A deep learning drug screening framework for integrating local-global characteristics: A novel attempt for limited data

Abstract

At the beginning of the "Disease X" outbreak, drug discovery and development are often challenged by insufficient and unbalanced data. To address this problem and maximize the information value of limited data, we propose a drug screening model, LGCNN, based on convolutional neural network (CNN), which enables rapid drug screening by integrating features of drug molecular structures and drug-target interactions at both local and global (LG) levels. Experimental results show that LGCNN exhibits better performance compared to other state-of-the-art classification methods under limited data. In addition, LGCNN was applied to anti-SARS-CoV-2 drug screening to realize therapeutic drug mining against COVID-19. LGCNN transcends the limitations of traditional models for predicting interactions between single drug targets and shows new advantages in predicting multi-target drug-target interactions. Notably, the cross-coronavirus generalizability of the model is also implied by the analysis of targets, drugs, and mechanisms in the prediction results. In conclusion, LGCNN provides new ideas and methods for rapid drug screening in emergency situations where data are scarce.

Keywords: COVID-19; Deep learning; Direct-acting antiviral drugs; Host-targeted antiviral drugs; Molecular docking; Virtual drug screening.

Graphical abstract

Fig. 1

The overall workflow of the study. The combination of drug substructure features and drug-target association features is embedded as comprehensive feature vectors, which is trained by a convolutional neural network composed of double convolutional architecture. The performance of LGCNN and the accuracy of the predicted results are evaluated by multiple cross-validation and molecular docking simulation experiments. Drug-target network analysis and therapeutic mechanism analysis are used to explain the reliability of the predicted results.

Fig. 2

Diagram of feature extraction and integration. Drug substructure features and drug-target association features are integrated into local-global comprehensive features.

Fig. 3

The architecture of the CNN model.

Fig. 4

LGCNN performance diagram. (A) Loss values and accuracy change figures of the training set and test set (Positive sample: negative sample = 1:1). (B) ROC and PR curves of the LGCNN. (C) LGCNN cross-validation results based on different datasets.

Fig. 5

ROC curves obtained by LGCNN and other competing methods on the limited datasets based on 5-fold CV.

Fig. 6

Drug-target interaction network. (A) Complete drug-target interaction network predicted by LGCNN. (B) Drug-target interaction subnetwork for multi-target drugs (cid). (C) Drug-target interaction subnetworks for targets modulated by multiple drugs (cid).

Fig. 7

Virtual docking diagrams of ligands and receptors. 3D interactions of ligand-receptor complex (left) and 2D interactions of ligand-receptor complex (right) in each subdiagram. (A) CHEMBL2263447-DHODH docking complex. (B) SCHEMBL18785395-SIG1R docking complex. (C) CHEMBL473984-MAPK14 docking complex. (D) CHEMBL474178-MAPK14 docking complex.

Fig. 8

Compound-mechanism-target diagram. This diagram illustrates the possible pathways through which high-scoring predictive drugs may treat COVID-19 via target engagement. Each target is assigned a serial number. Targets involved in viral infection replication processes are indicated by blue boxes, while targets implicated in the inflammatory response triggered by viral invasion are marked with brown boxes. Targets involved in both processes are labeled green. The example compound CHEMBL ID and the chemical formula corresponding to each pathway are highlighted in red. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)