When label-free morphology is sufficient for targeted cellular measurements

How cells respond to genetic perturbations reveals gene function, but each imaging experiment can measure only a limited set of molecular markers. Predicting additional readouts from label-free images could expand what these experiments reveal, but when such predictions can replace direct measurements remains unclear. Here we introduce MorphoSuff, a framework for assessing whether predicted cellular readouts support specific biological tasks. It evaluates reference re-producibility and repeats the intended analyses with predicted rather than measured responses. Rules applied to calibration perturbations select uses for testing on separate, held-out perturbations. Using nearly 10 million same-cell phase–reporter observations spanning 1,000 gene knock-outs, 52 fluorescent reporters and 73 A549 screens, we find information about reporter states beyond average perturbation responses and cell size, shape and density. Disrupting same-cell pairing within perturbation and screen groups during training reduced the median correlation between predicted and measured cell phenotypes from 0.76 to 0.29. Predictions often preserved perturbation rankings while compressing response magnitudes. Within the atlas, ranking-based acceptance supported more successful prioritization uses than acceptance based on Pearson correlation alone. In primary hepatocytes, brightfield predictions recovered 10 of the 11 strongest metabolic-activity losses among 217 held-out compounds, compared with 8 of 11 for a matched Ridge predictor using fluorescence-derived cell counts. MorphoSuff evaluates predictors by the scientific uses they support, using reference reproducibility and held-out task outcomes to assess whether inferred measurements are adequate.