setting_true/next_batch/iter_i.csv.gzwith no labels because they are unknown to the uploader. For easy record keeping, the commit message will be "Requesting new labels for iteration i.". - Oracle data manager will then update their local repository, run processing scripts to generate new labels (see generate_pstp_labels.py), then upload the new labels.
One file will be uploaded:
results/setting_true/training_data/cumulative_(i+1).csv.gzcontaining new labeled training data which includes recently requested batch. For easy record keeping, the commit message will be "Generated new labels for iteration (i+1).". Note that the oracle has access to all previous requests and labeling.
All PstP chemical screening data comes from the following manuscript, which should be cited if the dataset is reused:
Identification and characterization of inhibitors of the tuberculosis phosphatase PstP.
Chase Riedel, Jeremy Rahkola, Matthew Reichlen, Hunter Ries, Spencer S Ericksen, Martin Voskuil, Anthony Gitter, Nathan Wlodarchak.
Journal of Biological Chemistry. 302:4, 2026.
The original data is available in PubChem.