Accelerating Diffusion-Based Graph Generative Models for De Novo Drug Design via Hessian Trace Approximation

Diffusion-based models have recently gained prominence in various domains, including graph-structured data and de novo drug design, where they enable the generation of novel molecular structures and optimization of pharmaceutical candidates. In particular, methods such as “Digress” (Vignac et al. ۲۰۲۳) effectively incorporate diffusion processes to model complex interactions within graph data, while Graph Diffusion Policy Optimization (GDPO) (Liu et al. ۲۰۲۴) extends this idea by integrating policy optimization strategies to achieve higher-quality solutions. Compared to Digress, GDPO typically demonstrates higher convergence rate and more robust performance which makes it potentially better solution for advancing graph-based drug design problem and related applications. However, both Digress and GDPO rely heavily on gradient-based optimization, which is not as fast as necessary. On the other hand, computing the Hessian matrix directly to make use of second-order methods and follow the curvature characteristics is computationally expensive. To address these issues, we introduce a novel idea to approximate the Hessian matrix with relatively low cost to better guide the optimization process, providing a more accurate and efficient estimation of curvature that leads to improved directions and lower number of iterations. We trained our model on the ZINC-۲۵۰k dataset, a widely used collection of small molecules, and compared its performance with Digress and GDPO. Our approach demonstrates enhanced efficiency and superior performance over the existing diffusion-based generative models in de novo drug discovery.

کلیدواژه ها:

diffusion-based models ، drug discovery ، policy optimization ، hessian approximation ، generative models