Communication: Understanding molecular representations in machine learning: The role of uniqueness and target similarity.

By Bing Huang, O. A. von Lilienfeld. 2016

280 citations recorded by the source corpus.

The predictive accuracy of Machine Learning (ML) models of molecular properties depends on the choice of the molecular representation. Inspired by the postulates of quantum mechanics, we introduce a hierarchy of representations which meet uniqueness and target similarity criteria. To systematically control target similarity, we simply rely on interatomic many body expansions, as implemented in universal force-fields, including Bonding, Angular (BA), and higher order terms. Addition of higher order contributions systematically increases similarity to the true potential energy and predictive accuracy of the resulting ML models. We report numerical evidence for the performance of BAML models trained on molecular properties pre-calculated at electron-correlated and density functional theory level of theory for thousands of small organic molecules. Properties studied include enthalpies and free energies of atomization, heat capacity, zero-point vibrational energies, dipole-moment, polarizability, HOMO/LUMO energies and gap, ionization potential, electron affinity, and electronic excitations. After training, BAML predicts energies or electronic properties of out-of-sample molecules with unprecedented accuracy and speed.

Subfields: Atomic and Molecular Physics, and Optics, Materials Chemistry, Physical and Theoretical Chemistry, Representation & Self-Supervised Learning, Computational Chemistry & Drug Discovery

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Atomic and Molecular Physics, and Optics

1. Evgeny V. Podryabinkin
2. Bing Huang
3. Konstantin Gubaev
4. Sam Schoenholz
5. Oriol Vinyals
6. Justin Gilmer
7. George E. Dahl
8. Ivan S. Novikov

In Materials Chemistry

1. Sam Schoenholz
2. Evgeny V. Podryabinkin
3. Konstantin Gubaev
4. Ivan S. Novikov
5. Alexander V. Shapeev
6. Steven M. Kearnes
7. Oriol Vinyals
8. Patrick F. Riley

In Physical and Theoretical Chemistry

1. Sam Schoenholz
2. Evgeny V. Podryabinkin
3. Konstantin Gubaev
4. Oriol Vinyals
5. Justin Gilmer
6. George E. Dahl
7. O. V. Lilienfeld
8. Ivan S. Novikov

In Representation & Self-Supervised Learning

1. Sam Schoenholz
2. Oriol Vinyals
3. Justin Gilmer
4. George E. Dahl
5. O. V. Lilienfeld
6. Felix A. Faber
7. Steven M. Kearnes
8. Yoshua Bengio

In Computational Chemistry & Drug Discovery

1. Sam Schoenholz
2. Patrick F. Riley
3. Oriol Vinyals
4. Steven M. Kearnes
5. Justin Gilmer
6. George E. Dahl
7. Ivan S. Novikov
8. O. V. Lilienfeld

Citations in the corpus

• Neural Message Passing for Quantum Chemistry (2017, methodology): “…representations include the Coulomb Matrix (CM, Rupp et al. (2012)), Bag of Bonds (BoB, Hansen et al. (2015)), Bonds Angles, Machine Learning (BAML, Huang & von Lilienfeld (2016)), Extended Connectivity Fingerprints (ECPF4, Rogers & Hahn (2010)), and “Projected Histograms” (HDAD, Faber et al.…”
• The MLIP package: moment tensor potentials with MPI and active learning (2020, background): “These are symbolic-regression potentials [29], non-conservative force fields [8,38], force-fields for a fixed-size molecular system (typically applied to one or two organic molecules) [9, 42], on-lattice potentials [66], and cheminformatics-type models [10, 14,19,21,28,30,62,64,71], see also a recen”
• Prediction Errors of Molecular Machine Learning Models Lower than Hybrid DFT Error. (2017, background): “11,61 Errors for most models shown decay with roughly the same slopes, indicating similar exponents in the power-law of error decay.”