From cyst to target and back

Code, data, trained models, and analysis outputs accompanying the manuscript “From Phenotypic Screening to Target and Compound Prioritization for Autosomal Dominant Polycystic Kidney Disease.”

The repository is organized in the order of the study workflow. Precomputed notebook outputs and figures are committed so that the reported results can be inspected without rerunning computationally expensive or network-dependent steps.

Quick start

1. Clone the complete repository

Git LFS is required for four large CSV files. Install Git LFS before cloning, then run:

git lfs install
git clone https://github.com/CDDLeiden/cyst-to-target-and-back.git
cd cyst-to-target-and-back
git lfs pull
git lfs fsck

If you already cloned without Git LFS, install it and run git lfs pull from inside the clone.

2. Create the locked environment

Install Pixi, then run from the repository root:

pixi install --locked

Pixi installs the exact Linux or macOS packages recorded in pixi.lock; there is no separate Conda activation or pip installation step. The lock currently supports linux-64 and Apple Silicon (osx-arm64).

Use pixi shell if you want an interactive shell inside the environment.

GPU model training

A GPU is not needed to inspect results or rerun the phenotypic analyses. The locked default environment intentionally uses CPU XGBoost on both supported platforms. Full QSAR retraining in modelTrain-adora1_nr3c2.ipynb was performed with XGBoost 2.1.0 on NVIDIA GPUs. To retrain, use a Linux environment with the matching GPU-enabled XGBoost/CUDA build and change the notebook's device values to an available device (for example, cuda:0). The exact training/validation subsets, optimized model files, parameters, metrics, and Optuna study databases are committed, so model claims can be inspected without GPU retraining.

What can be reproduced?

The repository supports three different levels of verification:

1. Inspect the reported outputs. The notebooks retain their executed cells, plots, model metrics, and statistical results. Figures, trained models, Optuna studies, and processed tables are also committed.
2. Rerun analyses from supplied data. The screening, target-validation, compound-exploration, gene-expression, and model-analysis inputs are included. Use the working directories listed below.
3. Regenerate upstream datasets. Some steps use live PubChem, ChEMBL, or UniProt services and Papyrus 05.6. These can be slower and may differ as external services evolve. SmallWorld analogue searches and supplier stock searches are represented by their supplied result tables; the commercial queries themselves are not replayed by this repository. Experimental assay generation is outside the computational workflow.

See data/README.md for file-level descriptions and provenance.

Workflow and execution order

Notebook paths are relative to their own numbered directory. Start Jupyter from that directory (not from the repository root) so those paths resolve correctly.

Stage	Purpose	How to inspect or rerun
01_ADPKD_ScreeningProcess	Reprocess the published phenotypic screens and identify hits	From the repository root: pixi run python 01_ADPKD_ScreeningProcess/screening_process.py, then pixi run python 01_ADPKD_ScreeningProcess/hit_identification.py. These steps use live PubChem/ChEMBL queries; the exact precomputed hit table is supplied in data/adpkd_screening/identified_hits/.
02_TargetID_and_Prioritization	Link screened compounds to Papyrus 05.6 bioactivities and prioritize targets	pixi run jupyter-target-identification. For the supplied-data route, use the notebook section Loading from saved bioactivityLinker. The first-time route downloads Papyrus and uses UniProt.
02_TargetID_and_Prioritization/gene_expression	Plot expression of prioritized targets in mIMCD3 cells	pixi run jupyter-gene-expression.
03_Target_Validation	Normalize, plot, and statistically analyze the target-validation screen	Run pixi run jupyter-target-validation. The normalization audit can be regenerated with pixi run audit-normalization; it writes the per-well raw and normalized values and within-plate Mann–Whitney audit. See 03_Target_Validation/README.md.
04_Virtual_Screening	Train/evaluate A1AR and MR QSAR models, screen analogues, and analyze selected compounds	pixi run jupyter-virtual-screening, then open modelTrain-adora1_nr3c2.ipynb, virtualScreening.ipynb, or ADPKD-ExplorationScreening-analysis.ipynb. Full training requires a configured GPU; committed models and outputs allow inspection without retraining. See 04_Virtual_Screening/README.md.
05_A1R_Screening	Report radioligand displacement results for selected A1AR compounds	See 05_A1R_Screening/README.md. This is an experimental result summary, not a computational pipeline.

To register the environment as a selectable Jupyter kernel if needed:

pixi run python -m ipykernel install --user --name cystToTarget --display-name "cystToTarget"

Repository layout

• data/ — supplied raw/processed computational inputs and audit tables
• figures/ — manuscript figures and molecular drawings
• paper_tables/ — machine-readable manuscript tables
• src/papyrusBioactivityLinker/ — local compound-to-Papyrus linking package
• 01_... through 05_... — ordered analysis stages described above

Key supplied outputs

• Target-prioritization figures: figures/ADPKD-Prioritized_Target_Set-prioritized-geneName.svg and figures/ADPKD-Prioritized_Target_Set-prioritized-proteinName.svg
• Target-validation statistical results and figures: 03_Target_Validation/README.md
• QSAR, virtual-screening, and compound-exploration results: 04_Virtual_Screening/README.md
• A1AR predicted and experimental affinity summary: 05_A1R_Screening/README.md

License

The code is released under the MIT License. Please cite the accompanying paper when using this repository or its data.