How can diffusion models process 3D geometries in a coarse-to-fine manner, akin to our multiscale view of the world?

Predicting low-energy molecular conformations given a molecular graph is an important but challenging task in computational drug discovery. Existing state-of-the-art approaches either resort to large scale transformer-based models that diffuse over conformer fields, or use computationally expensive methods to generate initial structures and diffuse over torsion angles.

Code will be released upon acceptance.

In particular, torsional diffusion does not address the longstanding difficulty that existing cheminformatics methods have with macrocycles--rings with 12 or more atoms that have found several applications in drug discovery [Driggers et al., 2008].

Accurate prediction of likely 3D geometries of flexiblemolecules is along standing goal of computational chemistry, with broad implications for drug design, biopolymer research, and QSAR analysis.

Let's assume we apply a random CCW torsion rotation of angle We detail here the formulae used in section section 2.4. Similar to AlphaFold [Senior et al., 2020], we fit distances using normal distributions and angles Such cases require a special treatment. So far, we haven't tackled the following difficulty: Examples are hydrogen groups as in Figure 1. We propose a new loss function based on eq. The EMD computation cannot be parallelized in mini-batches in the current version of the library, but everything else is batch-parallelizable in our model (e.g., The training stage happens without assembling the full conformer.