AI Scientist System CellForge Autonomously Designs and Validates Virtual Cell Models, Matching Human Expert Performance
CellForge uses multi-agent collaboration to autonomously generate model architectures and code for single-cell perturbation prediction, with its research plans rated as more scientifically sound and innovative than those from single LLMs by human evaluators.
Researchers from Yale University and other institutions have released version 3 of CellForge, demonstrating that a team of autonomous agents can complete the full scientific research cycle from hypothesis to validation. For the complex task of single-cell perturbation modeling, the system generates executable model hypotheses and architectures end-to-end without humans specifying the design, loss function, or training protocol.
The core innovation of CellForge is that "the architecture itself is the hypothesis." Through critic-mediated discussion among agents specialized in data modeling, single-cell biology, and deep learning, it formulates a model hypothesis, then writes training code, runs experiments, and returns a trained predictor including architecture, code, and validation results. Across six perturbation datasets spanning transcriptomic, epigenomic, and proteomic contexts, CellForge yields models competitive with hand-crafted state-of-the-art results and extends the same workflow to paired multi-omic readouts.
Blinded human evaluators rated CellForge's research plans as more scientifically sound and innovative than those from single large language models or deep-research systems. For instance, in drug perturbation tasks, CellForge improved Pearson correlation by 20%; on sparse, high-dimensional scATAC-seq datasets, it achieved approximately a 16-fold gain in Pearson correlation on differentially expressed genes compared to the second-best baseline.
This study proves that AI can effectively build another AI in a complex setting that currently requires a team of human experts. However, current results are based on single runs, with each dataset producing only one architecture instance, indicating this is an existence proof rather than a guarantee of optimal design.