(364j) Generating 3D Conformer Ensembles of Molecular Graphs with End-to-End Differentiable Networks
AIChE Annual Meeting
2021
2021 Annual Meeting
Topical Conference: Applications of Data Science to Molecules and Materials
Applications of Data Science in Molecular Sciences II
Tuesday, November 9, 2021 - 5:18pm to 5:30pm
Prediction of small molecule 3D conformer ensembles plays an important role in many areas of cheminformatics, from the calculation of accurate reaction rates to the drug discovery pipeline. Traditional distance geometry-based generators have several issues including the need for computationally-expensive energy minimization, over-parameterization of the final structure, poor coverage of the true conformer distributions (i.e. lack of diversity in the generated ensembles), and iterative and non-deterministic treatment of tetrahedral chiral centers.
We propose a novel machine learning approach to generate distributions of low-energy 3D conformers directly from molecular graphs. Our approach comprises three steps. First, we predict the local 3D structure of all non-terminal atoms, which we deem local environments, by combining self-attention layers and message passing neural networks. These predictions correspond the the bond distances and bond angles of the final conformer. Next, we predict the dihedral angles that join all pairs of neighboring local environments. Importantly, because we know all bond angles of local environments, we only predict a single torsion angle value, which allows us to recover all the dihedral angles between local environments. That is, we develop a canonical representation of dihedral angles, which allows us to predict exactly the number of degrees of freedom of the conformer. Finally, we assemble all predicted pairs of local environments together to construct the conformer.
Since we build the conformer from predictions of local environments, we can deterministically handle tetrahedral chiral centers without the need for iterative optimization. To promote diverse conformer ensembles, we train our network end-to-end using a loss function inspired by optimal transport. Empirically, we outperform popular baselines, showcasing the benefit of our model. Finally, we discuss possible extensions and applications of our differentiable 3D structure generator to downstream tasks such as modeling drug-target interactions.
We propose a novel machine learning approach to generate distributions of low-energy 3D conformers directly from molecular graphs. Our approach comprises three steps. First, we predict the local 3D structure of all non-terminal atoms, which we deem local environments, by combining self-attention layers and message passing neural networks. These predictions correspond the the bond distances and bond angles of the final conformer. Next, we predict the dihedral angles that join all pairs of neighboring local environments. Importantly, because we know all bond angles of local environments, we only predict a single torsion angle value, which allows us to recover all the dihedral angles between local environments. That is, we develop a canonical representation of dihedral angles, which allows us to predict exactly the number of degrees of freedom of the conformer. Finally, we assemble all predicted pairs of local environments together to construct the conformer.
Since we build the conformer from predictions of local environments, we can deterministically handle tetrahedral chiral centers without the need for iterative optimization. To promote diverse conformer ensembles, we train our network end-to-end using a loss function inspired by optimal transport. Empirically, we outperform popular baselines, showcasing the benefit of our model. Finally, we discuss possible extensions and applications of our differentiable 3D structure generator to downstream tasks such as modeling drug-target interactions.