machine learning

Proteochemometric models (PCMs) are used in computational drug discovery to employ both protein and ligand representations jointly for bioactivity prediction. While machine learning (ML) and deep learning (DL) have come to dominate PCMs, often serving as a basis for scoring functions, rigorous evaluation standards have not always been consistently applied. In this study, using kinase-ligand bioactivity prediction as a model system, we highlight the critical roles of data set curation, permutation testing, class imbalances, and various data splitting strategies for mitigating plausible data leakage and embedding quality in determining model performance. Our findings indicate that data splitting and class imbalances are the most critical factors affecting PCM performance, emphasizing the challenges in the generalizing ability of ML/DL-PCMs. We evaluated various protein–ligand descriptors and embeddings, including those augmented with multiple sequence alignment information. However, permutation testing consistently demonstrated that protein embeddings contributed minimally to PCM efficacy. This study advocates for the adoption of stringent evaluation standards to enhance the generalizability of models to out-of-distribution data and improve benchmarking practices.

AbstractThe potential energy of molecular species and their conformers can be computed with a wide range of computational chemistry methods, from molecular mechanics to ab initio quantum chemistry. However, the proper choice of the computational approach based on computational cost and reliability of calculated energies is a dilemma, especially for large molecules. This dilemma is proved to be even more problematic for studies that require hundreds and thousands of calculations, such as drug discovery. On the other hand, driven by their pattern recognition capabilities, neural networks started to gain popularity in the computational chemistry community. During the last decade, many neural network potentials have been developed to predict a variety of chemical information of different systems. Neural network potentials are proved to predict chemical properties with accuracy comparable to quantum mechanical approaches but with the cost approaching molecular mechanics calculations. As a result, the development of more reliable, transferable, and extensible neural network potentials became an attractive field of study for researchers. In this review, we outlined an overview of the status of current neural network potentials and strategies to improve their accuracy. We provide recent examples of studies that prove the applicability of these potentials. We also discuss the capabilities and shortcomings of the current models and the challenges and future aspects of their development and applications. It is expected that this review would provide guidance for the development of neural network potentials and the exploitation of their applicability.This article is categorized under:Data Science > Artificial Intelligence/Machine LearningMolecular and Statistical Mechanics > Molecular InteractionsSoftware > Molecular Modeling

The ability to accurately and efficiently compute quantum-mechanical partial atomistic charges has many practical applications, such as calculations of IR spectra, analysis of chemical bonding, and classical force field parametrization. Machine learning (ML) techniques provide a possible avenue for the efficient prediction of atomic partial charges. Modern ML advances in the prediction of molecular energies [i.e., the hierarchical interacting particle neural network (HIP-NN)] has provided the necessary model framework and architecture to predict transferable, extensible, and conformationally dynamic atomic partial charges based on reference density functional theory (DFT) simulations. Utilizing HIP-NN, we show that ML charge prediction can be highly accurate over a wide range of molecules (both small and large) across a variety of charge partitioning schemes such as the Hirshfeld, CM5, MSK, and NBO methods. To demonstrate transferability and size extensibility, we compare ML results with reference DFT calculations on the COMP6 benchmark, achieving errors of 0.004e- (elementary charge). This is remarkable since this benchmark contains two proteins that are multiple times larger than the largest molecules in the training set. An application of our atomic charge predictions on nonequilibrium geometries is the generation of IR spectra for organic molecules from dynamical trajectories on a variety of organic molecules, which show good agreement with calculated IR spectra with reference method. Critically, HIP-NN charge predictions are many orders of magnitude faster than direct DFT calculations. These combined results provide further evidence that ML (specifically HIP-NN) provides a pathway to greatly increase the range of feasible simulations while retaining quantum-level accuracy.