Fig. 4: Experimental screening for PIM1 and CDK1. | Nature Machine Intelligence

Fig. 4: Experimental screening for PIM1 and CDK1.

From: Molecular deep learning at the edge of chemical space

Fig. 4: Experimental screening for PIM1 and CDK1.The alternative text for this image may have been generated using AI.

Compound selection was based on unfamiliarity scores and uncertainty estimates averaged over 10-fold Monte Carlo cross-validation (10% validation samples). All kinase activity measurements represent the mean of three technical replicates, except for positive control compounds (AZD1208 and dinaciclib). a, The ten most promising PIM1 inhibitors were selected from a library of ~180,000 compounds using three combinations of uncertainty and unfamiliarity (A: most uncertain and least unfamiliar B: least uncertain and least unfamiliar; C: least uncertain and most unfamiliar). b, Measured PIM1 activity across selected molecules at 10 µM and their maximum Tanimoto similarity (on ECFPs) to PIM1 training molecules. Lower PIM1 activity means stronger inhibition. c, Box plot of measured PIM1 activity (n = 10 molecules per method). The solid line represents PIM1 activity without any screening compound, while the dashed line represents PIM1 activity with a potent control inhibitor (AZD1208). Statistically significant differences (α = 0.05) are denoted with an asterisk, and were determined by paired, two-sided, Wilcoxon signed-rank tests. Box plots show the median (centre line), 25th and 75th percentiles (box bounds) and 1.5× IQR (whiskers). d, The ten most promising CDK1 inhibitors for each selection method. e, Measured CDK1 activity across selected molecules at 10 µM and their maximum Tanimoto similarity (on ECFPs) to CDK1 training molecules. f, Box plot of measured CDK1 activity (n = 10 molecules per method). The solid line represents CDK1 activity without any screening compound, while the dashed line represents CDK1 activity with a potent control inhibitor (dinaciclib). Box plots show the median (centre line), 25th and 75th percentiles (box bounds) and 1.5× IQR (whiskers). Statistically significant differences (α = 0.05) are denoted with an asterisk, and were determined by paired, two-sided, Wilcoxon signed-rank tests. g, Measured protein activity across all 60 screened compounds at 10 µM (method A: 1–10 and 31–40; method B: 11–20 and 41–50; method C: 21–30 and 51–60). Selected compounds (identifier highlighted in boldface) displayed in panel h) were further characterized for their dose–response curve. h, Structures and determined IC50 of the six most promising compounds for PIM1 and CDK1. IC50 values that could not be determined are denoted as NA.

Source data

Back to article page