The Importance of the Regression Model in the Structure-Based Prediction of Protein-Ligand Binding

Abstract

Abstract. Docking is a key computational method for structure-based design of starting points in the drug discovery process. Recently, the use of nonparametric machine learning to circumvent modelling assumptions has been shown to result in a large improvement in the accuracy of docking. As a result, these machine-learning scoring functions are able to widely outperform classical scoring functions. The latter are characterized by their reliance on a predetermined theory-inspired functional form for the relationship between the variables that characterise the complex and its predicted binding affinity.In this paper, we demonstrate that the superior performance of machinelearning scoring functions comes from the avoidance of the functional form that all classical scoring functions assume. These scoring functions can now be directly applied to the docking poses generated by AutoDock Vina, which is expected to increase its accuracy. On the other hand, as it is well known that the assumption of additivity does not hold in some cases, it is expected that the described protocol will also improve other classical scoring functions, as it has been the case with Vina. Lastly, results suggest that incorporating ligand-and protein-only properties into a model is a promising avenue for future research.

Protein–Ligand Scoring with Convolutional Neural Networks

“…The risk of overfitting increases the importance of rigorous validation, 26,27 but the inherent increase in flexibility allows machine learning methods to outperform more constrained methods when trained on an identical input set. 28 The choice of input features can limit the expressiveness of a machine learning method. Features such as atom interaction counts, 22 pairwise atom distance descriptors, 13 interaction fingerprints, 21 or “neural fingerprints” generated by learned atom convolutions 24 necessarily eliminate or approximate the information inherent in a protein-ligand structure, such as precise spatial relationships.…”

Protein Family-Specific Models Using Deep Neural Networks and Transfer Learning Improve Virtual Screening and Highlight the Need for More Data

“…Traditional approaches have typically used experimental data to parametrize a physically inspired function. − While interpretable, these techniques are inherently limited in their ability to capture complex interactions due to the use of rigid functional forms. Many machine learning-based scoring functions reuse the features of traditional approaches , but exploit the greater flexibility in model structure to produce better representations of the same input data . However, this can lead to overfitting and often results in a loss of interpretability.…”

“…Many machine learning-based scoring functions reuse the features of traditional approaches 8,17 but exploit the greater flexibility in model structure to produce better representations of the same input data. 18 However, this can lead to overfitting and often results in a loss of interpretability. In addition, the use of specific features, such as descriptors 17,19 or fingerprints, 20 both biases the model to the choice of features and leads to an unnecessary loss of information through the elimination or approximation of the raw structural data.…”

Learning RMSD to Improve Protein-Ligand Scoring and Pose Selection

“…Machine learning based models, on the other hand, are not necessarily bound by such constraints and have the potential to learn non linear relationships and capture binding features that are hard to model explicitly [16]. This has become an increasingly popular approach to model protein-ligand scoring functions.…”