Structure prediction of protein-ligand complexes from sequence information with Umol.

Protein-ligand docking is an established tool in drug discovery and development to narrow down potential therapeutics for experimental testing. However, a high-quality protein structure is required and often the protein is treated as fully or partially rigid. Here we develop an AI system that can predict the fully flexible all-atom structure of protein-ligand complexes directly from sequence information. We find that classical docking methods are still superior, but depend upon having crystal structures of the target protein. In addition to predicting flexible all-atom structures, predicted confidence metrics (plDDT) can be used to select accurate predictions as well as to distinguish between strong and weak binders. The advances presented here suggest that the goal of AI-based drug discovery is one step closer, but there is still a way to go to grasp the complexity of protein-ligand interactions fully. Umol is available at: https://github.com/patrickbryant1/Umol .

Topological Analysis of Molecular Dynamics Simulations using the Euler Characteristic.

Molecular dynamics (MD) simulations are used in diverse scientific and engineering fields such as drug discovery, materials design, separations, biological systems, and reaction engineering. These simulations generate highly complex data sets that capture the 3D spatial positions, dynamics, and interactions of thousands of molecules. Analyzing MD data sets is key for understanding and predicting emergent phenomena and in identifying key drivers and tuning design knobs of such phenomena. In this work, we show that the Euler characteristic (EC) provides an effective topological descriptor that facilitates MD analysis. The EC is a versatile, low-dimensional, and easy-to-interpret descriptor that can be used to reduce, analyze, and quantify complex data objects that are represented as graphs/networks, manifolds/functions, and point clouds. Specifically, we show that the EC is an informative descriptor that can be used for machine learning and data analysis tasks such as classification, visualization, and regression. We demonstrate the benefits of the proposed approach through case studies that aim to understand and predict the hydrophobicity of self-assembled monolayers and the reactivity of complex solvent environments.

Structural features of interfacial water predict the hydrophobicity of chemically heterogeneous surfaces.

The hydrophobicity of an interface determines the magnitude of hydrophobic interactions that drive numerous biological and industrial processes. Chemically heterogeneous interfaces are abundant in these contexts; examples include the surfaces of proteins, functionalized nanomaterials, and polymeric materials. While the hydrophobicity of nonpolar solutes can be predicted and related to the structure of interfacial water molecules, predicting the hydrophobicity of chemically heterogeneous interfaces remains a challenge because of the complex, non-additive contributions to hydrophobicity that depend on the chemical identity and nanoscale spatial arrangements of polar and nonpolar groups. In this work, we utilize atomistic molecular dynamics simulations in conjunction with enhanced sampling and data-centric analysis techniques to quantitatively relate changes in interfacial water structure to the hydration free energy (a thermodynamically well-defined descriptor of hydrophobicity) of chemically heterogeneous interfaces. We analyze a large data set of 58 self-assembled monolayers (SAMs) composed of ligands with nonpolar and polar end groups of different chemical identity (amine, amide, and hydroxyl) in five mole fractions, two spatial patterns, and with scaled partial charges. We find that only five features of interfacial water structure are required to accurately predict hydration free energies. Examination of these features reveals mechanistic insights into the interfacial hydrogen bonding behaviors that distinguish different surface compositions and patterns. This analysis also identifies the probability of highly coordinated water structures as a unique signature of hydrophobicity. These insights provide a physical basis to understand the hydrophobicity of chemically heterogeneous interfaces and connect hydrophobicity to experimentally accessible perturbations of interfacial water structure.

Machine learned coarse-grained protein force-fields: Are we there yet?

The successful recent application of machine learning methods to scientific problems includes the learning of flexible and accurate atomic-level force-fields for materials and biomolecules from quantum chemical data. In parallel, the machine learning of force-fields at coarser resolutions is rapidly gaining relevance as an efficient way to represent the higher-body interactions needed in coarse-grained force-fields to compensate for the omitted degrees of freedom. Coarse-grained models are important for the study of systems at time and length scales exceeding those of atomistic simulations. However, the development of transferable coarse-grained models via machine learning still presents significant challenges. Here, we discuss recent developments in this field and current efforts to address the remaining challenges.

Identifying nonadditive contributions to the hydrophobicity of chemically heterogeneous surfaces via dual-loop active learning.

Hydrophobic interactions drive numerous biological and synthetic processes. The materials used in these processes often possess chemically heterogeneous surfaces that are characterized by diverse chemical groups positioned in close proximity at the nanoscale; examples include functionalized nanomaterials and biomolecules, such as proteins and peptides. Nonadditive contributions to the hydrophobicity of such surfaces depend on the chemical identities and spatial patterns of polar and nonpolar groups in ways that remain poorly understood. Here, we develop a dual-loop active learning framework that combines a fast reduced-accuracy method (a convolutional neural network) with a slow higher-accuracy method (molecular dynamics simulations with enhanced sampling) to efficiently predict the hydration free energy, a thermodynamic descriptor of hydrophobicity, for nearly 200 000 chemically heterogeneous self-assembled monolayers (SAMs). Analysis of this dataset reveals that SAMs with distinct polar groups exhibit substantial variations in hydrophobicity as a function of their composition and patterning, but the clustering of nonpolar groups is a common signature of highly hydrophobic patterns. Further molecular dynamics analysis relates such clustering to the perturbation of interfacial water structure. These results provide new insight into the influence of chemical heterogeneity on hydrophobicity via quantitative analysis of a large set of surfaces, enabled by the active learning approach.

Predicting Hydrophobicity by Learning Spatiotemporal Features of Interfacial Water Structure: Combining Molecular Dynamics Simulations with Convolutional Neural Networks.

The hydrophobicity of functionalized interfaces can be quantified by the structure and dynamics of water molecules using molecular dynamics (MD) simulations, but existing methods to quantify interfacial hydrophobicity are computationally expensive. In this work, we develop a new machine learning approach that leverages convolutional neural networks (CNNs) to predict the hydration free energy (HFE) as a measure of interfacial hydrophobicity based on water positions sampled from MD simulations. We construct a set of idealized self-assembled monolayers (SAMs) with varying surface polarities and calculate their HFEs using indirect umbrella sampling calculations (INDUS). Using the INDUS-calculated HFEs as labels and physically informed representations of interfacial water density from MD simulations as input, we train and evaluate a series of neural networks to predict SAM HFEs. By systematically varying model hyperparameters, we demonstrate that a 3D CNN trained to analyze both spatial and temporal correlations between interfacial water molecule positions leads to HFE predictions that require an order of magnitude less MD simulation time than INDUS. We showcase the power of this model to explore a large design space by predicting HFEs for a set of 71 chemically heterogeneous SAMs with varying patterns and mole fractions.