Silent methodological failures in AI-assisted science: benchmarking large language model agents for virtual gene knockout analysis

Abstract Large language models (LLMs) are increasingly deployed as autonomous agents in computational biology, yet their reliability in executing complex scientific workflows remains poorly characterized. Here, we benchmarked five LLM agents (DeepSeek V4 Pro, GLM 5.1, HY3 Preview, Kimi K2.6, and Minimax M2.7) on virtual gene knockout analyses using four computational frameworks across two single-cell RNA-sequencing datasets. All agents successfully completed the assigned workflows, but their quantitative outputs varied substantially: execution times spanned a 31fold range (61 − 1,920 s), while cross-model coefficients of variation reached 142.1%. Notably, four of five agents (80%) exhibited “silent methodological failures,” generating computationally valid, error-free code that produced biologically implausible results. Among the evaluated agents, GLM 5.1 showed the most balanced performance across computational efficiency and biological plausibility. These findings demonstrate that successful code execution alone does not ensure scientific validity and highlight cross-model concordance and biological validation as important quality-control strategies for Artificial Intelligence (AI)-assisted computational biology workflows.