Extended Data Fig. 6: Evaluation of transitive predictions using CLIPn integrated space (related to Figs. 4, 5).

(a) Comparison of the classification probability distribution from k-NN predictions using either raw profiles or CLIPⁿ embeddings for previously validated compounds in Fig. 4f. Likelihood was estimated based on the percentage of the 10 nearest neighbors sharing the same label as the predictions. P value = 5.8E-25, two-sided t test. (b) Visualization of original profiles for reference compounds (colored dots) and uncharacterized compounds selected for validation (crosses). Predicted confidence is shown as the percentage of the 10 nearest neighbors with the same labels as the predictions. (c) Experimental validation results for different categories. Intensities of function-specific markers in each assay are shown for DMSO (blue), reference compounds (orange), and predicted hits (green). P values are derived by comparing DMSO to each predicted hit using a one-sided Mann-Whitney U test. Compound 221019 was predicted to EGFR at high dose (10 uM) and proteasome at lower dose (1 uM). (d) Left: The number of known compounds in each predicted category from the ChEMBL database. Right: Tanimoto similarities based on MACCS features between two ChEMBL compounds and between ChEMBL compounds and experimentally validated hits.