A Mixed Quantum Chemistry/Machine Learning Approach for the Fast and Accurate Prediction of Biochemical Redox Potentials and Its Large-Scale Application to 315 000 Redox Reactions

Abstract

A quantitative understanding of the thermodynamics of biochemical reactions is essential for accurately modeling metabolism. The group contribution method (GCM) is one of the most widely used approaches to estimate standard Gibbs energies and redox potentials of reactions for which no experimental measurements exist. Previous work has shown that quantum chemical predictions of biochemical thermodynamics are a promising approach to overcome the limitations of GCM. However, the quantum chemistry approach is significantly more expensive. Here, we use a combination of quantum chemistry and machine learning to obtain a fast and accurate method for predicting the thermodynamics of biochemical redox reactions. We focus on predicting the redox potentials of carbonyl functional group reductions to alcohols and amines, two of the most ubiquitous carbon redox transformations in biology. Our method relies on semiempirical quantum chemistry calculations calibrated with Gaussian process (GP) regression against available experimental data and results in higher predictive power than the GCM at low computational cost. Direct calibration of GCM and fingerprint-based predictions (without quantum chemistry) with GP regression also results in significant improvements in prediction accuracy, demonstrating the versatility of the approach. We design and implement a network expansion algorithm that iteratively reduces and oxidizes a set of natural seed metabolites and demonstrate the high-throughput applicability of our method by predicting the standard potentials of more than 315 000 redox reactions involving approximately 70 000 compounds. Additionally, we developed a novel fingerprint-based framework for detecting molecular environment motifs that are enriched or depleted across different regions of the redox potential landscape. We provide open access to all source code and data generated.

Comparison of computational chemistry methods for the discovery of quinone-based electroactive compounds for energy storage

“…In addition to the above findings, the results presented here are expected to be useful for generating accurate and large quantity chemical data on compounds, which can later be utilized by data-driven machine learning models for expanding the boundaries of search space during candidate compound explorations. Firstly, as has been argued in recent studies 27,29 , using the computationally costly quantum chemical calculations for millions of molecules, which is required for building powerful ML models, is still a major bottleneck. One of the key findings of this work is that the DFTB method is nearly as accurate as DFT when it comes to the prediction of quinone redox potentials.…”

“…Therefore, there is a need to determine the trade-offs between the prediction accuracy and the computational cost. While there has been a significant increase in the number of HTCS efforts for RFBs [25][26][27][28][29][30][31][32] , to the best of our knowledge, an analysis of the factors that affect prediction accuracy, such as the level of theory for optimization of molecular geometry, inclusion/exclusion of solvation effects, and the level of theory for the calculation of chemical descriptors, are not available in the current literature.…”

Machine learning dihydrogen activation in the chemical space surrounding Vaska's complex

“…The catalysis of a particular reaction is typically proven for only a few tens (or less) of metal complexes. The possibility of learning from synthetic data generated in silico [99][100][101] is appealing, but in practice accurate QC calculations are expensive. In this context, optimizing the ratio between accuracy and the size of the training data is imperative.…”

High throughput virtual screening of 230 billion molecular solar heat battery candidates

“…Select structures are investigated further at the M06-2X (Zhao & Truhlar, 2008)/6-31G(d) (Hehre, Ditchfield & Pople, 1972) level of theory (using Gaussian09). This level of theory was chosen as a good compromise between computational efficiency and accuracy as judged by comparison to DLPNO-CCSD(T) and CCSD(T)-F12a calculations (Koerstz, Elm & Mikkelsen, 2017).…”