AIVIVE

AI-aided translation of in vitro transcriptomic responses into in vivo profiles for toxicological applications

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

AIVIVE translates in vitro transcriptomic profiles into synthetic in vivo profiles using a GAN-based translator and biologically informed local optimization.

Overview

AIVIVE (AI-aided In Vitro–In Vivo Extrapolation) is a generative AI framework designed to translate in vitro transcriptomic responses into corresponding synthetic in vivo profiles.

The framework combines a GAN-based translator with module-specific local optimizers based on biologically meaningful gene clusters to improve in vitro–in vivo extrapolation of toxicogenomic responses.

AIVIVE was developed using rat liver transcriptomic data from Open TG-GATEs and evaluated across multiple levels of biological organization, including global expression-profile agreement, differentially expressed genes, pathway-level concordance, adverse outcome pathway-associated genes, and necrosis prediction.

AIVIVE at a Glance


Framework	GAN-based translator + local optimizers
Input	Rat liver in vitro transcriptomic profiles
Output	Synthetic in vivo transcriptomic profiles
Dataset	Open TG-GATEs
Gene space	Rat S1500+ and biologically defined gene modules
Profile evaluation	Cosine similarity, RMSE, MAPE
Gene-level evaluation	Differentially expressed genes
Pathway evaluation	KEGG pathway concordance
Mechanistic evaluation	AOP-associated gene expression
Toxicological application	Necrosis prediction

Quick Start

1. Clone AIVIVE

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

2. Obtain the transcriptomic data

The RMA-normalized rat liver in vitro and in vivo training/test datasets used by AIVIVE are available through Zenodo:

Download AIVIVE Data

The data were divided into training (80%) and held-out test (20%) sets based on unique compounds.

3. Choose a reproducibility path

Goal	Start from
Reproduce preprocessing from raw data	Open TG-GATEs CEL files
Train AIVIVE from the beginning	RMA-normalized Zenodo inputs
*Generate synthetic in vivo* profiles**	Trained AIVIVE generator and local optimizers
Run biological applications	Real and AIVIVE-generated profiles

All analysis scripts are listed once in the workflow below.

Workflow & Scripts

Open TG-GATEs
      │
      ▼
Preprocessing
RMA + Annotation
      │
      ▼
 In Vitro Profiles
      │
      ▼
GAN-Based Translator
      │
      ▼
Initial Synthetic In Vivo Profiles
      │
      ▼
Module-Specific Local Optimizers
      │
      ▼
Optimized Synthetic In Vivo Profiles
      │
      ├──────────────┬──────────────┬──────────────┐
      ▼              ▼              ▼              ▼
   Profile          DEG          Pathway        Toxicity
 Evaluation       Analysis       / AOP          Prediction
      │
      ▼
Cosine • RMSE • MAPE

Scripts can be executed after configuring their required local input and output paths.

Stage	Analysis	Script(s)
1A	RMA normalization	`rma_vivo_single.R`
1B	Rat gene annotation	`rat_annotation.R`
2A	GAN translator training	`vitro_vivo_GAN.py`
2B	Generator predictions	`train_test_samples.py`
2C	Module-specific local optimization	`training/modules/`
2D	Optimized module predictions	`module_test_evals.py`
3A	Negative-control benchmark	`negative_control.py`
3B	AIVIVE cosine similarity	`cosine.py`
3C	AIVIVE RMSE	`rmse.py`
3D	AIVIVE MAPE	`mape.py`
3E	Positive-control benchmark	`positive_control.py`
4A	Real-profile fold change	`real_fold.py`
4B	Synthetic-profile fold change	`gen_fold.py`
4C	Real-profile DEG identification	`real_deg.py`
4D	Synthetic-profile DEG identification	`gen_union.py`
4E	DEG overlap	`overlap_deg.py`
5A	KEGG pathway enrichment	`kegg_pathway.R`
5B	KEGG pathway overlap	`overlap_pathways.py`
6	AOP-associated gene analysis	`AOP.py`
7	Necrosis prediction	`necrosis_prediction.py`

Methods

Expand the sections below for additional methodological detail.

🧬 Transcriptomic preprocessing

Gene-expression data are prepared before AIVIVE model development using RMA (Robust Multi-array Average) normalization and rat gene annotation.

RMA normalization

Raw microarray expression data are normalized using RMA to generate processed gene-expression profiles suitable for downstream analysis.

Gene annotation

Probe identifiers are mapped to relevant rat gene identifiers, and the resulting annotations are processed to generate a unique gene-level expression dataset.

These preprocessing steps generate the expression profiles used for AIVIVE model development and evaluation.

🧠 GAN translator and local optimization

AIVIVE consists of two principal modeling stages.

GAN-based translator

The GAN-based translator learns a mapping from:

In Vitro Transcriptomic Profile
              │
              ▼
      GAN-Based Translator
              │
              ▼
Synthetic In Vivo Profile

The generator is trained to produce synthetic in vivo transcriptomic profiles from corresponding in vitro responses.

Local optimizers

Following global GAN translation, biologically meaningful gene modules are processed using module-specific local neural-network optimizers.

These local optimizers refine the translated profiles at the module level and are designed to improve the biological correspondence between synthetic and real in vivo gene-expression patterns.

📏 Global expression-profile evaluation

AIVIVE performance on the held-out test set is evaluated using three agreement metrics:

Cosine similarity
Root mean squared error (RMSE)
Mean absolute percentage error (MAPE)

Performance is interpreted relative to two real-data benchmarks.

Negative Control

Agreement is calculated between two real in vivo profiles while excluding biological replicates.

This provides a lower-reference benchmark for profile similarity.

AIVIVE

Agreement is calculated between each optimized synthetic in vivo profile generated by AIVIVE and its corresponding real in vivo profile.

Positive Control

Agreement is calculated between biological replicates within a treatment.

This provides a benchmark representing the level of agreement observed between real biological samples.

🧬 Differential gene-expression analysis

Differentially expressed genes (DEGs) are evaluated in both real and AIVIVE-generated synthetic profiles.

The same general workflow is used for the Rat S1500+ and module-specific gene sets, with the relevant input files changed depending on the analysis.

Fold change

Fold change is calculated as:

$$\text{Fold Change} = \text{Treatment Expression} - \text{Control Expression}$$

DEG identification

A gene is classified as differentially expressed when:

$$|\text{Fold Change}| > 1$$

DEG overlap

Agreement between DEGs identified using real and generated profiles is evaluated as:

$$\text{DEG Overlap} = \frac{ |\text{DEGs}_{\mathrm{real}} \cap \text{DEGs}_{\mathrm{generated}}| }{ |\text{DEGs}_{\mathrm{real}}| }$$

For synthetic profiles, the union of DEGs across subgroups of a treatment group is used where applicable.

🛣️ KEGG pathway analysis

KEGG (Kyoto Encyclopedia of Genes and Genomes) pathway enrichment analysis is used to evaluate whether AIVIVE-generated transcriptomic profiles preserve pathway-level biological interpretation.

The analysis consists of two stages.

Pathway enrichment

Differentially expressed genes are mapped to known KEGG pathways to identify enriched biological processes.

Pathway overlap

The overlap between pathways enriched using real in vivo profiles and those enriched using AIVIVE-generated profiles is then evaluated.

This provides a pathway-level assessment of biological concordance beyond individual gene-expression values.

🔗 AOP-associated gene-expression analysis

The Adverse Outcome Pathway (AOP) analysis evaluates expression agreement for AOP-associated genes represented within the Rat S1500+ gene set.

Percentage error between real and AIVIVE-generated gene-expression values is calculated as:

$$\%\text{ Error} = \left( \frac{ \text{Real} - \text{Synthetic} }{ \text{Real} } \right) \times 100$$

The resulting errors are visualized using heatmaps to identify patterns of agreement between real and synthetic profiles across AOP-associated genes.

🧪 Necrosis prediction

A downstream predictive model is used to assess whether AIVIVE-generated synthetic in vivo profiles retain information relevant to a toxicological endpoint.

Model development

The necrosis prediction model is trained using real in vivo training profiles.

Model evaluation

The trained model is evaluated independently on:

Real in vivo test profiles
AIVIVE-generated optimized synthetic in vivo test profiles

This analysis evaluates whether synthetic profiles generated from in vitro measurements preserve transcriptomic information relevant to necrosis prediction.

Data & Outputs

All AIVIVE data-related information is consolidated here.

Raw Open TG-GATEs Data

AIVIVE was developed using rat liver transcriptomic data from Open TG-GATEs.

Raw Open TG-GATEs data can be obtained from:

Open TG-GATEs Download Page

RMA-Normalized AIVIVE Inputs

The rat liver in vitro and in vivo single-dose transcriptomic profiles used for model development are distributed through Zenodo:

Download AIVIVE Data from Zenodo

The data are RMA normalized and split into:

Training set: 80%
Held-out test set: 20%

The split is based on unique compounds.

Repository Data

Supporting data required for individual analyses are retained in data/.

📂 Show supporting data files

File	Purpose
`final_rat_genes.csv`	Rat S1500+ gene set used for AIVIVE development and training
`open_tggates_cel_file_attribute.csv`	Open TG-GATEs metadata
`modules_genes/`	Module gene sets used for local optimization
`aop_overlap_genes.csv`	AOP-associated genes overlapping the Rat S1500+ gene set
`necrosis_df.csv`	Open TG-GATEs pathological findings used for necrosis analysis

Repository Organization

📁 Show repository structure

AIVIVE/
│
├── preprocessing/        # RMA normalization and gene annotation
│
├── training/             # GAN translator training and prediction
│   └── modules/          # Module-specific local optimizers
│
├── applications/         # Evaluation, KEGG, AOP, and necrosis analyses
│
├── degs/                 # Differential gene-expression analyses
│
├── data/                 # Supporting data and gene sets
│
├── README.md
└── LICENSE

Individual scripts are listed in the Workflow & Scripts section.

Environment

AIVIVE was developed using:

Software	Version
Python	3.11.7
TensorFlow-GPU	2.4.1
R	4.4.1
Bioconductor	3.19
rat2302.db	3.13.0

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

Reproducibility

AIVIVE supports multiple levels of reproducibility depending on the desired starting point.

Full reproduction

Raw Open TG-GATEs Data
          ↓
RMA + Annotation
          ↓
Train GAN Translator
          ↓
Train Local Optimizers
          ↓
Generate Synthetic Profiles
          ↓
Run Biological Analyses

Model reproduction

Zenodo Inputs
      ↓
Train AIVIVE
      ↓
Generate Synthetic Profiles
      ↓
Run Evaluation Pipeline

The repository separates preprocessing, model development, prediction, profile-level evaluation, and downstream biological applications so that individual stages can be reproduced independently.

Citation

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

AIVIVE Paper

AIVIVE: AI-aided In Vitro-In Vivo Extrapolation. Toxicological Sciences. https://doi.org/10.1093/toxsci/kfaf100

This article was selected as a ToxSpotlight article by Toxicological Sciences.

AIVIVE Dataset

AIVIVE rat liver in vitro/in vivo transcriptomic data and associated model outputs. Zenodo. https://doi.org/10.5281/zenodo.14984579

License

This project is licensed under the MIT License.

See the LICENSE file for details.

AIVIVE
AI-aided in vitro–in vivo extrapolation for toxicogenomic applications

Name		Name	Last commit message	Last commit date
applications		applications
data		data
degs		degs
docs		docs
preprocessing		preprocessing
training		training
LICENSE		LICENSE
README.md		README.md

Latest commit

History

Folders and files

Repository files navigation

AIVIVE

Overview

AIVIVE at a Glance

Quick Start

1. Clone AIVIVE

2. Obtain the transcriptomic data

Download AIVIVE Data

3. Choose a reproducibility path

Workflow & Scripts

Methods

RMA normalization

Gene annotation

GAN-based translator

Local optimizers

Negative Control

AIVIVE

Positive Control

Fold change

DEG identification

DEG overlap

Pathway enrichment

Pathway overlap

Model development

Model evaluation

Data & Outputs

Raw Open TG-GATEs Data

RMA-Normalized AIVIVE Inputs

Download AIVIVE Data from Zenodo

Repository Data

Repository Organization

Environment

Reproducibility

Full reproduction

Model reproduction

Citation

AIVIVE Paper

AIVIVE Dataset

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages