[Paper Review] Enabling Inverse Design in Chemical Compound Space: Mapping Quantum Properties to Structures for Small Organic Molecules

Alessio Fallani, Leonardo Medrano Sandonas+1 more|arXiv (Cornell University)|Sep 1, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper presents a variational autoencoder (VAE) framework that jointly encodes molecular structures and quantum mechanical (QM) properties into a shared latent space, enabling inverse design by generating 3D molecular structures from target QM properties. The method achieves accurate reconstruction of molecular geometry and composition using extensive QM properties as intrinsic coordinates in chemical space.

ABSTRACT

Computer-driven molecular design combines the principles of chemistry, physics, and artificial intelligence to identify novel chemical compounds and materials with desired properties for a specific application. In particular, quantum-mechanical (QM) methods combined with machine learning (ML) techniques have accelerated the estimation of accurate molecular properties, providing a direct mapping from 3D molecular structures to their properties. However, the development of reliable and efficient methodologies to enable \emph{inverse mapping} in chemical space is a long-standing challenge that has not been accomplished yet. Here, we address this challenge by demonstrating the possibility of parametrizing a given chemical space with a finite set of extensive and intensive QM properties. In doing so, we develop a proof-of-concept implementation that combines a Variational Auto-Encoder (VAE) trained on molecular structures with a property encoder designed to learn the latent representation from a set of QM properties. The result of this joint architecture is a common latent space representation for both structures and properties, which enables property-to-structure mapping for small drug-like molecules contained in the QM7-X dataset. We illustrate the capabilities of our approach by conditional generation of \emph{de novo} molecular structures with targeted properties, transition path interpolation for chemical reactions as well as insights into property-structure relationships. Our findings thus provide a proof-of-principle demonstration aiming to enable the inverse property-to-structure design in diverse chemical spaces.

Motivation & Objective

• Address the longstanding challenge of inverse molecular design—predicting molecular structures from desired quantum properties.
• Develop a unified latent space representation that jointly encodes molecular structures and their quantum mechanical properties.
• Enable conditional generation of novel molecular structures with targeted properties using the learned parameterization of chemical compound space.
• Investigate the interpretability of QM properties as intrinsic coordinates for navigating chemical space.
• Demonstrate the utility of the model in generating transition state geometries and interpolating reaction pathways.

Proposed method

• Train a Variational Auto-Encoder (VAE) on molecular structures represented as Coulomb matrices to learn a low-dimensional latent representation of 3D geometry.
• Implement a separate property encoder to map a set of quantum mechanical properties (e.g., energy, dipole, HOMO-LUMO gap) into the same latent space.
• Jointly train the VAE and property encoder to align the latent representations of structures and their corresponding QM properties.
• Use the trained property encoder in conjunction with the VAE decoder to perform conditional generation of molecular structures from desired property vectors.
• Leverage the differentiability of the latent space to identify key properties influencing molecular reconstruction and clustering.
• Apply the model to interpolate transition states in chemical reactions by generating intermediate geometries and validating them with nudged elastic band (NEB) calculations.

Experimental results

Research questions

• RQ1 Can quantum mechanical properties serve as effective intrinsic coordinates for parameterizing chemical compound space?
• RQ2 To what extent can a shared latent space representation enable accurate inverse mapping from properties to molecular structures?
• RQ3 Which quantum mechanical properties are most influential in determining molecular geometry and composition in the latent space?
• RQ4 Can the learned latent space support meaningful interpolation of reaction pathways and transition states?
• RQ5 How do extensive versus intensive properties compare in their ability to guide conformational navigation in chemical space?

Key findings

• The model successfully reconstructs molecular geometries and heavy atom compositions from QM property vectors with reasonable accuracy on the QM7-X dataset.
• The joint latent space enables conditional generation of novel molecular structures with targeted quantum properties, demonstrating feasibility for inverse design.
• The exchange and kinetic energy components were identified as the most influential QM properties in clustering molecules by chemical composition and bond types.
• Extensive properties outperformed intensive properties as local coordinates for conformational navigation, likely due to the geometric nature of the Coulomb matrix representation.
• Generated transition state geometries were successfully used as initial guesses for NEB calculations, with RMSD between interpolated and optimized geometries ranging from 0.14 Å to 0.35 Å.
• Geodesic interpolation using the latent space produced energy barriers that were consistently overestimated compared to NEB-calculated values, highlighting the need for refinement in path reconstruction.