Predicting compound activity from phenotypic profiles and chemical structures

The code release:

The related preprint Predicting compound activity from phenotypic profiles and chemical structures.

The dataset, which includes all the related data is on Zenodo

Experimental setup

The overview structure of the project

.
├── analysis            Jupyter and R notebooks for analysis and plotting
├── assay_data          Assay matrix, and lists with broad_ids\SMILES
├── plots               Some of the produced plots
├── predictions         Prediction files: aggregated. Raw files are too big for GitHub, all prediction files are available on Zenodo (dataset).
├── python              Python scripts for specific tasks.
├── scripts             Training and prediction shell and python scripts (for running ChemProp).
├── LICENSE
└── README.md

Analysis notebooks

├── analysis  Jupyter and R notebooks for analysis and plotting
    ├── 01-PUMA_feature_transform  Example of how sphering batch correction was done for morphological data.
    ├── 02-dataset-filtering  Example of how data filtering was performed with hierarchical clustering (for assays) and PAINS\hit frequency (for compounds). We share already filtered assay data: 270 assays and 16170 compounds. 
    ├── 03-generate_splits_and_UMAP This notebook is used to generate training-validation scripts for different types of experiments, some dataset stats and UMAP plots.
    ├── 04-late_fusion Generate late fusion predictions.
    ├── 05-tables_for_HitRates_plot Generate tables with hit rates and top-rank hit rates. Used in the next notebook.
    ├── 06-Hit-Rates_EF Enrichment factor plots as in Supplementary Figure 7.
    ├── 07-assay_modality_AUC_plots Generate plots with correlations between training data size and AUC scores. 
    ├── 08-exceed_auc Generate cluster map plots as in Supplementary Figure 14.
    ├── 19-Venn_diagrams Generate Venn diagrams for single and combined modalities. 
    ├── 10-AUC_comparison Generate plots for Supplementary Figure 4 of the paper (also figure 2A) to compare the AUROC scores of single modalities. 
    ├── 11-Figure_auc R notebook to generate single assay ROC plots as in Figure 4. 
    ├── 12-Boxplots Generate boxplots for single and combined modalities. 

Figures to data-scripts-notebooks or dataset

This section points to the data and code that was used to generate specific figures. Also, see the previous section for the description of the analysis notebooks.

• Figure 1B Can be generated with assay metadata ./assay_data/assay_metadata.csv, which also can be found in the Excel file (see paper's dataset on Zenodo).
• Figure 1C This data is obtained with help of notebook ./analysis/03-generate_splits_and_UMAP.ipynb.
• Figure 1D That is the final plot of ./analysis/02-dataset-filtering.ipynb, similar plots can be produced with ./analysis/08-exceed_auc.ipynb.
• Figure 1E Generated in ./analysis/03-generate_splits_and_UMAP.ipynb.
• Figure 1F This data can be found in the Excel file (see paper's dataset on Zenodo).
• Figure 2A Generated in ./analysis/10-AUC_comparison.ipynb with predictions/scaffold_median_AUC.csv.
• Figure 2B Generated in ./analysis/09-Venn_diagrams.ipynb with predictions/scaffold_median_AUC.csv (Venn diagram) and Excel file (bar plots, see paper's dataset on Zenodo).
• Figure 2C Generated in ./analysis/12-Boxplots.ipynb, also see script ./python/mean_and_median.py and data predictions/scaffold_median_AUC.csv.
• Figure 2D See Excel file (see paper's dataset on Zenodo) and predictions/scaffold_median_AUC.csv).
• Figure 2E See Excel file (see paper's dataset on Zenodo).
• Figure 3A Generated in ./analysis/09-Venn_diagrams.ipynb with predictions/scaffold_median_AUC.csv (Venn diagram) and Excel file (bar plots, see paper's dataset on Zenodo).
• Figure 3B See Excel file (see paper's dataset on Zenodo).
• Figure 3C See data predictions/scaffold_median_AUC.csv.
• Figure 3D Obtained with ./python/retrospective.py
• Figure 3E Generated in ./analysis/12-Boxplots.ipynb, also see script ./python/mean_and_median.py and data predictions/scaffold_median_AUC.csv.
• Figure 4 - Generated ./analysis/12-Figure_auc.Rmd with prediction files (see paper's dataset on Zenodo).
• Supplementary Figure 3 Generated in ./analysis/09-Venn_diagrams.ipynb with predictions/scaffold_median_AUC.csv (Venn diagram) and Excel file (bar plots, see paper's dataset on Zenodo), retrospective data with ./python/retrospective.py.
• Supplementary Figure 4 Generated in ./analysis/07-assay_modality_AUC_plots.ipynb with predictions/scaffold_median_AUC.csv.
• Supplementary Figure 5 Generated in ./analysis/10-AUC_comparison.ipynb.
• Supplementary Figure 7 Generated in ./analysis/06-Hit_Rates_EF.ipynb, the output of ./analysis/05-tables_for_HitRates_plot.ipynb is needed.
• Supplementary Figure 8 Generated in ./analysis/07-assay_modality_AUC_plots.ipynb.
• Supplementary Figure 10 B and C are generated in ./analysis/03-generate_splits_and_UMAP.ipynb, A was generated separately, see also 01-PUMA_feature_transform.ipynb. The left plot in A is before batch correction (file population_normalized.csv.gz in the paper's dataset on Zenodo), and right after batch correction.
• Supplementary Figure 11 Scripts plot_all_sim.py and plots_m.py. The input data for those scripts can be found in the paper's dataset on Zenodo.
• Supplementary Figure 12 See Excel file (see paper's dataset on Zenodo).
• Supplementary Figure 13 That is the final plot of ./analysis/02-dataset-filtering.ipynb, similar plots can be produced with ./analysis/08-exceed_auc.ipynb.
• Supplementary Figure 14 Generated with ./analysis/08-exceed_auc.ipynb.
• Supplementary Figure 15 Can be generated with assay metadata ./assay_data/assay_metadata.csv, which also can be found in the Excel file (see paper's dataset on Zenodo).

*Excel file is in /misc/Tables for PUMA.xlsx in paper's Zenodo dataset.

Addendum

See addendum for clarifications about the reported results.