Evaluation of Custom REINVENT Priors Trained on Synthetically Accessible Chemical Spaces for Generative Molecular Design
Тип публікації :
Препринт
Дата випуску :
9 березня 2026 р.
Автор(и) :
Sergey Sosnin
Oleksandr Zhadovets
Anna Kapeliukha
Mariana Perebyinis
Olha Semenenko
Mykola Protopopov
Olga O. Tarkhanova
eKNUTSHIR URL :
Журнал :
ChemRxiv
Цитування :
[APA 7] Sergey, S., Oleksandr, Z., Anna, K., Mariana, P., Olha, S., Mykola, P., & Olga, O. T. (2026). Evaluation of Custom REINVENT Priors Trained on Synthetically Accessible Chemical Spaces for Generative Molecular Design. ChemRxiv,. https://doi.org/10.26434/chemrxiv.15000188/v2
[ДСТУ] Evaluation of Custom REINVENT Priors Trained on Synthetically Accessible Chemical Spaces for Generative Molecular Design / S. Sergey та ін. ChemRxiv. 2026. DOI: 10.26434/chemrxiv.15000188/v2 (дата звернення: 11.09.2026).
Generative molecular design is increasingly popular as a promising way to create novel drug candidates with desired properties. The REINVENT approach, originally proposed by AstraZeneca, is a widely used framework for such generative design. However, the synthetic accessibility of generated molecules often remains uncertain. Combinatorial spaces provide access to hundreds of billions of molecules with guaranteed synthetic feasibility and a flat pricing model, making them attractive for planning and executing drug-discovery projects. This study addresses how generative molecular design can be used to ensure that generated molecules remain within a desired chemical space. To achieve this, we developed a new approach and demonstrated the feasibility of training REINVENT Priors on samples from Freedom Space 4.0 and REAL Space. We evaluated the performance of finetuned REINVENT Agent models derived from these Priors using both structure-based and ligand-based computational objectives. Our approach allows the generation of molecules that exhibit the desired activity and remain within the combinatorial space, ensuring synthetic accessibility. The Priors developed in this work are freely released to the community. We hope that the outcomes of this research will boost the practical applications of generative molecular design in academia and industry all over the world.
Якщо не вказано інше, ця робота розповсюджується на умовах ліцензії Attribution 4.0 International