Diffusion models have demonstrated remarkable generative capabilities across diverse domains, including computer vision and structural biology. However, achieving fine-grained control and precise steering over their generative trajectories remains an open challenge, particularly in complex domains such as de novo protein design. To address this, this research investigates mechanistic interpretability methods to guide generative dynamics through the analysis and manipulation of internal model activations. We will systematically evaluate steering mechanisms across visual and biological representations to identify universal control principles that generalize across modalities. The proposed framework will provide unified tools for cross-domain generative control, culminating in the experimental laboratory validation of computationally designed proteins.