Gregory L. Warren

Toxicological profiling is a critical component of the regulatory process required to inform the hazard characterisation and labelling of chemicals, thus guiding their safe handling and use. Skin irritation is a key endpoint for regulatory registrations, traditionally assessed by using the Draize rabbit test, which was introduced in 1944. Advances in New Approach Methodologies (NAMs), and particularly those based on reconstructed skin models, have led to more human-relevant alternatives, as compared to the in vivo test method. Here, we report on some groundbreaking high-throughput adaptations to the validated in vitro Skin Irritation Test (SIT; OECD TG 439) with the commercially available EpiDerm™ reconstructed human epidermis (RhE) model from MatTek Corporation. These minor adaptations enabled the safety evaluation of over 3000 chemicals in just ten months. This was necessary due to the large number of chemicals that needed to be pre-screened for dermal safety, after the insect repellent or insecticidal activity was predicted and assessed. The data obtained in the SIT were then used to train a new Artificial Intelligence (AI) tool, designed to predict the skin irritation potential of large libraries of molecules within (but not limited to) the chemical space which the AI tool is based upon. Our work exemplifies the integration of AI with existing non-animal-based approaches, in order to further develop chemical safety assessment by taking advantage of this new technology.

Next Generation Risk Assessment (NGRA) promotes animal-free, exposure-informed, and hypothesis-driven approaches to chemical safety assessment. In silico tools, such as quantitative structure-activity relationship (QSAR) models, are valuable new approach methodologies (NAMs) for use in NGRA. However, the practical implementation of in silico NAMs remains limited by challenges in data availability, heterogeneity, and regulatory acceptance. In this study, federated learning is introduced to advance chemical safety assessment while leveraging proprietary data domains. Federated learning is a decentralised machine learning approach where multiple organisations, devices or servers collaboratively train a model while keeping their data locally, sharing only model updates to preserve confidentiality and privacy. Three use cases were simulated with the Flower open-source federated learning framework, namely (i) federated analytics for dermal permeability (log Kp) screening; (ii) federated convolutional neural networks (CNNs) for mutagenicity prediction from SMILES strings, and (iii) federated eXtreme Gradient Boosting (XGBoost) models for predicting skin sensitisation potential using molecular fingerprints and descriptors. The results show that federated learning approaches can yield predictive performance comparable to centralised models while mitigating concerns over the visibility of, and access to, commercially sensitive data. Open challenges related to data curation, interpretability, and model governance, as well as future directions, are discussed. This work demonstrates that federated learning can facilitate secure collaboration across organisations, enhance the utility of distributed chemical datasets, and accelerate the adoption of in silico NAMs.

Free energy calculations are rapidly becoming indispensable in structure-enabled drug discovery programs. As new methods, force fields, and implementations are developed, assessing their accuracy on real-world systems (benchmarking) becomes critical to provide users with an assessment of the accuracy expected when these methods are applied within their domain of applicability, and developers with a way to assess the expected impact of new methodologies. These assessments require construction of a benchmark—a set of well-prepared, high quality systems with corresponding experimental measurements designed to ensure the resulting calculations provide a realistic assessment of expected performance. To date, the community has not yet adopted a common standardized benchmark, and existing benchmark reports suffer from a myriad of issues, including poor data quality, limited statistical power, and statistically deficient analyses, all of which can conspire to produce benchmarks that are poorly predictive of real-world performance. Here, we address these issues by presenting guidelines for (1) curating experimental data to develop meaningful benchmark sets, (2) preparing benchmark inputs according to best practices to facilitate widespread adoption, and (3) analysis of the resulting predictions to enable statistically meaningful comparisons among methods and force fields. We highlight challenges and open questions that remain to be solved in these areas, as well as recommendations for the collection of new datasets that might optimally serve to measure progress as methods become systematically more reliable. Finally, we provide a curated, versioned, open, standardized benchmark set adherent to these standards (protein-ligand-benchmark) and an open source toolkit for implementing standardized best practices assessments (arsenic) for the community to use as a standardized assessment tool. While our main focus is benchmarking free energy methods based on molecular simulations, these guidelines should prove useful for assessment of the rapidly growing field of machine learning methods for affinity prediction as well.

Part of early stage drug discovery involves determining how molecules may bind to the target protein. Through understanding where and how molecules bind, chemists can begin to build ideas on how to design improvements to increase binding affinities. In this retrospective study, we compare how computational approaches like docking, molecular dynamics (MD) simulations, and a non-equilibrium candidate Monte Carlo (NCMC) based method (NCMC+MD) perform in predicting binding modes for a set of 12 fragment-like molecules which bind to soluble epoxide hydrolase. We evaluate each method's effectiveness in identifying the dominant binding mode and finding any additional binding modes (if any). Then, we compare our predicted binding modes to experimentally obtained X-ray crystal structures. We dock each of the 12 small molecules into the apo-protein crystal structure and then run simulations up to 1 microsecond each. Small and fragment-like molecules likely have smaller energy barriers separating different binding modes by virtue of relatively fewer and weaker interactions relative to drug-like molecules, and thus likely undergo more rapid binding mode transitions. We expect, thus, to see more rapid transitions between binding modes in our study.

To compare ordered water positions from experiment with those from molecular dynamics (MD) simulations, a number of MD models of water structure in crystalline endoglucanase were calculated. The starting MD model was derived from a joint X-ray and neutron diffraction crystal structure, enabling the use of experimentally assigned protonation states. Simulations were performed in the crystalline state, using a periodic 2 × 2 × 2 supercell with explicit solvent. Water X-ray and neutron scattering density maps were computed from MD trajectories using standard macromolecular crystallography methods. In one set of simulations, harmonic restraints were applied to bias the protein structure toward the crystal structure. For these simulations, the recall of crystallographic waters using strong peaks in the MD water electron density was very good, and there also was substantial visual agreement between the boomerang-like wings of the neutron scattering density and the crystalline water hydrogen positions. An unrestrained simulation also was performed. For this simulation, the recall of crystallographic waters was much lower. For both restrained and unrestrained simulations, the strongest water density peaks were associated with crystallographic waters. The results demonstrate that it is now possible to recover crystallographic water structure using restrained MD simulations but that it is not yet reasonable to expect unrestrained MD simulations to do the same. Further development and generalization of MD water models for force-field development, macromolecular crystallography, and medicinal chemistry applications is now warranted. In particular, the combination of room-temperature crystallography, neutron diffraction, and crystalline MD simulations promises to substantially advance modeling of biomolecular solvation.