سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / اولین همایش بین المللی تحول خدمات سلامت: تلفیق مراقبت، درمان، بازتوانی و امیدبخشی

Generative AI for Multi-Objective ADMET Optimization in De Novo Drug Design: A Systematic Review of Molecular Generation Strategies and Evaluation Metrics

نویسندگان :

Shabnam Shabankare¹ Alireza Zarei²

1- موسسه آموزش عالی آپادانا 2- دانشگاه بین المللی امام رضا علیه السلام

کلمات کلیدی :

Artificial Intelligence،Deep Learning،Drug Discovery،Drug Design،Pharmacokinetics،Toxicity Tests

چکیده :

Abstract: Background: Generative Artificial Intelligence is driving a paradigm shift in de novo drug design. While these models efficiently traverse vast chemical spaces ({10}^{60} potential compounds), the primary bottleneck remains late-stage clinical attrition due to suboptimal Pharmacokinetics and Toxicity. This review evaluates the intersection of generative architectures and multi-objective ADMET optimization to ensure clinical viability. Method: Adhering strictly to PRISMA 2020 guidelines, a systematic search was conducted across PubMed, Scopus, IEEE Xplore, and Web of Science targeting literature from 2025-2026. Following rigorous screening of 600 initial records, 23 foundational studies focusing on molecular generation strategies, ADMET algorithms, and evaluation metrics were included for narrative synthesis. Finding: The literature reveals a transition from 1D sequence-based models to sophisticated 3D Equivariant Graph Neural Networks (E-GNNs) and Diffusion models. For multi-objective ADMET optimization, conditional generation currently outperforms Reinforcement Learning by mitigating “reward hacking” when balancing multiple pharmacokinetic constraints. Furthermore, synthesizing mathematically optimized compounds necessitates strict evaluation metrics; integrating SA Scores, QED, and AI-driven retrosynthetic pathway analysis is now mandatory to bridge the gap between in silico generation and in vitro feasibility. Conclusion: Generative ADMET frameworks have evolved from standalone mathematical generators into indispensable computational co-pilots. By integrating physical priors and rigorous synthesizability constraints, these systems are establishing “closed-loop” autonomous pipelines poised to drastically reduce clinical attrition rates and accelerate hit-to-lead optimization.