Fig. 5: Features of dual ligands designed by the chemical language models (CLM).

a–c Predicted binding modes of the most active dual ligands 1, 6 and 7. (FXR: protein data bank (pdb) ID 6a60⁵³, sEH: pdb ID 5ali⁵², THRβ: pdb ID 1nax⁵⁴, PPARδ: pdb ID 5y7x). d Structural and pharmacophore comparison of the dual FXR/THRβ ligand 6 and the dual PPAR/sEH modulator 9 with the most similar ligands of the targets of interest annotated in ChEMBL (pretraining set). Similarity refers to Tanimoto similarity computed on Morgan fingerprints³⁴. Common substructures in the designs and most similar ligands are highlighted (FXR—red, THRβ—blue, violet—both and yellow—sEH, green—PPARδ, orange—both). e Violin plots of quantitative estimation of drug-likeness (QED)³⁸ and synthetic accessibility scores of the CLM designs were favorable and resembled the fine-tuning molecules. Stars represent the fine-tuning molecules and designed dual ligands, respectively, the lines represent the 1st and 3rd quartiles and the mean. Numbers of the ligands/designs for PPARδ: 5, sEH: 6, PPARδ/sEH: 12. f Violin plots of basic chemical features of the CLM designs resembling the fine-tuning molecules. Stars represent the fine-tuning molecules and designed dual ligands, respectively, the lines represent the 1st and 3rd quartiles and the mean. Numbers of the ligands/designs for PPARδ: 5, sEH: 6, PPARδ/sEH: 12. Source data are provided as a Source Data file.