Kernel Methods and Similarity Learning Applied in Computational Chemistry

Over the last two decades, data-powered machine learning (ML) tools have profoundly transformed numerous scientific fields. In computational chemistry, machine learning applications have permitted faster predictions of chemical properties and provided powerful analytical tools, facilitating the exploration of the chemical space. The original work presented in this thesis leverages the paradigm-shifting influence of ML and focuses on bridging the divide between unsupervised and supervised learning with the overarching objective of improving the predictive power of similarity-based machine learning algorithms such as kernel regression.Despite their widespread use in chemistry, current implementations of kernel regression suffer from biased definitions of similarity between chemical environments. This problem originates from the rigidity of current numerical approaches for encoding molecular information, based on expert-crafted representations. Moreover, it is amplified by the incorrect (yet generalized) assumption that increasing the amount of information encoded in molecular representations unequivocally improves the evaluation of molecular similarity. As a result, the performance of kernel models can be sub-optimal reducing their broad applicability. To overcome such limitations, we introduce a series of statistical tools and methodologies based on supervised dimensionality reduction and metric learning capable of filtering and adapting the features of common molecular representations. This allows tailoring the notion of "molecular similarity" in order to optimize the prediction of specific chemical targets.Using examples such as the exploration of the free-energy landscape of oligopeptides or the prediction of subtle properties associated with the outcome of chemical reactions (for example, enantiomeric excess), we demonstrate how the methods proposed in this thesis unlock the optimal performance of kernel regression and, more generally, of any similarity-based algorithm. Overall, the work within is part of a larger, more comprehensive effort aimed at extending the capabilities of computational modeling to increasingly complex chemical situations by exploiting the latest advances in statistical learning.

Applications of Artificial Intelligence to Computational Chemistry

Nicholas John Browning

The calculation of the electronic structure of chemical systems, necessitates computationally expensive approximations to the time-independent electronic Schrödinger equation in order to yield static properties in good agreement with experimental results. These methods can also be coupled with molecular dynamics, to provide a first principles description of thermodynamic properties, dynamics and chemical reactions. Evidently, the cost of the underlying electronic structure method limits the time frame over which a system can be studied, and hence, certain chemical processes may be out-of-reach using a particular method. Furthermore, when one is interested in designing new molecules with interesting properties, a systematic enumeration of an inordinately large chemical space is typically required. The combination of both expensive electronic structure calculations and large chemical spaces results in an insurmountable barrier in computational cost. The application of artificial intelligence (AI) in computational chemistry has, over the past 20 years, seen an explosion in interest and scope with respect to these two issues. Intelligent algorithms capable of efficiently sampling chemical spaces, coupled with machine learning (ML) techniques to cheapen the calculation of electronic structure evaluations, enable both rapid throughput to search for new molecules with particular properties and in the case of ML, an increase in the timescales that can be simulated via molecular dynamics. In this thesis, computer programs have been developed that enable the application of AI algorithms to chemical and biological problems. In particular, a versatile evolutionary algorithm toolbox called EVOLVE has been developed. As a first test-case study, genetic algorithms were used to efficiently sample the vast chemical sequence space of an isolated ¿-helical peptide, from which insights are gained to rationalise the stability of particular genetically optimised peptides in a variety of implicit solvent environments. Genetic algorithms were then applied to the compositional optimisation of training sets used in machine learning models of molecular properties. The resulting optimal training sets are shown to significantly reduce out-of-sample errors on all thermodynamic and electronic properties considered. Furthermore, they reveal that there are systematic trends in the distribution of these optimally-representative molecules. Inspired by the success of machine learning, an ML-enhanced multiple time step approach for performing accurate ab initio molecular dynamics was developed. Two schemes representing different force partitioning were investigated. In the first scheme, the ML method provides an estimation of the slow (high level) force components acting on a system, while in the second, the ML forces are added to the fast (low level) components and a high level ab initio method is used to correct the error induced by the ML model. In both schemes significant overall speedups are obtained with respect to standard Velocity-Verlet integration, all-the-while maintaining the accuracy of the high level ab initio method.

EPFL2019

Pushing the Frontiers of First-Principles Based Computer Simulations of Chemical and Biological Systems

Negar Ashari Astani, Elizabeth Claire Brunk, Pablo Campomanes Ramos, Basile Curchod, Polydefkis Diamantis, Andrey Laktionov, Marilisa Neri, Giulia Palermo, Thomas James Penfold, Ursula Röthlisberger, Ivano Tavernelli, Stefano Vanni

The Laboratory of Computational Chemistry and Biochemistry is active in the development and application of first-principles based simulations of complex chemical and biochemical phenomena. Here, we review some of our recent efforts in extending these methods to larger systems, longer time scales and increased accuracies. Their versatility is illustrated with a diverse range of applications, ranging from the determination of the gas phase structure of the cyclic decapeptide gramicidin S, to the study of G protein coupled receptors, the interaction of transition metal based anti-cancer agents with protein targets, the mechanism of action of DNA repair enzymes, the role of metal ions in neurodegenerative diseases and the computational design of dye-sensitized solar cells. Many of these projects are done in collaboration with experimental groups from the Institute of Chemical Sciences and Engineering (ISIC) at the EPFL.

2011

Computational Quantum Chemical Studies on Radicals

Peter Rudolf Tentscher

Radicals play an important role in many areas of chemistry, such as atmospheric, aquatic, polymer, and biological, to name a few. Radicals are often highly reactive species which are short-lived and therefore harder to study by experimental techniques. In principle, computational quantum chemistry enables the study of arbitrary chemical species, including elusive radicals. However, the approximate quantum chemical methods which are used for production simulations need to be validated for the type of problem in question. First, it is necessary to validate that highly accurate reference methods commonly applied to closed-shell systems are equally applicable to radicals. I show that the “gold standard” for closed-shell systems (coupled cluster with singles, doubles, and perturbative triples excitations (CCSD(T))) can produce molecular geometries and vibrational frequencies suitable for (sub-)kJ/mol thermochemistry for radical species as well. However, additional diagnosis is necessary, especially for vibrational frequency calculations. Coupled cluster methods are then used to compute benchmark binding energies between small radicals and water or hydrogen fluoride as a proxy to the aqueous condensed phase. I find that many common density functional theory (DFT) methods cannot reproduce these binding energies to within chemical accuracy (1 kcal/mol). However, some functionals, for example MPW1K and BHandHLYP, perform better than others. If more accurate ab initio methods are computationally too demanding for the system of interest, these DFT methods should be used for (micro-)solvated radicals. Complexes of radicals with polar solvent are found to be more strongly bound than expected. This can not be explained by dispersive and electrostatic interactions alone. Using natural bond orbitals, I show that additional donor-acceptor interactions may be responsible for the strong intermolecular binding observed. One of the most fundamental processes involving aqueous organic radicals is the one-electron oxidation of closed-shell organics. I study the vertical ionization of neutral organics. That is, the ionization is faster than the reorganization of the solvent (and solute), and the conformation is assumed to be unchanged during the ionization process. I find that the difference between the gas-phase and aqueous-phase vertical ionization energy, ∆VIE = VIEaq − VIEgas, ranges from ≈ 0 to -1.0 eV for small aromatic compounds. The differences in ∆VIE between different compounds seem to be caused by the solvent structures of the ≈ 70 water molecules closest to the solute. In natural surface waters, oxidized organics arise through the reaction of photochemically created radicals with organic molecules already present in the water column. One example for such a reaction is the oxidation of sulfonamide antibiotics. I show that upon oxidation of sulfadiazine and related compounds, an aniline radical cation is formed. The reactivity of this oxidized species is strikingly different from its reduced counterpart: the barrier for a nucleophilic attack on the aniline ring is drastically lowered, and a fast rearrangement reaction can take place. I demonstrate that current quantum chemical protocols are suitable for mechanistic studies of solvated radicals, and I successfully apply them to such problems. To achieve chemical accuracy for quantum computational protocols, further developments are necessary. To validate novel approaches, the benchmark data presented in this thesis may be used.