Preeti
Deep neural networks have shown promise in accurately predicting molecular structures based on descriptors for various molecular systems. The DiG model aims to go beyond predicting the most probable structure by also generating diverse structures with probabilities under the equilibrium distribution. Inspired by the heating-annealing paradigm, DiG breaks down this challenge into simpler problems using a diffusion process. This process involves transforming the target distribution of a molecule towards a simple distribution through a time period, followed by a reverse diffusion process to transform it back. Deep neural networks are used to predict updates for the reverse process based on the given descriptors, processed into node and pair representations. The model leverages the Graphormer architecture for structure prediction.
Additionally, DiG can be trained using a pre-training algorithm called PIDP, which optimizes the model based on an initial set of candidate structures and the energy function of each system. The model can also be trained with molecular structure samples obtained from experiments or simulations. DiG can estimate the density function of the equilibrium distribution and enable property-guided structure generation by predicting conditional structural distributions based on microscopic properties. The model can also interpolate between states to approximate reaction paths and find intermediate states along the path. Overall, DiG offers a versatile approach to predicting molecular structures and properties efficiently and accurately.
Source link
Source link: https://www.nature.com/articles/s42256-024-00837-3
GIPHY App Key not set. Please check settings