AetherCell

AetherCell is a generative foundation model—described by its authors as a virtual-cell engine—for predicting transcriptomic responses to chemical and genetic perturbations and supporting in vivo drug discovery. Released as a March 2026 bioRxiv preprint by researchers at Sun Yat-sen University, including the State Key Laboratory of Ophthalmology at Zhongshan Ophthalmic Center, it aims to bridge the gap between simple cell-line screens and the complex biology of organoids, patients, and whole organisms.

The model unifies clinical RNA-seq and perturbation-assay data into a shared transcriptomic manifold, allowing it to predict responses to compounds and genetic perturbations it has not seen during training. A central goal is generalization across biological scales: AetherCell translates signals measured in cell lines to three-dimensional organoids and patient-derived systems, producing high-fidelity, whole-transcriptome predictions rather than narrow readouts.

AetherCell joins a wave of generative virtual-cell models (such as Lingshu-Cell and STATE) but distinguishes itself by emphasizing translation across biological scales and end-to-end validation: the authors report in vivo confirmation of repurposing candidates, including teriflunomide for dry eye disease and dabigatran for ulcerative colitis.

#Key Features

• Unseen-perturbation generalization: Predicts responses to compounds and genetic perturbations not present in training by embedding clinical RNA-seq and perturbation-assay data in a shared transcriptomic manifold.
• Cross-scale translation: Translates perturbation signals from simple cell lines to 3D organoids and patient-derived systems, supporting whole-transcriptome prediction across biological scales.
• Phenotype-knowledge mixture-of-experts: A mixture-of-experts strategy combining phenotype and knowledge signals enables precision drug repurposing.
• In vivo-validated discovery: Repurposing predictions were confirmed in animal models, identifying teriflunomide for dry eye disease and dabigatran for ulcerative colitis.

#Technical Details

AetherCell is a generative model that learns a foundational transcriptomic manifold from clinical RNA-seq and perturbation-assay data, enabling whole-transcriptome response prediction for chemical and genetic perturbations. For drug repurposing it applies a phenotype-knowledge mixture-of-experts strategy, and it is designed to map signals from cell lines onto more complex systems including patient-derived organoids and clinical cohorts. The authors validate predictions experimentally, reporting in vivo discovery of teriflunomide for dry eye disease and dabigatran for ulcerative colitis. Detailed architecture specifics, parameter counts, and benchmark tables are described in the preprint. Inference code and pretrained weights are released on GitHub and Hugging Face under a non-commercial research license (the AetherCell Research License), and the bulk RNA-seq pre-training corpus is archived on Zenodo under a CC-BY-4.0 license; training code is not part of the release.

#Applications

AetherCell is aimed at translational researchers and drug-discovery teams seeking to prioritize and repurpose therapeutics. By predicting whole-transcriptome perturbation responses and translating them across biological scales, it can support target identification, lead prioritization, and drug repurposing—particularly where organoid or patient-derived models are the relevant endpoint. The demonstrated ophthalmology and gastroenterology applications illustrate its intended use in disease-specific drug discovery.

#Impact

AetherCell stands out among generative virtual-cell models for coupling computational prediction with in vivo experimental validation of repurposing candidates, a step many perturbation models do not take. Its emphasis on cross-scale translation toward organoids and patients addresses a key gap between cell-line predictions and clinically relevant biology. Inference code, pretrained weights, and the pre-training corpus have been released—the code and weights under a non-commercial research license—though the lack of released training code means full reproducibility remains only partial.