Overcoming cisplatin resistance via deep learning-assisted de novo design of platinum complexes

Resistance to cisplatin poses a significant challenge in cancer treatment for patients globally. Despite the continuous efforts of both the pharmaceutical industry and academia, no new platinum drugs have been approved for the treatment of cancers that have developed resistance to cisplatin-based chemotherapy regimens. Even in the new era of AI, the potential to accelerate the discovery of metal-based anticancer therapeutics has remained largely unexplored. Here, we present the first implementation of deep learning models for activity prediction of de novo designed platinum anticancer agents. Our approach overcomes historic limitations through a manually curated dataset of 4,078 platinum complexes with 18,995 IC50 values, as well as specialized algorithms resolving metal-specific representation challenges. For experimental validation, we predicted the likelihood of potential anticancer activity for 214,373 previously published platinum complexes in MCF-7, A549, A2780 and A2780cis cancer cell lines, synthesized and tested 20 of them, demonstrating strong agreement between predicted and observed activity. Additionally, using de novo fragment assembly approach we designed and successfully synthesized a novel bis-carbene platinum(II) complex, PlatinAI. The experimentally determined cytotoxicity of PlatinAI in MCF-7, A549, A2780 and A2780cis cancer cell lines was in a perfect agreement with the model prediction. PlatinAI demonstrated superior activity compared to cisplatin in cisplatin-resistant cell lines and exhibited a distinct mechanism of action, not reliant on the formation of DNA adducts. This research sets the stage for the development of AI-assisted platforms that enable data-driven design of next-generation platinum therapies, offering improved efficacy and the ability to overcome resistance to current anticancer treatments.