Fig. 4: Early assessment and adjustment of CHIR dose by supervised ML on 0–12 h bright-field image streams.

a Titration of CHIR concentration and duration for three batches. Each dot’s color represents the mean percentage of cTnT⁺ cells on day 12 (n = 3 wells for each condition). Correlation analysis was conducted between CHIR concentrations and durations for high-efficiency wells (percentage of cTnT⁺ cells ≥ 50%), with Spearman’s ρ value. b Typical bright-field image streams under different CHIR concentrations from 0 h to 12 h. A partially enlarged view is provided below each image. Scale bar, 1 mm. c Schematic overview of the CHIR dose assessment and adjustment by supervised ML on 0–12 h live-cell bright-field image streams. Upper panel: each whole-well image stream was represented by a 21-D feature. Middle panel: In the training stage, the logistic regression model was trained with pairs of bright-field image streams and their CHIR concentration label; in the testing stage, the trained model made predictions of CHIR concentration (low, optimal, high) for new bright-field images. Lower panel: misdifferentiated cells can be rescued by further intervention. d LDA on 0–12 h image streams represented by 21-D features. LDA linearly projects the 21-D features into a 2-D plane spanned by the most discriminant axes (denoted by LD1 and LD2). n = 384 wells. e Classification performance of the logistic regression model (using all the features) on the test set. Precision, recall, F1 score, and AUC were macro-averaged over the three categories. n = 116 wells. f Comparison of predicted Deviation Scores (upper panel) and true ΔCHIR Concentrations (lower panel) for each CHIR dose condition in a testing batch, with Pearson’s r value. n = 12 CHIR doses. g Classification performance of cross-batch validation with Pearson’s r value (under a CHIR duration of 24 h). The color of the dots represents different testing batches. n = 20 CHIR doses. h The effects of adjusting CHIR duration based on ML prediction in different cell lines and batches. The percentage of cTnT⁺ cells for wells under a preset CHIR condition (8 μM, 48 h) and under an optimized condition (8 μM, ML-selected duration) were shown. n = 2–6 wells. Statistical significance was determined by t-test, ***P < 0.001; ****P < 0.0001. i Effect of CHIR concentration switching on differentiation efficiency for three batches. In the middle of the CHIR treatment, the initial CHIR concentrations (shown on the x-axis) were switched to other concentrations (shown on the y-axis).