Machine learning–driven optimization of drug candidates remains a central challenge in medicinal chemistry, particularly when attempting to improve potency without relying on external scoring functions. In this manuscript, the authors show that incrementally fine-tuning chemical language models along a structure-activity series enables the de novo design of PPARγ and RORγ modulators that surpass existing ligands in activity, demonstrating that CLMs can internalize SAR relationships and long-range dependencies for prospective molecular optimization.
- Tim Hörmann
- Domenic Mayer
- Daniel Merk