Structure-Based Drug Design via Diffusion Models Guided by Non-Differentiable Metrics

This work introduces a diffusion-guided molecular generation approach that seamlessly incorporates non-differentiable metrics—such as empirical property scores or domain heuristics—into the sampling process. Unlike traditional gradient-based optimization, our method leverages a conditional diffusion prior and a gradient-free feedback loop to navigate chemical space effectively, producing candidate molecules with enhanced validity and alignment with multiple design objectives. This project highlights how generative diffusion frameworks can be adapted for realistic drug design scenarios where scoring functions are irregular, discontinuous, or otherwise non-differentiable.