Abstract
<jats:p>Determining protein structures and dynamics has been a central pursuit in biomedical research for decades. However, the intrinsic complexity of protein systems and the limited ability of experimental techniques to capture transient conformational states have made molecular dynamics (MD) simulations indispensable for exploring protein conformational landscapes and elucidating structure–function relationships at the atomic level. Despite their broad utility, MD simulations are fundamentally limited by high computational cost and dependence on the accuracy of empirical force fields. To address these challenges, we introduce TSS-TDM, a diffusion-based protein trajectory sampling framework that integrates spatial and temporal attention to model protein conformational dynamics. TSS-TDM enables efficient and unbiased exploration of protein conformational space and generates protein transition pathways between distinct functional states. The performance of TSS-TDM was validated across three representative molecular systems: alanine dipeptide, a well-established model system for evaluating conformational sampling; the Nuclear Coactivator Binding Domain (NCBD), a molten-globule-like Intrinsically Disordered Domain (IDD) that exhibits multiple conformational states; and T4 lysozyme, an enzyme that undergoes hingebending motions between open and closed conformations. Upon validation of the generated trajectories, TSS-TDM outperforms its predecessor, GeoTDM. Notably, although TSS-TDM is trained without explicit energy information, it generates structures with OpenMM potential energy profiles that are closer to those observed in MD simulations and more stable than those generated by GeoTDM, indicating improved structural stability. In addition, TSS-TDM reproduces Ramachandran conformational distributions that are highly consistent with those sampled by real MD trajectories. For NCBD and T4 lysozyme, parallel sampling enables TSS-TDM to explore frontier conformational space in a layer-by-layer and unbiased manner, while trimming sampling allows efficient generation of transition pathways between distinct functional states. Benchmarking results demonstrate that TSS-TDM achieves substantially faster sampling than conventional MD simulations while maintaining an extremely low atomic clash rate in the generated structures. Overall, TSS-TDM provides an efficient alternative to MD simulation for unbiased protein trajectory generation and conformational space exploration.</jats:p>