In-silico 3D molecular editing through physics-informed and preference-aligned generative foundation models

Generating molecular structures towards desired properties is a critical task in computer-aided drug and material design. As special 3D entities, molecules inherit non-trivial physical complexity, and many intrinsic properties may not be learnable through pure data-driven approaches, hindering the transaction of powerful generative artificial intelligence (GenAI) to this field. To avoid existing molecular GenAI's heavy reliance on domain-specific models and priors, in this research, we derive theoretical guidelines to bridge the methodological gap between GenAI for images and molecules, allowing pre-training of foundation models for 3D molecular generation. Difficulties due to symmetry, stability and entropy, which are critical for molecules, are overcome through a simple and model-agnostic training protocol. Moreover, we apply physics-informed strategies to force MolEdit, a pre-trained multimodal molecular GenAI, to obey physics laws and align with contextual preferences, and thus suppress undesired model hallucinations. MolEdit can generate valid molecules with comprehensive symmetry, strikes a better balance between configuration stability and conformer diversity, and supports complicated 3D scaffolds which frustrate other methods. Furthermore, MolEdit is applicable for zero-shot lead optimization and linker design following contextual and geometrical specifications. Collectively, as a foundation model, MolEdit offers flexibility and developability for AI-aided editing and manipulation of molecules serving various purposes.