Abstract
The crystallographic challenge of structure determination is nowadays effectively supported by advanced computational methods, such as direct methods and Patterson techniques, implemented in sophisticated software. With the rapid expansion of artificial intelligence (AI) across diverse scientific domains, exploring its potential contribution to structure solution and its ability to overcome the limitations of traditional approaches has become increasingly compelling. This work builds upon and extends the findings of two recent studies on AI-driven phasing. The first, by Larsen et al. [Science (2024), 385, 522-528], focused on designing and applying a neural network architecture to solve small structures (with unit-cell volumes up to 1000 Å(3)), primarily within the most common centrosymmetric space group P2(1)/c. The second, by Carrozzini et al. [Acta Cryst. (2025), A81, 188-201], introduced a novel phase-seeding method applicable to both centro-symmetric and non-centrosymmetric crystal structures of varying complexity, from small molecules to proteins. Although designed with AI integration in mind, this latter method had not yet been tested within an AI framework. In this paper, we apply the method proposed by Carrozzini et al. to cases where seed phases are generated by the AI network developed by Larsen et al. We demonstrate that this combined approach, termed AI-PhaSeed, successfully extends the applicability of Larsen's neural network to structures with unit-cell volumes exceeding 1000 Å(3), even under conditions of limited experimental resolution. The proposed procedure has been extensively tested on a set of structures taken from the Crystallography Open Database, proving it to be a powerful and reliable tool for structure solution. We also provide insights into the use of AI for crystallographic phasing and introduce statistical tools to evaluate the robustness of the solution process based on AI-calculated phases.