GanCtrl

Study-aligned synthetic controls from time-matched treatment data for preclinical toxicology using generative AI

Quick Start • Workflow & Scripts • Methods • Data & Outputs • Citation

---

GanCtrl generates time-matched synthetic control profiles from treatment-derived clinical pathology data.

---

Overview

GanCtrl (GAN-based synthetic control) is a conditional VAE-GAN framework that translates high-dose treatment-derived clinical-pathology profiles into their time-matched control equivalents.

Developed using the Open TG-GATEs rat in vivo repeat-dose dataset, GanCtrl models 38 clinical pathology measurements using a context-conditioned encoder, an attention-aware biologically informed decoder, and an adversarial discriminator.

The framework is designed to generate physiologically coherent synthetic controls while preserving biologically relevant relationships required for downstream toxicological evaluation.

---

GanCtrl at a Glance

Model	Conditional VAE-GAN
Input	High-dose treatment clinical-pathology profiles
Output	Time-matched synthetic control profiles
Dataset	Open TG-GATEs rat in vivo repeat-dose studies
Measurements	38 clinical pathology endpoints
Conditioning	Body weight, timepoint, replicate identity, study-specific clusters
Evaluation	Cosine similarity, RMSE, biological co-elevation, toxicity concordance
Benchmarks	Inter-laboratory, intra-laboratory, replicate control, VCG, VCG-LR

---

Quick Start

1. Clone GanCtrl

git clone https://github.com/CHANDMX20/GanCtrl.git
cd GanCtrl

2. Obtain the input data

The preprocessed training and held-out test datasets used by GanCtrl are available through Zenodo:

Download GanCtrl Input Data

Configure the required input and output paths in the relevant scripts for your local environment.

3. Choose a reproducibility path

Goal	Start from
Train GanCtrl from the beginning	Preprocessed Zenodo inputs
Generate synthetic controls	Trained GanCtrl checkpoints
Reproduce downstream analyses only	Generated predictions included in data/

All scripts required for these workflows are listed below.

---

Workflow & Scripts

Open TG-GATEs
      │
      ▼
Preprocessed Inputs
      │
      ▼
GanCtrl Training
      │
      ▼
Synthetic Controls
      │
      ├───────────────┬─────────────────┐
      ▼               ▼                 ▼
 Agreement        Biological         Toxicity
 Evaluation       Co-elevation      Concordance
      │
      ▼
 VCG / VCG-LR
 Benchmark

Scripts can be executed with python <script-path> after configuring the required local input/output paths.

Stage	Analysis	Script(s)
1	GanCtrl training	training/ganctrl_training.py
2	Synthetic-control generation	training/train_test_samples.py
3A	Inter-laboratory agreement	interlab_cosine.py · interlab_rmse.py
3B	Intra-laboratory agreement	intralab_cosine.py · intralab_rmse.py
3C	Replicate-control agreement	replicate_control_cosine.py · replicate_control_rmse.py
3D	GanCtrl vs real-control agreement	cosine.py · rmse.py
4	Biological co-elevation	co-elevation.py
5A	Concordance threshold calibration	train_concordance.py
5B	Held-out toxicity concordance	test_concordance.py
6A	VCG benchmark	vcg_baseline.py
6B	Laboratory-relaxed VCG	vcg-lr_baseline.py

---

Methods

Expand the sections below for methodological details.

🧠 GanCtrl architecture and training

GanCtrl is a one-sided conditional VAE-GAN trained to map high-dose treatment clinical-pathology profiles to their time-matched control equivalents.

Conditioning variables

• Body weight
• Timepoint
• Replicate identity
• Study-specific clusters

Training objectives

• Variance-aware Gaussian negative log likelihood
• Adversarial loss
• TBIL range-constraint loss
• Biological correlation-preservation loss
• Batch-level mean matching

These components are designed to preserve organ-level biological relationships and realistic variability while generating physiologically coherent synthetic-control profiles.

The implementation uses fixed random seeds and deterministic TensorFlow operations to support reproducibility.

---

📏 Synthetic vs real-control agreement

Generated synthetic controls are compared with corresponding real controls using:

• Cosine similarity
• Root mean squared error (RMSE)

GanCtrl is interpreted relative to three real-control benchmarks:

Inter-laboratory

Cross-study real-control agreement using the same vehicle across different laboratories.

Intra-laboratory

Within-study real-control agreement using the same vehicle and laboratory.

Replicate control

Agreement between biological replicates within a treatment, providing a benchmark for real biological variability.

GanCtrl

Agreement between each generated synthetic-control profile and its corresponding real-control profile.

---

🧬 Biological co-elevation analysis

GanCtrl is evaluated for its ability to preserve literature-anchored hepatotoxicity and nephrotoxicity biological conclusions.

Treatment-associated measurement elevations are identified using:

• One-sided Welch t-tests
• Benjamini-Hochberg false discovery rate correction

Coordinated liver responses

• ALT–AST
• ALP–TBIL
• ALP–GGT/GTP
• ALT–AST–LDH

Coordinated kidney response

• BUN–CRE

Agreement between real- and synthetic-control conclusions is evaluated using:

Metric	Definition
Recall	Fraction of real-control co-elevations also detected using synthetic controls
Specificity	Fraction of real-control non-elevations remaining non-elevated using synthetic controls
Balanced Accuracy	Mean of recall and specificity

---

🧪 Toxicity concordance

The concordance analysis evaluates whether synthetic controls reproduce the sample-level abnormal/normal classifications obtained using real concurrent controls.

For each compound-time group and clinical pathology measurement:

$z = \dfrac{x_{\text{treatment}} - \mu_{\text{control}}}{\sigma_{\text{control}}}$

The decision threshold is:

1. Calibrated using the training set
2. Fixed before evaluation
3. Applied to the independent held-out test set

Synthetic-control classifications are compared with real-control classifications using:

• True positives (TP)

• True negatives (TN)

• False positives (FP)

• False negatives (FN)

  $\text{Concordance Accuracy} = \dfrac{TP + TN}{TP + TN + FP + FN}$

Primary question: If the concurrent control arm is replaced with GanCtrl-generated synthetic controls, are the same toxicological conclusions reached?

---

📚 Historical-control VCG benchmark

GanCtrl is benchmarked against virtual control groups (VCGs) constructed from historical control data.

Historical controls are drawn from training-set concurrent controls. Controls originating from the same compound as the test group are excluded to prevent information leakage.

VCG

Matched on:

• Sacrifice time
• Vehicle
• Laboratory

VCG-LR

Matched on:

• Sacrifice time
• Vehicle

while allowing controls from different laboratories.

For each test group:

• The VCG contains the same number of animals as the corresponding real concurrent control group.
• Sampling is repeated 100 times using independent draws from the eligible historical-control pool.

---

Data & Outputs

All GanCtrl data-related information is consolidated in this section.

Raw Open TG-GATEs Data

GanCtrl was developed using rat in vivo repeat-dose clinical pathology data from Open TG-GATEs.

The raw Open TG-GATEs data are not redistributed through this repository.

Open TG-GATEs Download Page

---

Preprocessed GanCtrl Inputs

The exact preprocessed training and held-out test inputs used by GanCtrl are distributed through Zenodo because of their size.

Download Preprocessed GanCtrl Inputs

The deposit contains treatment/control inputs, metadata, and molecular descriptor features required by the model.

📂 Show principal input files

repeat_train_treatment_2d.csv
repeat_train_control_2d.csv
repeat_test_treatment_2d.csv
repeat_test_control_2d.csv

---

Generated Synthetic Controls

Generated decoded synthetic-control predictions are included in data/, allowing downstream analyses to be reproduced without retraining GanCtrl.

📊 Show prediction files

File	Description
generated_liver_train.csv	Liver synthetic controls — training set
generated_liver_test.csv	Liver synthetic controls — held-out test set
generated_kidney_train.csv	Kidney synthetic controls — training set
generated_kidney_test.csv	Kidney synthetic controls — held-out test set
generated_predictions_merged_train.csv	Combined liver + kidney predictions for training-set analyses
generated_predictions_merged_test.csv	Combined liver + kidney predictions for held-out test analyses

---

💾 Show model and inference outputs

Depending on the training and inference configuration:

g_model1_*.h5
d_model1_*.h5
composite_model1_*.h5
predictions_encoded/
predictions_decoded/
samples/generated_samples_s{K}_*.csv

Output	Description
g_model1_*.h5	Generator checkpoints
d_model1_*.h5	Discriminator checkpoints
composite_model1_*.h5	Composite-model checkpoints
predictions_encoded/	Encoded-space predictions, if enabled
predictions_decoded/	Fully decoded synthetic-control predictions

---

🔬 Show clinical pathology variables

GanCtrl evaluates 38 clinical pathology measurements spanning clinical chemistry and hematology.

Key metadata

COMPOUND_NAME
DOSE_LEVEL
SACRIFICE_PERIOD
INDIVIDUAL_ID

Liver-associated measurements

Abbreviation	Measurement
ALP	Alkaline phosphatase
ALT	Alanine aminotransferase
AST	Aspartate aminotransferase
GTP/GGT	Gamma-glutamyl transferase
LDH	Lactate dehydrogenase
TBIL	Total bilirubin
DBIL	Direct bilirubin

Kidney-associated measurements

Abbreviation	Measurement
BUN	Blood urea nitrogen
CRE	Creatinine
Ca	Calcium
Cl	Chloride
Na	Sodium
IP	Inorganic phosphorus
K	Potassium

---

Repository Organization

📁 Show repository structure

GanCtrl/
│
├── training/      # Model training and synthetic-control generation
├── evaluation/    # Agreement and toxicological analyses
├── baseline/      # Real-control agreement benchmarks
├── vcg/           # Historical-control benchmarks
├── data/          # Generated synthetic-control predictions
├── plots/         # Study-design and supporting figures
├── README.md
└── LICENSE

Individual scripts are listed in the Workflow & Scripts section.

---

Environment

Software	Version
Python	3.11.7
TensorFlow-GPU	2.4.1
R	4.4.1
Bioconductor	3.19

Additional Python and R dependencies are imported within the corresponding analysis scripts.

---

Reproducibility

GanCtrl supports two reproducibility paths.

Full reproduction Zenodo Inputs ↓ Train GanCtrl ↓ Generate Synthetic Controls ↓ Run Evaluation Pipeline

Downstream reproduction Provided Predictions ↓ Agreement Analysis ↓ Biological Evaluation ↓ Toxicity Evaluation

The repository uses fixed random seeds, deterministic TensorFlow operations, explicit training/test splits, publicly available input data, and provided synthetic-control predictions to support reproducibility.

---

Citation

If you use GanCtrl in your research, please cite the associated paper and dataset.

GanCtrl Paper

GanCtrl: A Generative AI Approach to Derive Study-Aligned Synthetic Controls for Reducing Concurrent Control Animal Use. Toxicological Sciences. https://doi.org/10.1093/toxsci/kfag099

GanCtrl Dataset

Chandra, M. GanCtrl: Synthetic Control Predictions for Liver and Kidney Clinical-Pathology Profiles (Open TG-GATEs). Zenodo. https://doi.org/10.5281/zenodo.17883691

---

License

This project is licensed under the MIT License.

See the LICENSE file for details.

---

GanCtrl
Generative AI for study-aligned synthetic controls in preclinical toxicology