CellForge: Agentic Design of Virtual Cell Models: Multi-Agent Framework Autonomously Designs Cell Models, Achieving Architectural Innovation Beyond Human Experts
In August 2025, CellForge proposed a multi-agent framework that automatically designs neural network architectures for single-cell data and perturbation tasks through collaborative reasoning. On six datasets (gene knockout, drug treatment, cytokine stimulation), the generated models are highly competitive with existing baselines, and discovered novel architectures such as trajectory-aware encoders and perturbation diffusion modules. Code is open-sourced.
This paper demonstrates the potential of multi-agent collaboration in scientific method design: unlike traditional hyperparameter search or single LLM prompting, CellForge autonomously generates executable, high-quality neural network architectures through iterative reasoning of multiple specialized agents (e.g., architecture search, data understanding, verification). This marks a paradigm shift from human design to AI-autonomous design of scientific methods, a milestone in computational biology.
CellForge comprises four core agents: Task Analyzer (parses data and objectives), Architecture Designer (proposes network structure), Code Generator (implements executable code), and Verifier (tests and provides feedback). They collaborate through structured dialogue to iteratively optimize architectures. On multimodal data such as scRNA-seq, scATAC-seq, and CITE-seq, the generated models achieve or exceed state-of-the-art methods in tasks like predicting gene expression changes and drug responses. A key innovation is the agents' ability to combine novel components, such as integrating attention mechanisms with diffusion processes to form perturbation diffusion modules, which were absent in human designs.
This framework can be generalized to other scientific fields (e.g., materials, chemistry, physics) for AI-driven experimental design and model development. For the biopharmaceutical industry, CellForge can accelerate the construction of target discovery and drug response prediction models. The open-source code lowers the barrier to entry, but requires computational resources and domain knowledge to customize agents.
Computational biology teams are advised to try the CellForge framework for automating model development processes, reducing manual trial and error. Pharmaceutical companies can integrate it into drug discovery pipelines to predict compound effects on specific cell types. The open-source nature allows internal customization, but engineering resources are required.
Focus on CellForge's performance on larger-scale datasets (e.g., Human Cell Atlas); efficiency and interpretability of agent collaboration; and whether it can generate fundamentally new architectures beyond existing paradigms. Additionally, the reproducibility and biological plausibility of generated models need evaluation.