Impact of Chemist-In-The-Loop Molecular Representations on Machine Learning Outcomes

Abstract

The development of molecular descriptors is a central challenge in cheminformatics. Most approaches use algorithms that extract atomic environments or end-to-end machine learning. However, a looming question is that how do these approaches compare with the critical eye of trained chemists. The CAS fingerprint engages expert chemists to curate chemical motifs, which they deem could influence bioactivity. In this paper, we benchmark the CAS fingerprint against commonly used fingerprints using a well-established benchmark set of 88 targets. We show that the CAS fingerprint outperforms most of the commonly used molecular fingerprints. Analysis of the CAS fingerprint reveals that experts tend to select features that are rarely reported in the literature, though not all rare features are selected. Our analysis also shows that the CAS fingerprint provides a different source of information compared to other commonly used fingerprints. These results suggest that anthropomorphic insights do have predictive power and highlight the importance of a chemist-in-the-loop approach in the era of machine learning.

Physics-Inspired Structural Representations for Molecules and Materials

“…Even though this review has no intention of providing an exhaustive historical account of the development of descriptors for atomic structures, it is worth providing a brief overview. A “data-driven” philosophy emerged early in the field of chemical and molecular science, where the combinatorial extent of the space of possible molecules, and the possibility of accessing this space with comparatively simple synthetic strategies, encouraged the development of quantitative structure–property relationship (QSPR) techniques, attempting to map descriptors of molecular structurebased on cheminformatics fingerprints, , chemical-intuition driven descriptors, molecular graphs, or indicators obtained from quantum chemical calculationsto the behavior of a selected compound, usually focusing on properties of direct applicative interest − such as solubility, toxicity, or pharmacological activity. , …”

Artificial Intelligence in Chemistry: Current Trends and Future Directions

“…The US is the leading country of origin: 15 of the 34 papers in Tables 1−3 are affiliated with US organizations. Other countries with significant numbers of important AI documents are Germany (6) and Switzerland (5). Among organizations, the Massachusetts Institute of Technology (US) and the University of Basel (Switzerland) were the two biggest contributors.…”

Physics-inspired structural representations for molecules and materials

“…Even though this review has no intention of providing a comprehensive discussion of the development of descriptors for atomic structures, it is worth providing a brief overview. A "data-driven" philosophy emerged early in the field of chemical and molecular science, where the combinatorial extent of the space of possible molecules, 31 and the possibility of accessing this space with comparatively simple synthetic strategies, encouraged the development of quantitative structure/property relationships (QSPR) techniques, attempting to map 32 descriptors of molecular structure -based on cheminformatics fingerprints, 33,34 chemical-intuition driven descriptors 35 , molecular graphs, 36 or indicators obtained from quantum chemical calculations 37 -to the behavior of a selected compound, usually focusing on properties of direct applicative interest [38][39][40] such as solubility, toxicity, 41 or pharmacological activity. 42,43 Cartesian coordinates This approach should be contrasted with that of "bottom-up" predictions, that aim to use models of the interactions between the atomic constituents of a material to simulate the behavior of the system on an atomic time and length scale.…”

“…135 to learn interatomic forces. The fact that equivariant features of the form (35) follow O(3) transformation rules means that they can be combined using established relationships in the quantum theory of angular momentum. In particular, the coupled-basis representation used in the definition of Eqs.…”

“…This is well exemplified in Figs. 35, where the electron density prediction of C(8) hydrocarbons 22 is tested upon having trained the model on much smaller compounds, with only 4 carbon atoms. This approach has since been applied to more complex systems, such as oligopeptides 276 , and to the prediction of other scalar fields such as the on-top density 277 .…”