CoarsenConf is a novel architecture introduced for molecular conformer generation, a critical task in computational chemistry where the goal is to predict stable low-energy 3D molecular structures, or conformers, using a given 2D molecule. This is a crucial process for various applications, such as drug discovery and protein docking, which depend on accurate spatial and geometric characteristics.
The CoarsenConf employs the concept of coarse-graining, which simplifies a molecule’s representation by grouping the fine-grained atoms into individual coarse-grained beads. This process reduces dimensionality and enables autoregressive generation of conformer, which is done by generating all coordinates independently, unlike previous work in this field. Conditioning on 3D coordinates of previously generated subgraphs, this model generalizes better over chemically and spatially similar subgraphs. Moreover, unlike prior methods, the CoarsenConf can generate low-energy conformers directly from modeling atomic coordinates, distances, and torsion angles.
The architecture comprises various components. The encoder takes a fine-grained ground truth conformer, RDKit approximate conformer, and a coarse-grained conformer as inputs and outputs a variable-length equivariant CG representation. Additionally, Equivariant MLPs learn the mean and log variance of both the posterior and prior distributions. The posterior or prior is sampled and put into the Channel Selection module, where an attention layer learns the optimal pathway from CG to FG structure. Lastly, given the FG latent vector and the RDKit approximation, the decoder recovers the low-energy FG structure via autoregressive equivariant message passing. The whole model can be trained by optimizing the KL divergence of latent distributions and the reconstruction error of generated conformers.
Furthermore, the architecture employs an approach named Aggregated Attention to learn the optimal variable-length mapping from the latent CG representation to FG coordinates. In simple terms, the method aggregates 3D segments of FG information to form latent queries, enabling efficient transition from the latent CG representation to viable FG coordinates.
Molecular coarse-graining is a widely utilized technique in protein and molecular design, and its usage on a fragment-level or subgraph-level generation has proven significantly valuable in various 2D molecule design tasks. Therefore, the CoarsenConf uses a variable-length CG approach, which provides flexibility and supports any choice of coarse-graining. This approach implies a single CoarsenConf model can generalize to any coarse-grained resolution.
The architecture is a hierarchical Variational Autoencoder (VAE) with SE(3)-equivariant encoder and decoder. The encoder works on SE(3)-invariant atom features and SE(3)-equivariant atomistic coordinates. The key aspect of working with 3D structures are appropriate equivariance. Three-dimensional molecules are equivariant under rotations and translations. To maintain this, SE(3)-equivariance is enforced in the encoder, decoder, and latent space of CoarsenConf.
In terms of experimental results, the CoarsenConf produced the lowest average and worst-case error across the entire test set of DRUGS molecules. It was discovered that RDKit, with inexpensive physics-based optimization, achieved better coverage than most deep learning-based methods.
The full-paper, entitled “CoarsenConf: Equivariant Coarsening with Aggregated Attention for Molecular Conformer Generation,” is authored by Danny Reidenbach and Aditi S. Krishnapriyan and is available for further references on arXiv.
