Cancer, even of the same type, shows significant variation in treatment effectiveness due to differences in genetic information among patients. In particular, treating difficult cancers like triple-negative breast cancer is challenging because the targets are not clearly defined, making it hard to expect sufficient results from existing treatments.
Research team led by Professor Nam Ho-jeong from the Gwangju Institute of Science and Technology (GIST) has developed the world’s first generative artificial intelligence (AI) model that analyzes the genotypes of cancer patients to propose personalized anticancer drug candidates, it announced on the 3rd. This breakthrough could present new solutions for difficult cancers that do not respond well to existing treatments, as well as enable personalized precision medicine.
Previous studies on developing anticancer drugs using generative AI have faced several limitations. In complex diseases like cancer, the treatment targets are often unclear, leading to limited effectiveness of the generated drugs, and reliance on specific data that is hard to acquire in clinical settings has lowered the practical applicability.
To overcome these limitations, the research team developed a generative AI model called G2D-Diff, which learned from approximately 1.5 million chemical structures and 1.2 million drug response data. By inputting the genomic information that can be obtained in actual clinical practice and the level of drug response targeted, it automatically designs optimized anticancer drug candidates.
G2D-Diff exhibited overwhelming performance across all performance metrics compared to existing generative AI models. Particularly, when compared to IBM's PaccMannRL, known as the highest-performing model, G2D-Diff demonstrated superior results in diversity, feasibility, and condition suitability.
In the 'condition suitability' evaluation item assessing how well the generated compounds match the input genotype conditions, existing models showed an average error rate of about 51%, while G2D-Diff recorded an average error rate of approximately 1%. The generated molecular structures showed an average similarity in drug likeness (quantitative estimate of druglikeness (QED)) and synthetic accessibility (SAS) that was 35-44% higher than that of existing models, demonstrating a higher potential for development into actual new drugs.
The research team validated the practical applicability of the G2D-Diff model by applying it to triple-negative breast cancer, a representative example of difficult cancers. The drug candidates generated by inputting the genetic mutation information of the patients precisely targeted core proteins such as PI3K, HDAC, and CDK that inhibit cancer cell proliferation.
Professor Nam Ho-jeong noted, "This research opens new possibilities for personalized medicine, and AI technology is expected to provide new hope for patients with difficult cancers."
The research findings were published online in the international journal Nature Communications on the 1st.
References
Nature Communications (2025), DOI: https://doi.org/10.1038/s41467-025-60763-9