Application of Deep Neural Network Models in Drug Discovery Programs

Abstract

In silico driven optimization of compound properties related to pharmacokinetics, pharmacodynamics, and safety is a key requirement in modern drug discovery. Nowadays, large and harmonized datasets allow to implement deep neural networks (DNNs) as a framework for leveraging predictive models. Nevertheless, various available model architectures differ in their global applicability and performance in lead optimization projects, such as stability over time and interpretability of the results. Here, we describe and compare the value of established DNN-based methods for the prediction of key ADME property trends and biological activity in an industrial drug discovery environment, represented by microsomal lability, CYP3A4 inhibition and factor Xa inhibition. Three architectures are exemplified, our earlier described multilayer perceptron approach (MLP), graph convolutional network-based models (GCN) and a vector representation approach, Mol2Vec. From a statistical perspective, MLP and GCN were found to perform superior over Mol2Vec, when applied to external validation sets. Interestingly, GCN-based predictions are most stable over a longer period in a time series validation study. Apart from those statistical observations, DNN prove of value to guide local SAR. To illustrate this important aspect in pharmaceutical research projects, we discuss challenging applications in medicinal chemistry towards a more realistic picture of artificial intelligence in drug discovery.

Optimizing interactions to protein binding sites by integrating docking-scoring strategies into generative AI methods

“…Their potential difficulty is that a large set of ligands and affinities must be known, before a predictive model can be derived. Especially for novel target proteins, this is not always the case, while for ADMET properties, many validated models have already been described ( Wenzel et al, 2019 ; Goller et al, 2020 ; Aleksić et al, 2021 ; Grebner et al, 2021 ). The design guided by those models often only explores the already known chemical space for that particular target.…”

Interpretation of Structure–Activity Relationships in Real-World Drug Design Data Sets Using Explainable Artificial Intelligence

“…We have recently shown the good performance of GCNN models in drug design applications which motivates the employment of such an approach. 32 Methods for Explaining Neural Networks. It is not a trivial task to explain predictions made by neural networks.…”

“…This information is then compressed into a fixed size vector in the final dense layer, but the nature of the NN training allows the selection of relevant substructure information for the given task and thus is expected to generate a set of descriptors better suited to the machine-learning problem. We have recently shown the good performance of GCNN models in drug design applications which motivates the employment of such an approach …”

Prediction of Maximum Absorption Wavelength Using Deep Neural Networks

“…Other Models. We compare the performance of the FCNN and CNN with other three baseline models: the graph convolutional network (GCN), 31 Attentive FP, 32 and Chem-Fluo. 20 Specifically, the GCN includes a graph convolutional layer to learn the vector representation of molecules and a fully connected layer for regression.…”

“…Since the accuracy in training and prediction of photophysical properties relies on the proper translation of chemical structures into computer language, the other two DNNs suitable for chemical structures were included as well. One is GCNs, 31 a modified version of CNNs with better suitability for extrapolating to new chemical spaces, and the other is Attentive FP, 32 a small molecule representation framework based on a graph neural network.…”