A Bayesian machine learning approach for drug target identification using diverse data types

Abstract

Drug target identification is a crucial step in development, yet is also among the most complex. To address this, we develop BANDIT, a Bayesian machine-learning approach that integrates multiple data types to predict drug binding targets. Integrating public data, BANDIT benchmarked a ~90% accuracy on 2000+ small molecules. Applied to 14,000+ compounds without known targets, BANDIT generated ~4,000 previously unknown molecule-target predictions. From this set we validate 14 novel microtubule inhibitors, including 3 with activity on resistant cancer cells. We applied BANDIT to ONC201-an anti-cancer compound in clinical development whose target had remained elusive. We identified and validated DRD2 as ONC201's target, and this information is now being used for precise clinical trial design. Finally, BANDIT identifies connections between different drug classes, elucidating previously unexplained clinical observations and suggesting new drug repositioning opportunities. Overall, BANDIT represents an efficient and accurate platform to accelerate drug discovery and direct clinical application.

Fig. 1

BANDIT exploits the individual predictive powers of each data type. a Density plots showing how various different similarity scores correlate with one another, with darker area corresponding to a higher density of values. R² and P value were calculated using a pearson correlation. b Distributions of similarity scores across two sets—drug pairs known to share a target and those with no known shared targets. P values and D statistics were calculated using the Kolmogorov–Smirnov test. c Schematic of BANDIT’s method of integrating multiple data types to predict shared target drug pairs

Fig. 2

BANDIT can accurately predict shared targets and specific target interactions. a Area under the receiver-operating curve for different sets of data types. SE = Side effects; C = CMap; N = NCI60; B = Bioassays; S = Structure. b Ratio of true positives to false positives at different likelihood ratio cutoffs. c Schematic of the BANDIT voting schematic for predicting specific target interactions. d Accuracy level of BANDIT’s voting algorithm at various likelihood ratio cutoffs. e Schematic of two proposed operating scenarios for BANDIT

Fig. 3

Microtubules are a correct target of the newly identified small molecules. Effect of various compounds (1 μM) on the microtubule integrity of MDA-MB-231 cells after 6 h of treatment. a Control with DMSO (Scale bar: 5 μm), b Vinblastine as a positive control, c Compound #16, d Compound #15, e Compound #24 f Compound #2. g Dose dependent effect of Compound #12 and h Compound #13. i Box plot showing the % tubulin in the pellet compared to the supernatant for depolymerizing drugs at 1 and 10 μM. The median is denoted by the center line and the min/max are represented by the whiskers. Each individual replicate is represented by a point in the box plot

Fig. 4

BANDIT predicted small molecules can act on resistant cells. Effect of various compounds on the microtubule integrity of 1A9-ERB cells after 6 h of treatment: a Control with DMSO (Scale bar: 5 μm), 100 nM of b Eribulin and c Vinblastine, and 1 μM of d Compound #15, e Compound #16 and f Compound #24

Fig. 5

ONC201 is a selective DRD2 antagonist. a BANDIT target predictions for ONC201. Connections between ONC201 and known drugs are weighted based on the likelihood ratio and predicted targets are sized based on the prediction strength. b Antagonism of ligand-stimulated dopamine receptors by ONC201. c Schild analysis of DRD2L antagonism by ONC201 using arrestin recruitment or d cAMP modulation reporters. Error bars represent 1 standard deviation

Fig. 6

BANDIT can predict specific mechanisms of action and connections between drug classes. a Hierarchical clustering of drugs known to target microtubules and b drugs known to target protein kinases. c Network of drugs based on shared target interactions. Drugs are colored based on their most prevalent ATC code. Three specific clusters corresponding to beta-blockers and Parkinson’s medications, anti-retrovirals and statins, and opioids and antimicrotubule drugs are highlighted