GEMGen (Gene Expression-guided Molecular Generator) is a large language model-based framework for the inverse design of chemical perturbations conditioned on desired cellular state transitions.
Whereas conventional perturbation-response modeling asks what cellular state will result from a given perturbation, GEMGen addresses the inverse design problem of perturbation biology: given transcriptomic profiles representing the initial and target cellular states, it directly generates chemical perturbations (small molecules) predicted to induce the corresponding transition — without restricting the search to predefined compound libraries.
Cellular state transitions are represented as text-based, rank-ordered differential-gene instructions (<cell type>, up <gene list>, down <gene list>), which condition autoregressive molecular generation. The framework consists of the following core components:
- Generator: an instruction-tuned LLM that generates molecular candidates (SMILES) autoregressively, conditioned on the desired transcriptomic state transition.
- Filter: filters generated molecules for syntactic validity and drug-like properties (e.g., QED, Lipinski's Rule of Five, synthetic accessibility), producing experimentally tractable candidate lists.
- Scorer: a compound-response scoring model that estimates the likelihood that a candidate molecule reproduces the target transcriptional response, prioritizing candidates for experimental testing.
The framework is designed for flexible inference and can be easily adapted to different biological contexts and datasets — including chemical perturbation signatures, genetic perturbation (e.g., CRISPRi) signatures, and disease-associated state transitions.
The preprint of our paper is now publicly available on bioRxiv.
Read it here: https://www.biorxiv.org/content/10.64898/2026.01.03.697483v2.full.pdf
conda create -n gemgen python=3.10 -y
conda activate gemgenGEMGen depends on vllm==0.7.3. We recommend installing all dependencies via requirements.txt:
pip install -r requirements.txtAlternatively, you may install them manually:
pip install vllm==0.7.3
pip install transformers==4.57.6
pip install pandas==2.3.3
pip install rdkit==2024.3.1Note This model only supports inference on NVIDIA GPUs. A CUDA-enabled environment with a compatible NVIDIA GPU is required.
Pretrained model checkpoints are required for both the generator and the scorer.
-
Download the checkpoint folder from 10.5281/zenodo.18739230.
-
Place the checkpoint directory at a desired local path.
-
Update the corresponding checkpoint paths in the demo scripts:
demo/run_generator.shdemo/run_evaluation.shdemo/run_scorer.sh
GEMGen expects structured input data describing the desired cellular state transition (perturbation signature).
-
Generator inputs
- Cell type information
- Lists of up-regulated and down-regulated genes
- Prompt templates (see
data/generator_prompts.txtanddata/templates.txt)
-
Scorer inputs
- Generated molecules (e.g., SMILES)
- Corresponding phenotypic conditions
- Example format is provided in
data/scorer_test_demo.tsv
Please refer to the files in the data/ directory for templates and example data formats.
git clone https://github.com/DLS5-Omics/GEMGen.git
cd GEMGenThe generator produces candidate chemical perturbations (molecules) conditioned on the desired transcriptomic state transition.
bash demo/run_generator.shThe generator produces a single JSON file containing a list of results, where each entry corresponds to one input prompt from the multi-line prompts file. Specifically: The output is a list of lists. Each inner list has two elements:
- prompt (str): The original input prompt (exactly as read from the input prompts file).
- generated_smiles (List[str]): A list of SMILES strings representing the small molecules generated by the model in response to that prompt. The number of generated SMILES per prompt is determined by the user-specified sampling count parameter.
Example Structure:
[
[
"With the goal of altering gene expression patterns in K562, construct a compound that increases the expression of SAMD4B, MEIOB, ... and decreases the expression of ATPSCKMT, ...,MFHAS1.",
[
"CC1=CN=C(C(=C1OC)C)CS(=O)C2=NC3=C(N2)C=C(C=C3)OC",
"COC1=CC(=CC(=C1OC)OC)C2=C3C(=NC=N2)C=CC=C3N"
]
],
[
"In a cell population characterized as K562, formulate a compound that will upregulate the expression of PRG2, CEP85L, ..., KIAA0355 and downregulate the expression of USP32, ..., MYD88.",
[
"CC(=O)Nc1ccc(cc1)S(=O)(=O)N",
"C1COCCN1C(=O)C(=O)N"
]
]
]The evaluation step filters out duplicate and invalid molecules generated by the model. Optionally, you can control whether to compute chemical metrics for the generated small molecules and their similarity to real molecules.
bash demo/run_evaluation.sh-
Input: The JSON file generated by the generator (e.g., results/generated_molecules.json).
-
Output: A CSV file where each row corresponds to an input prompt and its associated valid generated molecules, along with evaluation metrics (if calculated).
The scorer evaluates the consistency between generated molecules and the target transcriptional response, providing a perturbation-response match score to prioritize candidates for experimental testing.
bash demo/run_scorer.shTypical outputs include:
- A list of matching scores between molecules and phenotypic inputs
- Prepare input data describing the desired cellular state transition (cell type + up-/down-regulated gene lists).
- Run the generator to produce candidate molecules.
- Filter and post-process generated molecules (remove duplicates, invalid structures, etc.).
- Feed filtered molecules into the scorer.
- Select high-scoring molecules for downstream analysis or validation.
This project is released under the terms of the license provided in the LICENSE file.