Text‐based representations with interpretable machine learning reveal structure–property relationships of polybenzenoid hydrocarbons

Abstract

New tools are developed and applied to enable the use of interpretable machine learning for investigation of structure–property relationships in polybenzenoid hydrocarbons (PBHs). A textual molecular representation, which is based on the annulation sequence of PBHs, is shown to be of utility either in its textual form or as a basis for a curated feature vector. Both forms display interpretability exceeding those achievable by standard SMILES representation; and the former also has increased predictive accuracy. A recently developed model, CUSTODI, was applied for the first time as an interpretable model, identifying important structural features that impact various electronic molecular properties. The resulting insights not only validate several well‐known “rules of thumb” of organic chemistry but also reveal new behaviors and influential structural motifs, thus providing guiding principles for rational design and fine‐tuning of PBHs.

Interpretable Deep-Learning Unveils Structure–Property Relationships in Polybenzenoid Hydrocarbons

“…We recently utilized this to formulate a text-based representation that can be used with interpretable machine learning models. 38 In this work, we combined this subunit-based approach with the graph-of-rings (GOR) representation (see the Molecular Representation section in Computational Methods for further details) to extract further chemical insight using interpretable deep-learning models. Such graph representations were successfully employed previously in the investigation of PBHs.…”

“…One of the main conclusions from this body of work was that many of the characteristics of cata -condensed PBHs are encoded in the sequence of tricyclic annulations, namely, linear or angular annulations. We recently utilized this to formulate a text-based representation that can be used with interpretable machine learning models . In this work, we combined this subunit-based approach with the graph-of-rings (GOR) representation (see the Molecular Representation section in Computational Methods for further details) to extract further chemical insight using interpretable deep-learning models.…”

A Crowning Achievement: The First Solution‐Phase Synthesis of Circumcoronenes

“…Certain relationships have already been uncovered. For example, the molecular electronic properties of cata ‐condensed PBHs are mostly determined by their longest linearly annulated subsection, [7, 8] while the properties of peri ‐condensed PBHs are determined by their size and edge‐attributes [9] . Indeed, a correlation has been found between the presence of zig‐zag edges, low band‐gap, and multiple spin‐states [10] .…”

COMPAS-3: a Data Set of peri-Condensed Polybenzenoid Hydrocarbons

“…The first installment, COMPAS-1, 60 focuses on groundstate cata-condensed Polybenzenoid Hydrocarbons; the second installment, COMPAS-2, 61 focuses on ground-state cata-condensed heterocyclic PASs. COMPAS-1 and COMPAS-2 have already been used to provide the first examples of interpretable machine and deeplearning models for PASs 62,63 and to demonstrate the first generative design of PASs with targeted properties. 64 Both data, as well as all future installments, are freely available for use, according to the FAIR 65 principles of data sharing.…”