AI/ML for Autonomous Control and Atomistic Modeling

Understanding protein conformation is crucial for gaining insights into their functional mechanisms. Proteins can adopt a wide range of conformations, from well-defined folded states to more dynamic and disordered unfolded regions. However, accurately capturing the conformational diversity of both folded and unfolded regions is a significant challenge for many popular coarse-grained (CG) force fields, which often oversimplify the intricate interactions governing protein structure. These limitations can lead to inaccurate representations of protein behavior, especially in cases where both structural flexibility or stability are essential.

To overcome these challenges, we developed CGDiFF, a Diffusion-model-based foundation model for protein CG force field parameterization. CGDiFF builds upon a recently developed deep generative protein ensemble sampler, which leverages state-of-the-art machine learning techniques to generate diverse protein conformations guided by experimental data. Powered by a simple fine-tuning process, CGDiFF can automatically parameterize a CG force field tailored to any protein sequence, effectively bridging the gap between folded and unfolded regions.

Such simple fine-tuning process makes CGDiFF adaptable for various protein systems and reducing the need for manual adjustments. CGDiFF represents a significant advancement in the field, providing a powerful tool to enhance the accuracy of CG simulations across a wide range of protein structures, from highly ordered domains to disordered regions critical for biological function.