Alchemical free energy methods with molecular mechanics (MM) force fields are now widely used in the prioritization of small molecules for synthesis in structure-enabled drug discovery projects because of their ability to deliver 1–2 kcal mol ⁻¹ accuracy in well-behaved protein-ligand systems. Surpassing this accuracy limit would significantly reduce the number of compounds that must be synthesized to achieve desired potencies and selectivities in drug design campaigns. However, MM force fields pose a challenge to achieving higher accuracy due to their inability to capture the intricate atomic interactions of the physical systems they model. A major limitation is the accuracy with which ligand intramolecular energetics—especially torsions—can be modeled, as poor modeling of torsional profiles and coupling with other valence degrees of freedom can have a significant impact on binding free energies. Here, we demonstrate how a new generation of hybrid machine learning / molecular mechanics (ML/MM) potentials can deliver significant accuracy improvements in modeling protein-ligand binding affinities. Using a nonequilibrium perturbation approach, we can correct a standard, GPU-accelerated MM alchemical free energy calculation in a simple post-processing step to efficiently recover ML/MM free energies and deliver a significant accuracy improvement with small additional computational effort. To demonstrate the utility of ML/MM free energy calculations, we apply this approach to a benchmark system for predicting kinase:inhibitor binding affinities—a congeneric ligand series for non-receptor tyrosine kinase TYK2 (Tyk2)—wherein state-of-the-art MM free energy calculations (with OPLS2.1) achieve inaccuracies of 0.93±0.12 kcal mol ⁻¹ in predicting absolute binding free energies. Applying an ML/MM hybrid potential based on the ANI2x ML model and AMBER14SB/TIP3P with the OpenFF 1.0.0 (“Parsley”) small molecule force field as an MM model, we show that it is possible to significantly reduce the error in absolute binding free energies from 0.97 [95% CI: 0.68, 1.21] kcal mol ⁻¹ (MM) to 0.47 [95% CI: 0.31, 0.63] kcal mol ⁻¹ (ML/MM).

A force field as accurate as quantum mechanics (QMs) and as fast as molecular mechanics (MMs), with which one can simulate a biomolecular system efficiently enough and meaningfully enough to get quantitative insights, is among the most ardent dreams of biophysicists—a dream, nevertheless, not to be fulfilled any time soon. Machine learning force fields (MLFFs) represent a meaningful endeavor in this direction, where differentiable neural functions are parametrized to fit ab initio energies and forces through automatic differentiation. We argue that, as of now, the utility of the MLFF models is no longer bottlenecked by accuracy but primarily by their speed, as well as stability and generalizability—many recent variants, on limited chemical spaces, have long surpassed the chemical accuracy of 1 kcal/mol—the empirical threshold beyond which realistic chemical predictions are possible—though still magnitudes slower than MM. Hoping to kindle exploration and design of faster, albeit perhaps slightly less accurate MLFFs, in this review, we focus our attention on the technical design space (the speed-accuracy trade-off) between MM and ML force fields. After a brief review of the building blocks (from a machine learning-centric point of view) of force fields of either kind, we discuss the desired properties and challenges now faced by the force field development community, survey the efforts to make MM force fields more accurate and ML force fields faster, and envision what the next generation of MLFF might look like.

In calculations of relative free energy differences, the number of atoms of the initial and final states is rarely the same. This necessitates the introduction of dummy atoms. These placeholders interact with the physical system only by bonded energy terms. We investigate the conditions necessary so that the presence of dummy atoms does not influence the result of a relative free energy calculation. On the one hand, one has to ensure that dummy atoms only give a multiplicative contribution to the partition function so that their contribution cancels from double-free energy differences. On the other hand, the bonded terms used to attach a dummy atom (or group of dummy atoms) to the physical system have to maintain it in a well-defined position and orientation relative to the physical system. A detailed theoretical analysis of both aspects is provided, illustrated by 24 calculations of relative solvation free energy differences, for which all four legs of the underlying thermodynamic cycle were computed. Cycle closure (or lack thereof) was used as a sensitive indicator to probing the effects of dummy atom treatment on the resulting free energy differences. We find that a naive (but often practiced) treatment of dummy atoms results in errors of up to k_BT when calculating the relative solvation free energy difference between two small solutes, such as methane and ammonia. While our analysis focuses on the so-called single topology approach to set up alchemical transformations, similar considerations apply to dual topology, at least many widely used variants thereof.

Enumerating protonation states and calculating microstate pK_a values of small molecules is an important yet challenging task for lead optimization and molecular modeling. Commercial and non-commercial solutions have notable limitations such as restrictive and expensive licenses, high CPU/GPU hour requirements, or the need for expert knowledge to set up and use. We present a graph neural network model that is trained on 714,906 calculated microstate pK_a predictions from molecules obtained from the ChEMBL database. The model is fine-tuned on a set of 5,994 experimental pK_a values significantly improving its performance on two challenging test sets. Combining the graph neural network model with Dimorphite-DL, an open-source program for enumerating ionization states, we have developed the open-source Python package pkasolver, which is able to generate and enumerate protonation states and calculate pK_a values with high accuracy.

The large neutral amino acid transporter 1 (LAT1) is a promising anticancer target that is required for the cellular uptake of essential amino acids that serve as building blocks for cancer growth and proliferation. Here, we report a structure-based approach to identify chemically diverse and potent inhibitors of LAT1. First, a homology model of LAT1 that is based on the atomic structures of the prokaryotic homologs was constructed. Molecular docking of nitrogen mustards (NMs) with a wide range of affinity allowed for deriving a common binding mode that could explain the structure−activity relationship pattern in NMs. Subsequently, validated binding hypotheses were subjected to molecular dynamics simulation, which allowed for extracting a set of dynamic pharmacophores. Finally, a library of ~1.1 million molecules was virtually screened against these pharmacophores, followed by docking. Biological testing of the 30 top-ranked hits revealed 13 actives, with the best compound showing an IC₅₀ value in the sub-μM range.