Predicting therapeutic and side effects from drug binding affinities to human proteome structures

Abstract

Evaluation of the binding affinities of drugs to proteins is a crucial process for identifying drug pharmacological actions, but it requires three dimensional structures of proteins. Herein, we propose novel computational methods to predict the therapeutic indications and side effects of drug candidate compounds from the binding affinities to human protein structures on a proteome-wide scale. Large-scale docking simulations were performed for 7,582 drugs with 19,135 protein structures revealed by AlphaFold (including experimentally unresolved proteins), and machine learning models on the proteome-wide binding affinity score (PBAS) profiles were constructed. We demonstrated the usefulness of the method for predicting the therapeutic indications for 559 diseases and side effects for 285 toxicities. The method enabled to predict drug indications for which the related protein structures had not been experimentally determined and to successfully extract proteins eliciting the side effects. The proposed method will be useful in various applications in drug discovery.

Keywords: Biocomputational method; Bioinformatics; Biological sciences; Natural sciences; Pharmacoinformatics.

Graphical abstract

Figure 1

Workflow of the proposed drug discovery methods using AlphaFold (A) Structures of the drugs were obtained from KEGG DRUG. (B) 3D structural data for all human proteins were retrieved from AlphaFoldDB. (C) Ligand-binding pockets were detected for all human protein structures, and the binding affinities for all drug-protein pairs were assessed. (D) Estimated binding affinities were compiled to generate the proteome-wide binding affinity score (PBAS) profiles. (E and F) Using PBAS profiles, the prediction of potential drug therapeutic indications (E) and side effects (F) was conducted.

Figure 2

Proteome-wide binding affinity score (PBAS) profile and clustering analysis The PBAS profiles created from the results of docking simulations for all drugs with all human proteins are shown in the center of the heatmap. The estimated binding affinities are expressed as a gradient, with the horizontal and vertical axes representing proteins (19,135 structures) and drugs (7,582 drugs), respectively. Proteins and drugs are clustered separately, and the labels are shown for each cluster. The dendrogram on the left side of and under the heatmap shows the clusters of drugs (D1–D8) and proteins (P1–P5), respectively. Enrichment analysis was performed for the protein and drug clusters. The table above the heatmap reflects the results of the enrichment analysis of the structural domain groups of InterPro for the protein clusters. The number of proteins in each structural domain group is shown as a gradient. The table on the right side of the heatmap contains the outcomes of the enrichment analysis for the Anatomical Therapeutic Chemical (ATC) classification system groups for the drug clusters. The number of drugs enriched in the ATC group is marked as a gradient.

Figure 3

Small part of the drug-protein-disease network predicted by the proposed methods using proteome-wide binding affinity score profiles Blue circles, green rhombuses, and red rectangles stand for drugs, diseases, and proteins, respectively. Orange broken and gray lines indicate the relationships between the drugs and proteins obtained from the docking simulation results and those between proteins and possible disease indications, respectively.

Figure 4

Protein-compound binding structures estimated by docking simulation Compounds and proteins are shown as a red rod and an ivory ribbon model, respectively. (A) Overall structure of PTGFR docked with cloprostenol. (B) Overall structure of SLC5A2 conjugated with ertugliflozin. (C) Overall structure of OPRM1 merged with buprenorphine. All protein structures were estimated using AlphaFold.

Figure 5

Performance evaluation on the side effect prediction Asterisks indicate level of statistical significance: ∗∗p < 0.01. Upper two panels show the AUC (left) and AUPR (right) scores for the SIDER dataset, and lower two panels show the scores for the FARES dataset. The results of replicates (n = 30) of cross-validation experiments for each of the seven profiles are shown. PBAS + Fingerprint, PBAS + TESS, and PBAS + TELR represent the integrated profiles that combine PBAS with the other three profiles.

Figure 6

Small part of the drug-protein-side effect network predicted using our method Blue circles, red rectangles, and green rhombuses indicate drugs, proteins, and side effects, respectively. Orange-dashed and green-dotted lines show the drug-protein relationships estimated from the docking simulation and the relationships between side effects and proteins with high weights in the predictive model, respectively.