Quantum chemical roots of machine-learning molecular similarity descriptors

Stefan Gugler, Markus Reiher

TL;DR

This paper bridges the gap between popular machine learning (ML) molecular descriptors and the fundamental laws of quantum chemistry. By deriving new descriptors—Coulomb Lists (CL) and Smooth Overlap of Electron Densities (SOED)—directly from electronic structure theory, the authors provide a more physically grounded way to represent molecules, offering better interpretability and sensitivity to the subtle changes that occur during chemical reactions.

What Problem Does This Paper Address?

In the world of "AI for Chemistry," we represent molecules as vectors called descriptors. Two of the most popular are the Coulomb Matrix (CM), which looks at nuclear repulsion, and SOAP, which represents a molecule as a collection of "fuzzy" Gaussian atoms.

However, these descriptors are often chosen for mathematical convenience rather than physical rigor. This leads to the "Similarity Paradox" in reaction chemistry: during a reaction, the Transition State (TS) is geometrically very similar to the reactants or products. Standard descriptors see this geometric similarity and assume the properties are similar. In reality, the electronic structure of a TS is vastly different (often involving stretched bonds and complex electron correlation), making its energy much harder to predict. The authors aim to ground these descriptors in quantum mechanics to see if we can better capture these "invisible" electronic effects.

Methods

The authors take two established ML frameworks and re-derive them using the language of quantum chemistry:

  1. From Coulomb Matrix to Coulomb Lists (CL): The standard CM requires a diagonalization step to ensure that the order of atoms doesn't change the result. The authors show that by splitting the electronic energy expression into atomic contributions, they can create "Coulomb Lists." These lists avoid the computationally expensive diagonalization and, crucially, keep the information tied to specific atoms, making the model much easier to interpret.
  2. From SOAP to SOED: SOAP uses arbitrary Gaussian functions to represent atoms. The authors replace these with the actual electron density of the molecule. Since electron density is a fundamental quantum observable (via the Hohenberg–Kohn theorem), this "Smooth Overlap of Electron Densities" (SOED) should theoretically be a more "complete" descriptor.
  3. Lobe Function Approximation: To make SOED computationally feasible, the authors used "lobe functions"—approximating complex orbital shapes (like p or d orbitals) using simple s-type Gaussians. This allows them to use existing, highly optimized SOAP software architectures to calculate their new electronic descriptor.

Datasets

Rather than using a massive "black box" dataset, the authors focused on a detailed analysis of three specific elementary reaction steps:

  • Reaction 1: A generic imidazole ring-opening (a structure found in the common QM9 dataset).
  • Reaction 2: The same reaction but with "push-pull" substituents (electron-donating and withdrawing groups) to test electronic sensitivity.
  • Reaction 3: The same reaction with bulky ethyl groups to test steric (spatial) sensitivity.

The reference data consisted of Intrinsic Reaction Coordinate (IRC) paths calculated using Density Functional Theory (DFT) at the PBE/SVP level.

Key Insights and Ideas

  • The Interpretability Win: Unlike the "mysterious" eigenvalues of a Coulomb Matrix, Coulomb Lists allow researchers to track exactly which atom is changing most during a reaction. For example, they could clearly see the "jump" in the descriptor value for a shifting proton.
  • The Alpha Sensitivity: In SOAP, the width of the Gaussian atoms ($\alpha$) is a tunable hyperparameter. In SOED, this width is physically determined by the basis set. The authors found that narrower Gaussians (higher $\alpha$) make the descriptor significantly more sensitive to small structural changes, which is vital for distinguishing transition states.
  • Redundancy is Real: The authors found that nuclear repulsion features and external potential features are highly correlated. This suggests that for many applications, we can use cheaper, nuclear-only descriptors without losing significant accuracy.

Key Results

  • Increased Sensitivity: SOED was shown to be much more sensitive to reaction progress than standard SOAP. It "dropped" more significantly as the molecule moved from reactant to transition state, providing a clearer signal for ML models to learn from.
  • Computational Efficiency: Coulomb Lists provide nearly identical information to CM eigenvalues but are cheaper to compute ($O(M \log M)$ for sorting vs. $O(M^3)$ for diagonalization) and offer better atom-specific insights.
  • The Paradox Persists: Despite being rooted in quantum theory, both descriptors still primarily track geometry. Even with electron density included, the descriptors still see the Transition State as being "more similar" to the reactants than the reactants are to the products, leaving the electronic "similarity paradox" partially unresolved.

Limitations and Future Work

The primary limitation is computational scaling. Because SOED uses basis functions rather than just atoms, the number of "centers" to calculate increases dramatically. For a large molecule, SOED scales at $O(m^4)$ (where $m$ is the number of basis functions), compared to $O(M^2)$ for SOAP (where $M$ is the number of atoms).

The authors suggest that future work should focus on:

  1. Basis Set Reduction: Finding ways to approximate the electron density with fewer Gaussians to speed up SOED.
  2. Large-Scale Testing: Applying these descriptors to massive datasets to see if the increased sensitivity actually leads to better energy predictions in real-world ML tasks.
  3. Encoding Correlation: Developing descriptors that explicitly capture electron correlation effects, which are currently the "missing link" in resolving the similarity paradox.

Summary updated

Abstract

In this work, we explore the quantum chemical foundations of descriptors for molecular similarity. Such descriptors are key for traversing chemical compound space with machine learning. Our focus is on the Coulomb matrix and on the smooth overlap of atomic positions (SOAP). We adopt a basic framework that allows us to connect both descriptors to electronic structure theory. This framework enables us then to define two new descriptors that are more closely related to electronic structure theory, which we call Coulomb lists and smooth overlap of electron densities (SOED). By investigating their usefulness as molecular similarity descriptors, we gain new insights in how and why Coulomb matrix and SOAP work. Moreover, Coulomb lists avoid the somewhat mysterious diagonalization step of the Coulomb matrix and might provide a direct means to extract subsystem information that can be compared across Born-Oppenheimer surfaces of varying dimension. For the electron density we derive the necessary formalism to create the SOED measure in close analogy to SOAP. Since this formalism is more involved than that of SOAP, we review the essential theory, but also introduce a set of approximations that eventually allow us to work with SOED in terms of the same implementation available for the evaluation of SOAP. We focus our analysis on elementary reaction steps, where transition state structures are more similar to either reactant or product structures than the latter two are with respect to one another. The prediction of electronic energies of transition state structures can, however, be more difficult than that of stable intermediates due to multi-configurational effects. The question arises to what extent molecular similarity descriptors rooted in electronic structure theory can resolve these intricate effects.
BibTeX
@article{gugler2022quantum,
  author = {Stefan Gugler and Markus Reiher},
  title = {{Quantum chemical roots of machine-learning molecular similarity descriptors}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2022},
  doi = {10.1021/acs.jctc.2c00718},
}
Figure