From jeopardy champion to drug discovery; semantic similarity artificial intelligence

ABSTRACT

We have employed artificial intelligence to streamline the small molecule drug screening pipeline and identified the cholesterol-reducing compound probucol in the process. Probucol augmented mitophagy and prevented loss of dopaminergic neurons in flies and zebrafish challenged with mitochondrial toxins. Further dissection of the mechanism of action led to the identification of ABCA1, the target of probucol, as a mitophagy modulator. Probucol treatment regulates lipid droplet dynamics during mitophagy and ABCA1 is required for these effects. Here we will summarize the combination of in silico and cell-based screening that led us to identify and characterize probucol as a compound that enhances mitophagy and include thoughts about future directions for the topics explored in our study.

Abbreviations: ABCA1: ATP binding cassette transporter protein 1; ATP: Adenosine tri-phosphate; CCCP: carbonyl cyanide m-chlorophenylhydrazone; DsRed: Discosoma red; FDA: Food and drug administration; GFP: Green fluorescent protein; LAMP: lysosome-associated membrane glycoproteins; LD: Lipid droplet; PD: Parkinson’s disease; PINK: PTEN-induced kinase

Since mitochondrial damage and mitophagy dysfunction are well-defined aspects of Parkinson’s disease (PD) pathogenesis, bolstering mitophagy, a process which mitigates mitochondrial damage may be advantageous for the survival of dopaminergic neurons and thus represents a therapeutic approach to reduce PD progression. This logic emboldened us, as well as other researchers, to search for genetic and pharmacological treatments that augment mitophagy. Currently, most molecules enhance mitophagy function by exerting mitochondrial damage, which is not be suitable for the development of therapeutics. To identify new molecules with more suitable mechanisms of action, we used artificial intelligence techniques based on natural language processing, instead of randomly screening large collections of molecules in high throughput.

A predictive model was created based on several previously characterized mitophagy agonists. We deployed our predictive model to select the best candidates from the Drugbank database, a repository of FDA-approved small molecules, biologics, nutraceuticals, and experimental drugs (www.drugbank.com). The major benefit of exploring this chemical space is the repurposing potential of the compounds. Since the initiation of our study, additional mitophagy-augmenting molecules have been identified, which presents the opportunity to re-train the predictive model and perform an updated analysis. However, the current predictive model can also be deployed against other repositories of molecules.

We selected the top 79 compounds and screened for their ability to augment the clearance of mitochondria damaged with carbonyl cyanide m-chlorophenylhydrazone (CCCP) in cells based on a fluorescence microscopy assay (Figure 1). The expectation for a mitophagy agonist is that it increases the percentage of cells that have lost mitochondrial signal, in the form of fluorophore DsRed targeted to the mitochondria. This approach differs from other mitophagy-related screens that have focused on upstream steps in the PINK1/Parkin mitophagy pathway, such as recruitment of Parkin to depolarized mitochondria. By using a readout at the endpoint of mitophagy, which is agnostic to the method of mitochondrial degradation, it is possible to cast a wider net for the identification of hits. One can then work backwards to determine precisely which step in mitophagy the compound impacts. Using this approach, the cholesterol-lowering drug probucol was identified as a mitophagy agonist [1].

Figure 1.

In silico screening is coupled to cell-based screening to identify mitophagy agonists, which may be tested on PD-relevant phenotypes in flies and zebrafish. Text fingerprints were generated for a list of positive control mitophagy-augmenting molecules in addition to a candidate list of chemicals from the DrugBank database. The similarity of the candidate compounds to the positive controls is evaluated and the most similar compounds are selected and arrayed in a 96-well format. This pilot library is tested on cells expressing GFP-Parkin and the mitochondrial marker mito-DsRed, to identify molecules that augment the removal of mitochondria damaged with CCCP. Screening in cells yielded the lead compound, probucol, for further testing in both fly and zebrafish models of PD. Aspects of organism health such as locomotor function, survival and dopaminergic neuron survival were evaluated.

Testing probucol in flies and zebrafish revealed improvements to several PD-relevant phenotypes such as dopaminergic neuron degeneration, and locomotor, and lifespan decline. Importantly, probucol augmented mitophagy in the dopaminergic neurons of flies that were challenged with the PD-causing mitochondrial toxin paraquat. Further dissection of the mechanism of action connected probucol’s target, the ATP binding cassette transporter protein 1 (ABCA1), and its effects on lipid droplet (LD) expansion under conditions of mitochondrial dysfunction, to probucol’s ability to enhance mitophagy. ABCA1 localizes to both the cell surface and to intracellular vesicles, so future studies should attempt to determine which subcellular population is relevant for mitophagy and LD dynamics.

We captured changes to LDs at two discrete time points-prior to and following extended mitochondrial depolarization with CCCP. Under basal conditions, probucol treatment increased contacts between LDs and mitochondria. LDs accumulate upon prolonged mitochondrial depolarization, but the extent to which this occurs is impaired by probucol treatment. In vivo, paraquat treatment elevates LD area, while the addition of probucol returned LD area back to basal levels. While these findings provide a starting point, temporal LD dynamics in response to probucol remain to be investigated. Capturing changes to LD area over the entire course of the 24 h mitochondrial depolarization using live-cell imaging could be informative, especially if markers for other organelles of interest are included. Probucol may prevent LD area expansion altogether, or LD expansion may occur as usual, but LDs simply undergo subsequent degradation by the 24 h time point. It remains to be determined whether the LDs interact with other organelles. LDs may interact with endosomes containing ABCA1 or with the lysosome-associated membrane glycoproteins (LAMP)-positive lysosomes, which are increased under basal conditions by probucol treatment. How these interactions are shaped by probucol treatment may help to further elucidate the mechanism of how this drug modulates mitophagy.

Probucol treatment increases the proximity of mitochondria and LDs, prior to the induction of mitochondrial damage. While LDs themselves are important for probucol-mediated mitophagy enhancement, the requirement for the LD-mitochondria contacts has not been established. Since probucol is an ABCA1 inhibitor and functions differently from modern cholesterol-reducing drugs such as statins, the effects of cholesterol-lowering drugs with diverse mechanisms of action on mitophagy and LD dynamics could be assessed in the future as potential therapeutic approaches in mitochondrial-dependent neurodegenerative diseases.

Funding Statement

The work was supported by the Canadian Institutes of Health Research.

Disclosure statement

No potential conflict of interest was reported by the author(s).