Kim Hyunho, Bae Bongsung, Park Minsu, Shin Yewon, Ideker Trey, Nam Hojung
Department of Electrical Engineering and Computer Science, Gwangju Institute of Science and Technology, Gwangju, Republic of Korea.
Division of Advanced Predictive Research, Korea Institute of Toxicology, Daejeon, Republic of Korea.
Nat Commun. 2025 Jul 1;16(1):5628. doi: 10.1038/s41467-025-60763-9.
Despite advances in precision oncology, developing effective cancer therapeutics remains a significant challenge due to tumor heterogeneity and the limited availability of well-defined drug targets. Recent progress in generative artificial intelligence (AI) offers a promising opportunity to address this challenge by enabling the design of hit-like anti-cancer molecules conditioned on complex genomic features. We present Genotype-to-Drug Diffusion (G2D-Diff), a generative AI approach for creating small molecule-based drug structures tailored to specific cancer genotypes. G2D-Diff demonstrates exceptional performance in generating diverse, drug-like compounds that meet desired efficacy conditions for a given genotype. The model outperforms existing methods in diversity, feasibility, and condition fitness. G2D-Diff learns directly from drug response data distributions, ensuring reliable candidate generation without separate predictors. Its attention mechanism provides insights into potential cancer targets and pathways, enhancing interpretability. In triple-negative breast cancer case studies, G2D-Diff generated plausible hit-like candidates by focusing on relevant pathways. By combining realistic hit-like molecule generation with relevant pathway suggestions for specific genotypes, G2D-Diff represents a significant advance in AI-guided, personalized drug discovery. This approach has the potential to accelerate drug development for challenging cancers by streamlining hit identification.
尽管精准肿瘤学取得了进展，但由于肿瘤异质性以及明确的药物靶点有限，开发有效的癌症治疗方法仍然是一项重大挑战。生成式人工智能（AI）的最新进展提供了一个有前景的机会，通过基于复杂基因组特征设计类似苗头的抗癌分子来应对这一挑战。我们提出了基因型到药物扩散（G2D-Diff），这是一种生成式人工智能方法，用于创建针对特定癌症基因型定制的基于小分子的药物结构。G2D-Diff在生成满足给定基因型所需疗效条件的多样、类药物化合物方面表现出卓越性能。该模型在多样性、可行性和条件适应性方面优于现有方法。G2D-Diff直接从药物反应数据分布中学习，无需单独的预测器即可确保可靠的候选物生成。其注意力机制提供了对潜在癌症靶点和途径的见解，增强了可解释性。在三阴性乳腺癌案例研究中，G2D-Diff通过关注相关途径生成了似是而非的类似苗头的候选物。通过将逼真的类似苗头分子生成与针对特定基因型的相关途径建议相结合，G2D-Diff代表了人工智能指导的个性化药物发现的重大进展。这种方法有可能通过简化苗头识别来加速针对具有挑战性癌症的药物开发。
