Performance of ex vivo drug screen in predicting clinical responses in vivo. The association between ex vivo and in vivo responses are shown for 21 patients with available clinical response data. Ex vivo normalized blast counts corresponding to each patient’s administered drugs were transformed using a robust z score to produce a drug sensitivity score. (A) A receiver operating characteristic (ROC) curve showing performance of score in predicting response after treatment. Area under the curve, 0.87 (95% CI, 0.72-0.98). (B) Optimal cut point as determined by maximizing Youden’s J statistic. (C) Violin plots comparing drug sensitivity scores between clinical responders (orange) and nonresponders (blue). Concordance (light gray) or discordance (dark gray) between ex vivo and in vivo responses are indicated. Dashed line represents optimal cut point determined in panel B. (D) Violin plots comparing normalized blast counts (without z-score transformation) between clinical responders (orange) and nonresponders (blue). Concordance (light gray) or discordance (dark gray) between ex vivo and in vivo responses indicated are based on predictions from panel C. Performance metrics for discovery data set: mean PPV, 0.92 (95% CI, 0.69-1.00); mean NPV, 0.82 (95% CI, 0.58-1.0); mean accuracy, 0.85 (95% CI, 0.71-0.95).