Virtual Screening of Molecules via Neural Fingerprint-based Deep Learning Technique

Abstract

A machine learning-based drug screening technique has been developed and optimized using convolutional neural network-derived fingerprints. The optimization of weights in the neural network-based fingerprinting technique was compared with fixed Morgan fingerprints in regard to binary classification on drug-target binding affinity. The assessment was carried out using six different target proteins using randomly chosen small molecules from the ZINC15 database for training. This new architecture proved to be more efficient in screening molecules that less favorably bind to specific targets and retaining molecules that favorably bind to it.

Scientific contribution: We have developed a new neural fingerprint-based screening model that has a significant ability to capture hits. Despite using a smaller dataset, this model is capable of mapping chemical space similar to other contemporary algorithms designed for molecular screening. The novelty of the present algorithm lies in the speed with which the models are trained and tuned before testing its predictive capabilities and hence is a significant step forward in the field of machine learning-embedded computational drug discovery.

Keywords: Artificial neural network; deep learning; drug design; machine learning; neural network-based fingerprinting; structure-based inhibitor design.

Figure 1

The folds of the enzymes and the active site-bound small molecules (shown in the insets) for a) ACEase with dihydrotanshinone I, b) GLTase with glutathione, c) PAPase with a-benzyl-aminobenzyl-phosphonic acid, d) PTPase with ligand ID: 527, e) QR1ase with ligand ID: ES936, and f) QR2ase with menadione.

Figure 2

The receiver operative characteristic (ROC) curves for the six models, namely ACEase, GLTase, PAPase, PTPase, QR1ase and QR2ase. The AUC values are given in Table 1.

Figure 3

The precision-recall (PR) curves for the six models, namely ACEase, GLTase, PAPase, PTPase, QR1ase and QR2ase. These curves are plotted based on the best selected model for each protein alongside corresponding PR-AUC values given in Table 1.

Figure 4.

Predictive enrichment probability graphs and the total number of molecules with which the PEP was calculated over the range of ${\Delta }_{\mathrm{b}\mathrm{i}\mathrm{n}\mathrm{d}}{G}^{\mathrm{o}}(aq)$ values in the binary ‘hit’ range. Note that the counts displayed as bar graphs are not the number of correctly predicted molecules but rather the total number of molecules in the test dataset at each ${\Delta }_{\mathrm{b}\mathrm{i}\mathrm{n}\mathrm{d}}{G}^{\mathrm{o}}(aq)$ value.

Figure 5

A visual assessment of the superiority of the neural fingerprint-based model in predicting substructure similarity between molecules. The similarity maps were calculated using random forest classification. The darker green contour lines indicate greater similarity, and the red atoms indicate dissimilarity.