A new system called CellForge uses a team of autonomous AI agents to build virtual-cell models from raw biological data. Given a dataset and a task, the agents survey the literature, parse the data, identify baselines, and debate a model hypothesis through critic-mediated discussion among specialists in data modeling, single-cell biology, and deep learning. They then write the training code, run experiments, and return a trained predictor with architecture, code, and validation results — all without a human specifying the design or training protocol.

The architecture itself is treated as the hypothesis, and CellForge produces it end-to-end. The researchers tested the system on six perturbation datasets spanning transcriptomic, epigenomic, and proteomic contexts. The resulting models were competitive with hand-crafted state-of-the-art results on established tasks, and the same workflow extended to paired multi-omic readouts.

Blinded human evaluators rated CellForge's research plans as more scientifically sound and innovative than those from single LLMs or deep-research systems. The authors conclude that autonomous agents can complete the full hypothesis-to-validation cycle in single-cell biology — effectively having an AI build another AI in a setting that normally requires a team of human experts.