🧬 biology

Inverse design of chemical perturbations for desired cellular state transitions with GEMGen

The paper introduces GEMGen, a large language model-based framework that solves the inverse design problem in perturbation biology by generating novel chemical structures to induce desired cellular state transitions, successfully identifying and validating potent antifibrotic compounds in animal models.

Original authors: Sida Shao, Qun Jiang, Xing Ye, Zekun Guo, Yingce Xia, Zequn Liu, Jiaying Xu, Peiran Jin, Fusong Ju, Huanhuan Xia, Shangya Feng, Haiguang Liu, Tao Qin, Pan Deng

Published 2026-11-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sida Shao, Qun Jiang, Xing Ye, Zekun Guo, Yingce Xia, Zequn Liu, Jiaying Xu, Peiran Jin, Fusong Ju, Huanhuan Xia, Shangya Feng, Haiguang Liu, Tao Qin, Pan Deng

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Scientists have long sought to understand how to steer the machinery of life. Cells are not static; they are dynamic systems that can shift from a healthy state to a diseased one, or from a resting state to an active one, in response to various signals. For decades, researchers have mapped these shifts by observing what happens when they poke a cell with a specific chemical or genetic nudge. This forward approach—applying a change and watching the result—has been incredibly successful, creating vast libraries of data that tell us how cells react to thousands of different substances. However, a more difficult challenge remains: working backward. If a researcher knows exactly how a cell should change to become healthy again, can they find or create the specific chemical that will force that change? This is the inverse design problem, a quest to start with a desired outcome and work backward to discover the tool that makes it happen.

A team of researchers has now tackled this challenge with a new artificial intelligence framework called GEMGen. Instead of simply scanning existing lists of chemicals to see which ones might work, this system learns to invent new ones. The researchers trained the model on a massive collection of data linking chemical compounds to the specific changes they cause in a cell's genetic activity. In this context, genetic activity refers to the thousands of genes inside a cell that are turned on or off, creating a unique signature that defines the cell's state. The model learned to read these signatures not as complex numbers, but as ordered lists of genes, much like reading a set of instructions. It then learned to translate those instructions into the chemical structures of molecules that could produce the desired effect.

The power of this approach lies in its ability to look beyond what is already known. Traditional drug discovery often involves testing millions of existing chemicals, hoping to find a match. GEMGen, by contrast, generates entirely new molecular structures from scratch. In a series of tests, the system was asked to find chemicals that could reverse specific genetic changes. When given the genetic signature of a known drug, the model successfully generated new molecules that were structurally similar to the original drug. More impressively, it also found molecules that were chemically different but worked through the same biological mechanism. This suggests the system learned the underlying logic of how chemicals affect cells, rather than just memorizing a list of known pairs.

To ensure these computer-generated ideas were not just theoretical, the researchers moved to the laboratory. They first tested the system on a genetic pathway involved in the body's response to stress. The model was given the genetic signature of a cell where a specific protective gene had been turned off. It proposed a list of new chemicals to restore that gene's activity. Among the candidates, the researchers selected a molecule called flavokawain C, which had not been previously known to activate this specific pathway. When they tested it in human cells, the molecule successfully turned on the protective genes, confirming that the AI had correctly identified a functional solution.

The most significant test involved a disease called fibrosis, a condition where healthy tissue becomes scarred and stiff, often leading to organ failure. The researchers defined the desired outcome as a return to a healthy, non-scarred state. They asked GEMGen to design chemicals that could reverse the genetic signature of fibrosis. The system generated a list of candidates, including some that were completely new to science. The team synthesized seventeen of these novel molecules from scratch. In cell experiments, eight of them successfully reduced the markers of scarring, a success rate that is remarkably high for new chemical discoveries.

One of these new molecules, named GEMGen-010, was selected for further study. The researchers tested it in mice that had been induced to develop liver fibrosis. The treatment was administered daily for two weeks. The results showed a clear reduction in the scarring of the liver tissue. The treated mice had lower levels of liver damage markers in their blood and showed a reversal of the genetic changes associated with fibrosis. Crucially, this molecule had not been modified or optimized after its initial generation; it worked exactly as the AI had designed it.

This work demonstrates that it is possible to use artificial intelligence to design chemical interventions that are not just variations of existing drugs, but entirely new structures capable of producing specific biological outcomes. By treating the desired state of a cell as a set of instructions, the system can navigate the vast space of possible chemical structures to find solutions that have never been seen before. While the study focused on specific conditions like liver fibrosis and stress responses, the method suggests a new path for medicine: one where the goal is defined first, and the tools to achieve it are invented to fit that goal.

Technical Summary: Inverse Design of Chemical Perturbations with GEMGen

Problem Statement
Perturbation biology traditionally focuses on "forward" modeling: predicting the cellular state resulting from a given genetic or chemical intervention. However, many biomedical applications, such as drug discovery and disease reprogramming, require the "inverse" problem: identifying specific perturbations capable of inducing a desired cellular state transition. Existing approaches typically rely on screening predefined compound libraries, which limits the search space to known chemical entities and fails to explore the vast, unenumerated chemical space (estimated at $10^{60}$ structures). Furthermore, current generative models often struggle to generalize across unseen chemical structures, experimental conditions, or perturbation modalities (e.g., translating genetic signatures to chemical interventions).

Methodology: The GEMGen Framework
The authors introduce GEMGen (Gene Expression-guided Molecular Generator), a large language model (LLM) framework designed to solve the inverse design problem by generating chemical perturbations conditioned on desired transcriptomic state transitions.

Architecture and Training:
- Foundation Model: GEMGen is built upon NatureLM, a 1-billion-parameter scientific foundation model trained on heterogeneous data including natural language, scientific literature, and structured biological databases.
- Instruction-Response Format: The framework converts the inverse design task into an instruction-response format. The input is a text-based, rank-ordered list of differentially expressed genes (upregulated and downregulated) representing the transition from an initial to a target cellular state. The output is a generated SMILES string representing a candidate molecule.
- Training Tasks: The model undergoes multi-stage fine-tuning:
  - Bioentity Representation: Learning descriptions of genes and cell types to establish a shared semantic representation of biological entities.
  - Chemical Perturbation Generation: Trained on the LINCS database (and L1000CDS²), where transcriptomic signatures are transformed into text-based gene lists to condition molecular generation.
  - Compound-Response Scoring: A separate module is fine-tuned to act as a binary classifier, scoring the likelihood that a generated molecule reproduces a specific transcriptional response.
Data Representation Strategy:
- Instead of using continuous numerical fold-change values, GEMGen utilizes text-based, rank-ordered gene lists. The authors demonstrate that this abstraction improves robustness to technical variations (e.g., batch effects across experimental plates) by grouping samples based on compound identity rather than numerical magnitude, while preserving the direction of transcriptional response.
Workflow:
- Generation: Given a target state transition, GEMGen autoregressively generates candidate molecules.
- Filtering: Candidates are filtered for syntactic validity and drug-likeness (e.g., Lipinski's Rule of Five).
- Scoring: The compound-response scorer prioritizes molecules predicted to match the target transcriptional signature.
- Output: A focused list of experimentally tractable candidates, including de novo structures not present in training libraries.

Key Contributions

Inverse Design via LLMs: The paper presents a novel application of scientific LLMs to directly generate chemical structures from transcriptomic instructions, bypassing the need for predefined compound libraries.
Cross-Modal Generalization: The framework is shown to generalize across different perturbation modalities, successfully using genetic perturbation signatures (CRISPRi) as inputs to design chemical interventions.
Prospective Experimental Validation: Unlike many generative models validated only in silico, GEMGen is prospectively validated through the synthesis and testing of de novo generated molecules.

Results
The authors evaluated GEMGen across three progressively challenging settings:

Recovery of Known Perturbations (In Silico Benchmarks):
- Novel Compounds: In a test set of 404 LINCS samples with compounds structurally distinct from training data, GEMGen recovered benchmark compounds or mechanistic analogs for 50% of samples (203/404) at a Tanimoto similarity threshold of 0.6.
- Novel Conditions: When tested on the Tahoe-100M single-cell dataset (640 samples), the model generalized from bulk training data to single-cell contexts, recovering hits for 37.8% (242/640) of samples.
- Mechanistic Analogs: When including mechanistic analogs (compounds sharing targets/pathways but with distinct scaffolds), recovery rates increased significantly (up to 85% in the novel compound set).
- Novelty: Approximately 61% of generated compounds were absent from both the training set and the ZINC 11M reference library, demonstrating the ability to explore unenumerated chemical space.
Cross-Modal Design (Genetic to Chemical):
- Using CRISPR interference signatures for 21 genes, GEMGen generated chemical perturbations that phenocopied the loss-of-function states.
- Case Study (KEAP1/NRF2): For the KEAP1 knockdown signature, the model identified flavokawain C, a chalcone derivative not previously reported as an NRF2 activator. Experimental validation confirmed it dose-dependently induced NRF2 target genes (e.g., NQO1) and stabilized NRF2 protein levels, validating the model's ability to translate genetic perturbations into chemical solutions.
Disease-Relevant Inverse Design (Fibrosis):
- Repurposing: The model identified existing molecules (HG-9-91-01 and indirubin-3′-monoxime) with previously unreported antifibrotic activities in TGFβ-stimulated LX-2 cells.
- De Novo Synthesis: From 17 de novo synthesized candidates (structurally distinct from known libraries), 8 compounds (47%) exhibited measurable antifibrotic effects, reducing fibrosis markers (FN1, Acta2).
- Lead Candidate (GEMGen-010): A quinazoline-2,4-diamine derivative, GEMGen-010, demonstrated sub-micromolar potency in vitro. In a CCl₄-induced mouse model of liver fibrosis, intraperitoneal administration of GEMGen-010 significantly reduced collagen deposition, lowered serum liver injury markers (ALT/AST), and suppressed fibrosis-associated gene expression, all without prior structural optimization.

Significance and Claims
The paper claims that GEMGen establishes a practical framework for inverse perturbation design, shifting the paradigm from predicting the consequences of interventions to designing interventions that produce specified biological outcomes. Key claims include:

Generalization: The text-based abstraction of state transitions allows the model to generalize across bulk and single-cell data and across perturbation modalities (genetic to chemical).
Actionability: The framework produces experimentally actionable candidates, evidenced by the high success rate of de novo synthesized molecules in cellular and in vivo assays.
Beyond Libraries: By generating molecules directly rather than searching libraries, GEMGen expands the scope of perturbation biology to explore chemical space beyond the constraints of existing collections.
Modesty: The authors acknowledge limitations, noting that transcriptomic similarity does not fully capture downstream physiological or safety outcomes, and that the model's performance is dependent on the diversity of available perturbation-response training data. They emphasize that the current implementation focuses on chemical generation but the framework is theoretically extensible to other modalities.

More like this