HLM

/code

Please refer to the code manual for complete information.

---

/data_record

Contains experimental data for GCN and DMPNN, organized by category.

GCN

• /data_record/GCN/GCN_no_data_selection
  Results of hyperparameter optimization using Optuna without data selection.

• /data_record/GCN/GCN_trying
  Results of hyperparameter optimization using different samplers in Optuna.

• /data_record/GCN/Remove_MW_TPSA_outlier
  Results after applying data selection based on MW and TPSA to remove outliers.

DMPNN

• /data_record/DMPNN/No_data_selection
  Results of hyperparameter optimization using Optuna without data selection.

• /data_record/DMPNN/N3_N7_ring_substructures
  Data selection based on compound substructures
  (N3 = three-membered ring, N7 = seven-membered ring).
  See the PPT inside the folder for details.

• /data_record/DMPNN/Even class training
  Training datasets generated by clustering and oversampling minor classes
  to achieve an even class distribution.
  See the PPT inside the folder for details.

• /data_record/DMPNN/remove_wwl_outlier
  Data selection based on the WWL distance matrix, using the IQR method
  to remove outliers. See the PPT inside the folder for details.

---

/Fingerpint

Fingerprint CSV files generated with padelpy for the datasets used in this study.

---

/dataset

Datasets categorized according to different data selection methods.
These files are directly used when running scripts in /code.

• eli_duplcate_hf.csv
  Main file used for model training (baseline, no data selection).

• regression_equal_no_ion_5_smaller.csv
  Original dataset passed down from a senior colleague.

---

/compound_figure

Compound structures visualized using RDKit, ordered according to
the SMILES column in eli_duplcate_hf.csv.

---

data_searching_protocol.docx

Documentation on how the data in regression_equal_no_ion_5_smaller.csv was collected.