Fitting quantum machine learning potentials to experimental free energy data: Predicting tautomer ratios in solution

• DOI: 10.1101/2020.10.24.353318
• Publication Date: 2020-10-25T16:15:27Z
• Authors: Marcus Wieder; Josh Fass; John D. Chodera
• Keywords: Thermochemistry; Potential energy surface; Quantum Chemistry

Abstract The computation of tautomer ratios druglike molecules is enormously important in computer-aided drug discovery, as over a quarter all approved drugs can populate multiple tautomeric species solution. Unfortunately, accurate calculations aqueous ratios—the degree to which these must be penalized order correctly account for tautomers modeling binding discovery—is surprisingly diffcult. While quantum chemical approaches computing using continuum solvent models and rigid-rotor harmonic-oscillator thermochemistry are currently state the art, methods still inaccurate despite their enormous computational expense. Here, we show that major source this inaccuracy lies breakdown standard approach accounting rigid rotor harmonic oscillator (RRHO) approximations, frustrated by complex conformational landscape introduced migration double bonds, creation stereocenters, introduction conformations separated low energetic barriers induced single proton. Using machine learning (QML) allow us compute potential energies with accuracy at fraction cost, how rigorous relative alchemical free energy used vacuum from limitations RRHO approximations. Furthermore, since parameters QML tunable, train correct underlying learned surface energies, enabling learn generalize across broader range predictions.