Target identification among known drugs by deep learning from heterogeneous networks

Abstract

Without foreknowledge of the complete drug target information, development of promising and affordable approaches for effective treatment of human diseases is challenging. Here, we develop deepDTnet, a deep learning methodology for new target identification and drug repurposing in a heterogeneous drug-gene-disease network embedding 15 types of chemical, genomic, phenotypic, and cellular network profiles. Trained on 732 U.S. Food and Drug Administration-approved small molecule drugs, deepDTnet shows high accuracy (the area under the receiver operating characteristic curve = 0.963) in identifying novel molecular targets for known drugs, outperforming previously published state-of-the-art methodologies. We then experimentally validate that deepDTnet-predicted topotecan (an approved topoisomerase inhibitor) is a new, direct inhibitor (IC₅₀ = 0.43 μM) of human retinoic-acid-receptor-related orphan receptor-gamma t (ROR-γt). Furthermore, by specifically targeting ROR-γt, topotecan reveals a potential therapeutic effect in a mouse model of multiple sclerosis. In summary, deepDTnet offers a powerful network-based deep learning methodology for target identification to accelerate drug repurposing and minimize the translational gap in drug development.

Fig. 1. A diagram illustrating the workflow of deepDTnet. DeepDTnet embeds the 15 types of chemical, genomic, phenotypic, and cellular networks and applies a deep neural network algorithm to learn a low-dimensional vector representation of the features for each node (see ESI Methods†). After learning the feature matrix X and Y for drugs and targets (i.e., each row in X and Y represents the feature vector of a drug or a target, respectively), deepDTnet applies PU-matrix completion to find the best projection from the drug space onto target (protein) space, such that the projected feature vectors of drugs are geometrically close to the feature vectors of their known interacting targets. Finally, deepDTnet infers new targets for a drug ranked by geometric proximity to the projected feature vector of the drug in the projected space (see Methods).

Fig. 2. A workflow illustrating the network embedding and performance of deepDTnet. (A) The deep neural networks model for graph representations (DNGR) consists of three major steps: (i) a random surfing model to capture the graph structural information and generate a probabilistic co-occurrence (PCO) matrix; (ii) calculation of the shifted positive pointwise mutual information (PPMI) matrix based on the probabilistic co-occurrence matrix; and (iii) a stacked denoising autoencoder to generate compressed, low-dimensional vectors from the original high-dimensional vertex vectors. The learned low-dimensional feature vectors encode the relational properties, association information, and topological context of each node in the heterogeneous drug–gene–disease network (see Methods). (B and C) Performance of deepDTnet was assessed by both (B) the area under the receiver operating characteristic curve (AUROC) and (C) the area under the precision-recall curve (AUPR) of deepDTnet against top k predicted list during cross-validation. The experimentally validated drug–target interactions (Table S3†) are used to evaluate the model performance.

Fig. 3. Visualization of the learned drug and target vectors. Visualization of the drug vector matrix and protein vector matrix learned by network embedding using the t-SNE (t-distributed stochastic neighbor embedding algorithm). (A) The two-dimensional (2D) representation of the learned vectors for 14 types of drugs grouped by the first-level of the Anatomical Therapeutic Chemical Classification System codes (http://www.whocc.no/atc/). We can observe that semantically similar drugs are mapped to nearby representations. We assigned the drugs with multiple ATC codes based on two criteria: (1) the majority rule of ATC codes, and (2) manually checked and assigned by experts based on common clinical uses. (B) An illustration of the learned vectors for four well-known drug target families: G-protein-coupled receptors (GPCRs), kinases, nuclear receptors (NRs), and ion channels (ICs), non-linearly projected to 2D space for visualization by the t-SNE algorithm.

Fig. 4. The uncovered drug–target network via deepDTnet. (A) Computationally predicted drug–target networks for four well-known drug target families: G-protein-coupled receptors (GPCRs), kinases, nuclear receptors (NRs), and ion channels (ICs). Drugs are grouped by the first-level of the Anatomical Therapeutic Chemical classification system (ATC) codes (http://www.whocc.no/atc/). Drug targets comprise four groups, GPCRs, kinases, NRs, and ICs. (B) An illustration of the mechanisms-of-action of the deepDTnet-predicted GPCRs for three approved drugs validated by a recent high-throughput screening assay, for characterizing the mechanisms-of-action of their clinically reported adverse events. The experimental data for the predicted the drug–target interactions and the target-adverse events were collected from a recent study. The clinically reported adverse events of known drugs were collected from metaADEDB.

Fig. 5. DeepDTnet-predicted topotecan is a novel ROR-γt antagonist. (A) The screening results of 18 deepDTnet-predicted drugs at 10 μM in Gal4-based ROR-γt luciferase assay. (B) Topotecan (TPT) exhibits dose-dependent inhibition of ROR-γt transcriptional activity in Gal4-based luciferase reporter system. (C) TPT reveals dose-dependent inhibition of ROR-γt LBD and cofactor peptide SRC1 interaction in HTRF assay. (D) Induced circular dichroism (CD) spectra reveals the direct binding of TPT to ROR-γt LBD. Data are representative of three independent experiments. (E) High-performance liquid chromatography (HPLC) experiment indicates the binding of TPT to recombinant ROR-γt-LBD. (F) The predicted ligand-protein binding mode between TPT and ROR-γt using molecular docking (see Methods).

Fig. 6. DeepDTnet-predicted topotecan (TPT) reverses experimental autoimmune encephalomyelitis (EAE), a mouse model of multiple sclerosis. (A) An illustration of induction and treatment of EAE. (B) Mean clinical scores of EAE in vehicle- or TPT-treated group (n = 10/group). TPT (10 mg kg⁻¹) or vehicle is intraperitoneal administered on day 11 after immunization every four days. Data are presented as the mean ± SEM of eight mice per group. Student's t-test is revealed, *P < 0.05, **P < 0.01. (C) The body weight of mice in vehicle- or TPT-treated group. Student's t-test is revealed, *P < 0.05. (D) Section of spinal cord tissue is prepared on day 20 post immunization and subjected to hematoxylin and eosin (H&E) staining and Luxol fast blue (LFB) staining. (E) In vivo imaging of myelination using myelin-binding dye, 3,3-diethylthiatricarbocyanine iodide (DBT) on day 20 after immunization. DBT dye readily enters the brain and specifically binds to myelinated fibers. (F) In vivo imaging of the blood–brain barrier integrity using Cy5.5-BSA on day 20 after immunization. Cy5.5-BSA uptake in the brain when the BBB (blood–brain barrier) integrity is disrupted. (G) ELISA analysis of IL-17 production of spinal cords and brain from vehicle- or TPT-treated EAE mice on day 20 after immunization. Data are presented as the mean ± SEM. Student's t-test is revealed, **P < 0.01. (H and I) Concentration of T0901317 in mice brain samples (H) and plasma (I). T0901317 (ref. 56), an orthosteric ligand of ROR-γt, was used as the tracer for assessing target occupancy of TPT in the mouse model. Student's t-test was performed and sterile water was used as vehicle.