This comment evaluates dela Cruz-Chuh et al.'s AI-based label-free workflow for analyzing T-cell mediated tumor killing via brightfield imaging.The study's core strengths include eliminating fluorescent labeling artifacts, achieving comparable consistency to conventional segmentation-based methods, and accommodating phenotypically diverse cancer cells without manual parameter tuning-addressing key bottlenecks in immunotherapy screening.However, critical considerations persist: the binary "killing/non-killing" classification framework may insufficiently resolve lowlevel or partial cytotoxicity, as evidenced by six false negatives; generalizability to nonadherent hematological malignancies or alternative effector cells remains untested; and the model lacks interpretability regarding morphological features driving cytotoxicity predictions.Additionally, performance across variable imaging platforms or culture conditions is unevaluated, limiting translational reproducibility.Despite these gaps, the workflow advances high-throughput immunotherapy screening efficiency.Future studies incorporating multi-class training, validation across diverse cancer types, and mechanistic decoding of AI-predicted features will strengthen its rigor and broader applicability in drug discovery.
Liu et al. (Sun,) studied this question.