Next-Gen AI Models for 3D Molecular Generation and Macrocycle Design

Generative chemical models are becoming powerful tools in drug discovery, enabling multi-objective molecular optimization and accelerating preclinical research. The open source REINVENT4 ecosystem, developed by AstraZeneca in Sweden, is a leading example and has supported the discovery of novel molecules across diverse drug targets in both industrial and academic settings. Through a previous EuroHPC grant (EHPC-AI-2024A06-040), the researchers advanced generative chemical modeling through open science on LUMI. This effort has already produced one paper accepted at ICML 2026, three preprints under review, and one additional manuscript in preparation. Building on this, the proposed project will focus on two high-impact areas: 3D molecular generation and macrocycle design. 

For 3D molecular generation, the team will build on FlexiFlow and a controllable flow-matching model, both developed during our previous EuroHPC allocation. Designing molecules with their 3D conformational ensembles conditioning on flexible pockets will enable the generation of novel small molecules that capture multiple three-dimensional configurations. In traditional structure-based drug design, protein-ligand interactions are often modeled from static crystallographic poses. This assumption neglects induced fit effects, i.e., the 3D changes of the protein when different ligands bind to it, and the dynamic nature of binding sites; accounting for these effects could improve protein-ligand complementarity and support the design of molecules with enhanced binding properties. Furthermore, controllable generation using reinforcement learning is essential, as it enables the model to be guided toward optimizing specific properties. In this project, the researchers will extend FlexiFlow in three directions: To scale model training to larger datasets and larger molecular systems;  to extend the model to allow for the design of ligands while accounting for binding site flexibility; and to enable controllable molecular generation, allowing the direct generation of 3D molecules with specified properties within the FlexiFlow framework. 

For macrocycle design, the team will build on OmniInvent (manuscript in preparation), a generalist drug design model also developed through our previous EuroHPC allocation. Macrocycles are an increasingly important pharmaceutical modality, but current AI-assisted design methods remain immature: they typically support only a single design strategy, most often de novo design, and generally lack the ability to generate molecules with user-defined properties. This capability is particularly important in pharmaceutical development, where compounds must be iteratively refined to achieve a balanced profile across multiple properties. The team will develop MacroInvent, an extension of OmniInvent tailored to macrocycles, as a unified system for both de novo macrocycle design and macrocyclization. Through integration with the REINVENT4 molecular optimization framework, this will enable the design of property-tailored macrocyclic molecules and expand the applicability of generative AI to a challenging and strategically important class of compounds.