AtlasAgent: Vision language model and Agent-guided Framework for Evaluation of Atlas-scale Single-cell Integration

As single-cell omics transitions into the era of AI-virtual cells (AIVC), where large-scale single-cell data integration becomes prevalent, the computational demands of integration evaluation emerge as critical scalability bottlenecks. Traditional integration evaluation pipelines, requiring metrics like k-nearest-neighbor batch effect test (kBET) and Local Inverse Simpsons Index (iLISI) employed by state-of-the-art scIB method, often demand large computational resources and long runtimes, making them infeasible for large scale integration studies. Herein, we present AtlasAgent, the first vision-language model (VLM)-powered and AI agent framework to accelerate atlas-scale integration evaluation at unprecedented speed and scale. We systematically evaluate batch correction quality, biological signal preservation and overcorrection risks using chain-of-thought reasoning in conjunction with few-shot and zero-shot prompting strategies. AtlasAgent completes evaluation within 32 seconds, in contrast to scIBs runtime of 5.55 hours in GPU, while identifying the scIB-determint best integration methods within the top-3 in 88.3% of the time, lowering evaluation time from hours to seconds while preserving alignment with domain expert reasoning. AtlasAgent pioneers the use of VLMs to realize scalable and rapid integration evaluation at atlas scale.