ENHANCING FULL-LENGTH DDX3X CRYSTALLIZATION: A CASE STUDY ON MACHINE LEARNING APPLICATIONS IN STRUCTURAL BIOLOGY

DDX3X is an ATP-dependent RNA helicase from the DEAD-box family, essential for RNA metabolism and associated with neurodevelopmental disorders. In this study, we aim to crystallize the full-length human DDX3X, a protein containing intrinsically disordered regions in both N- and C-termini, in complex with double-stranded RNA and ADP, with the goal of determining the 3D structure of these complexes by X ray diffraction. The recombinant protein was expressed in E. coli and purified through cobalt ion-affinity chromatography followed by size-exclusion chromatography. Dynamic light scattering (DLS) experiments confirmed a monodisperse protein sample with a hydrodynamic radius of 4.3 ± 0.5 nm and a polydispersity index of 13%, indicating suitability for crystallization trials. Initial crystallization screening was performed using the vapor diffusion method in 96-well plates, with commercial crystallization kits and a crystallization robot for high throughput screening. Crystal hits were observed in 49 and 35 out of the 576 tested conditions for DDX3X:ADP and DDX3X:ADP:RNA, respectively, as confirmed by UV imaging. We applied machine learning algorithms (Random Forest), combined with SHAP (SHapley Additive exPlanations) analysis, to identify key physicochemical variables and prioritize conditions most likely to yield appropriated crystals for diffraction experiments. This integrated approach accelerates the refinement of crystallization conditions and reduces the number of non-productive experiments. We also observed that specific conditions, particularly those containing buffer with pH ranging from 4.5 to 6.5, tend to promote the formation of biomolecular condensates through liquid–liquid phase separation (LLPS) in the absence of RNA.