Machine Learning for Discovery of GSK3β Inhibitors

Abstract

Alzheimer’s disease (AD) is the most common cause of dementia, affecting approximately 35 million people worldwide. The current treatment options for people with AD consist of drugs designed to slow the rate of decline in memory and cognition, but these treatments are not curative, and patients eventually suffer complete cognitive injury. With the substantial amounts of published data on targets for this disease, we proposed that machine learning software could be used to find novel small-molecule treatments that can supplement the AD drugs currently on the market. In order to do this, we used publicly available data in ChEMBL to build and validate Bayesian machine learning models for AD target proteins. The first AD target that we have addressed with this method is the serine–threonine kinase glycogen synthase kinase 3 beta (GSK3β), which is a proline-directed serine–threonine kinase that phosphorylates the microtubule-stabilizing protein tau. This phosphorylation prompts tau to dissociate from the microtubule and form insoluble oligomers called paired helical filaments, which are one of the components of the neurofibrillary tangles found in AD brains. Using our Bayesian machine learning model for GSK3β consisting of 2368 molecules, this model produced a five-fold cross validation ROC of 0.905. This model was also used for virtual screening of large libraries of FDA-approved drugs and clinical candidates. Subsequent testing of selected compounds revealed a selective small-molecule inhibitor, ruboxistaurin, with activity against GSK3β (avg IC₅₀ = 97.3 nM) and GSK3α (IC₅₀ = 695.9 nM). Several other structurally diverse inhibitors were also identified. We are now applying this machine learning approach to additional AD targets to identify approved drugs or clinical trial candidates that can be repurposed as AD therapeutics. This represents a viable approach to accelerate drug discovery and do so at a fraction of the cost of traditional high throughput screening.

Introduction

An estimated 50 million people globally are living with dementia, and that number is expected to triple by the year 2050.¹ Alzheimer’s disease (AD) is the most common cause of dementia, responsible for approximately 60% of cases, affects 10% of people over the age of 65, and is currently the sixth leading cause of death in the United States.² The available FDA-approved treatments for those suffering with AD are designed to slow the cognitive decline associated with the underlying neurodegeneration, not stop or slow the disease progress, and there have been no new treatments approved since 2003.³ Because of the societal and economic burden this disease continues to present, and despite mixed reports regarding the efficacy of some late-stage therapeutics,^4,5 the search continues for additional treatments for AD. One potential strategy to streamline this search is to repurpose molecules that have already been shown to have clinical safety. Combining this strategy with machine learning models can further reduce the time and cost required to identify potential new therapeutics from molecules already tested or approved for other diseases.

The beta isoform of the serine–threonine kinase glycogen synthase kinase 3 (GSK3β) is a promising drug target for several neurological diseases, including AD, Parkinson’s disease, schizophrenia, and bipolar disorder, as well as other diseases involving energy metabolism and cell death, like diabetes and cancer.⁶⁻⁹ The reason that this particular kinase is so broadly applicable in several diseases is that it is a constitutively active protein kinase with nearly 40 protein substrates and is regulated by the Wnt, insulin, and brain-derived neurotropic factor (BDNF) signaling pathways.¹⁰ Activated GSK3β has been found at high levels in the brains of AD patients and has been implicated in improper amyloid precursor protein (APP) processing and impaired lysosomal clearance of protein.¹¹⁻¹⁵ There is then a feed-forward mechanism whereby the neurotoxic amyloid-β peptide inhibits Wnt signaling, prohibiting the deactivating phosphorylation of GSK3β.¹⁶ GSK3β is also the main kinase responsible for phosphorylation of the microtubule-stabilizing protein tau.¹⁷⁻¹⁹ Hyperphosphorylated tau (pTau) is insoluble and forms paired helical filaments, which are the main component of the intracellular neurofibrillary tangles that are canonical hallmarks of AD.^18,20,21 GSK3β also plays a role in apoptosis through activation of the tumor suppressor protein p53, and inhibition of GSK3β in mouse models of AD reduces amyloid-β deposition and associated neuronal death.²²⁻²⁵ Numerous studies have looked at small-molecule GSK3β inhibitors as potential therapeutics in AD, and many of these displayed beneficial effects in mouse models of the disease.^23,25−30 Early GSK3β inhibitors displayed biliary toxicity in dogs and humans, and of the dozens of GSK3β inhibitors available, only a handful of these have made it to clinical trials.^27,31−33 Two of these, lithium and tideglusib, were tested for efficacy in treating AD, but neither proved to be an effective treatment for the disease.³⁴ GSKβ therefore remains an elusive target for drug discovery. One advantage of repurposing drugs for new diseases is that the safety and tolerability has already been determined, so trials can proceed more quickly to efficacy determination (phase II).

In this study, we have applied machine learning techniques using our software Assay Central³⁵⁻⁴⁴ to build models from publicly available screening data (CHEMBL and PubChem) to predict compounds that could potentially inhibit GSK3β. Using this approach, we scored thousands of compounds from the SuperDRUG2 library, which consists of approved drugs that have already demonstrated safety in clinical trials.⁴⁵ We selected five compounds from this library predicted by our model to test the in vitro inhibition of GSK3β activity as well as another 48 compounds from the MicroSource Spectrum screening compound library. From these, we identified three previously unreported inhibitors of GSK3β that demonstrate IC₅₀ values in the sub-micromolar to low-micromolar range. In order to ascertain selectivity, we used a GSK3α machine learning model to score the ability of these molecules to inhibit GSK3α and then verified the prediction scores with in vitro testing.

Results

Machine Learning Models

We built Bayesian machine learning models with Assay Central software using publicly available data from ChEMBL for GSK3β and GSK3α. The threshold selected by Assay Central for the GSK3β model was 732.8 nM, which divided the training set of 2368 compounds into 1220 active and 1148 inactive drugs, hence representing a well-balanced model with a five-fold cross validation recall, precision, and ROC of 0.807, 0.858, and 0.905, respectively (Figure 1A). This model’s performance was further assessed with 41 known inhibitors as an external test.⁴⁶⁻⁴⁸ The precision in predicting the activity of these compounds was 0.656, while recall, specificity, and ROC were 0.840, 0.313, and 0.635, respectively (Figure 1B).

Figure 1.

GSK3β model statistics and molecular predictions in Assay Central. (A) Five-fold ROC curve for the model. (B) External validation. (C) Visualization of the ranked predictions in the Assay Central interface. The color box to the right of the molecule represents the prediction score from the Bayesian model. Green indicates a higher prediction score, while the black bar under the color box indicates a higher overlap of features as well as the confidence in the activity against the target.

The GSK3α model had a calculated threshold of 1224 nM, which divided the dataset of 430 drugs into 226 active and 204 inactive compounds. The internal five-fold cross-validation resulted in a recall, precision, and ROC of 0.748, 0.924, and 0.885, respectively (Figure S1).

The GSK3β IC₅₀ prediction model was used to score the SuperDRUG2 database for active compounds. Each compound was assigned a “prediction score” from zero to one, with higher values indicating a higher confidence of activity for the compound against GSK3β (Figure 1C and Table 1). The molecules were scored for applicability, which indicates how many fingerprints of the molecule are present in the compounds used to build the model. We selected five of the highest-scoring compounds to purchase based on their prediction scores: ruboxistaurin (prediction score = 0.76, applicability = 0.82), nefopam (prediction score = 0.58, applicability = 0.61), dantrolene (prediction score = 0.65, applicability = 0.67), nifedipine (prediction score = 0.59, applicability = 0.65), and darifenacin (prediction score = 0.60, applicability = 0.67) (Figure 1C). We excluded the top-scoring compounds that were antibiotics or antineoplastic drugs, as we felt these would not be suitable therapeutics for a chronic disease such as AD.

Table 1. Model Fit and Inhibition Activity of Predicted Molecules for GSK3β.

Additional Machine Learning Algorithms

To evaluate the performance of the seven different machine learning models, internal five-fold cross-validation statistics for the different machine learning algorithms were compared (Figure S2 and Table S2). It can be concluded that there is some variability in individual model statistics across all datasets with different algorithms performing well for some metrics but not others. As observed in Table S2, Assay Central and random forest models performed the best for the GSK3α model, while support vector classification performed the best for the GSK3β model.

In Vitro Testing for GSK3β Inhibition

The five commercially available compounds, ruboxistaurin, nefopam, dantrolene, nifedipine, and darifenacin, were selected from the SuperDRUG2 library based on their model scores (Table 1) to test for activity in vitro using the Promega ADP-Glo assay. In this two-step assay, ATP that is converted to ADP during the kinase reaction is converted back into ATP and subsequently utilized in a luciferin/luciferase reaction that can be quantified using a luminometer. Luminescence produced in this reaction directly correlates with GSK3β kinase activity, and the pan-kinase inhibitor staurosporine was used as a positive control, showing 100% inhibition. Of the five compounds tested, ruboxistaurin displayed the highest inhibition (96%) at 100 μM (Figure 2). This is in agreement with the Assay Central predictions, where this compound had one of the highest prediction and applicability scores (Table 1). Nifedipine showed 69% inhibition at the highest concentration but displayed a very sharp drop in inhibition at lower concentrations, indicating that the activity might be due to interference with the luminescence signal rather than true inhibition of the assay (Figure S3). We also used our model to predict compounds from the MicroSource Spectrum screening library, and we selected 48 compounds with prediction scores between 0.59 and 0.51 to test for GSK3β inhibition using this assay (Table S1). Two of these compounds, nithiamide and thiabendazole, showed >70% inhibition at 100 μM (Table 1), and we considered these two compounds active and selected them for further testing.

Figure 2.

In vitro validation of GSK3β inhibitors predicted using Assay Central. Potential inhibitors were selected for in vitro testing for GSK3β inhibition using the Promega ADP-Glo assay. All inhibitors were tested at 100 μM (n = 3), except for the known GSK3β inhibitor staurosporine. Statistical analysis was performed in GraphPad Prism 8.4.3 using one-way ANOVA (Brown–Forsythe and Welch) with Dunnett’s T3 multiple comparison tests. Significance reported as adjusted P values: ns indicates P > 0.05, * indicates P ≤ 0.05, ** indicates P ≤ 0.01, and **** indicates P ≤ 0.0001.

We performed dose–response curves for ruboxistaurin as well as for the two MicroSource compounds nithiamide and thiabendazole. Ruboxistaurin displayed an IC₅₀ of 39.4 nM, while nithiamide and thiabendazole demonstrated IC₅₀ values of 8.5 and 12.1 μM, respectively (Figure 3). In order to confirm these activities, they were tested using the fluorescence-based Z’LYTE kinase assay (ThermoFisher). This assay measures kinase activity through the ability of a phosphorylated substrate to avoid proteolytic cleavage. Inhibition of GSK3β prevents phosphorylation of a FRET peptide, which is then cleaved by the development reagent, and FRET between the acceptor and donor fluorophores is disrupted. The donor fluorophore (coumarin) is then excited, and the ratio of donor emission to acceptor (fluorescein) emission is accepted as the ratio of cleaved to uncleaved peptide. The emission ratio remains low in the presence of active kinase but increases in the presence of small-molecule inhibition. Using this methodology, ruboxistaurin showed an IC₅₀ of 155.1 nM, and nithiamide and thiabendazole had IC₅₀ values of 5.0 and 54.8 μM, respectively (Figure 4). These compounds were additionally tested for inhibition of GSK3α, with ruboxistaurin having an IC₅₀ of 619.5 nM and nithiamide and thiabendazole having IC₅₀ values of 4.4 and 60.8 μM, respectively, suggesting that ruboxistaurin is four-fold more selective for GSK3β, while nithiamide and thiabendazole are non-selective and weaker inhibitors (Table S3 and Figure S4).

Figure 3.

Dose–response curves of three GSK3β inhibitors. A three-fold 10-point serial dilution of each inhibitor was performed (n = 2), beginning at 100 μM, and inhibitory activity against GSK3β was performed using the Promega ADP-Glo assay. Prism was used to generate a non-linear regression log(inhibitor) vs. response equation with four parameters. The Hill slopes are as follows: ruboxistaurin = −0.68, nithiamide = −0.80, and thiabendazole = −0.68.

Figure 4.

Secondary in vitro assay validation of GSK3β using Z’LYTE kinase assay. A three-fold 10-point serial dilution of each inhibitor was performed (n = 2), beginning at 100 μM for ruboxistaurin and 1 mM for nithiamide and thiabendazole. Prism was used to generate a non-linear regression log(inhibitor) vs. response equation with four parameters. The Hill slopes are as follows: ruboxistaurin = 0.98, nithiamide = 1.26, and thiabendazole = 1.10.

Discussion

This study demonstrates the potential of using machine learning models in repurposing molecules as inhibitors for GSK3β, a validated target for AD. Using this model, we selected 53 compounds (Table S1) to test in vitro that led to finding an inhibitor of GSK3β ruboxistaurin, with sub-micromolar affinity. Machine learning and Bayesian methods in particular have been previously demonstrated to be a valuable asset for drug discovery, learning from high throughput screening data and enabling the selection of compounds to prospectively test the models. In many cases, this has led to significant enrichment of active compounds.^35,49,50 In this study, public data was used to build models and predict molecules that can bind to GSK3β using the MicroSource Spectrum screening library and SuperDRUG2 library.

Ruboxistaurin is a protein kinase C (PKC) inhibitor with specificity for the beta isoform that has been tested in clinical trials for indications associated with diabetes, including macular edema, neuropathy, and diabetic retinopathy, and is currently under investigation as a novel therapeutic for heart failure.^48,51 After completing this work, we identified a US patent describing the use of ruboxistaurin as an inhibitor of GSK3β for the treatment of psychiatric and neurological disorders.⁵² However, this activity was not reported in any of the datasets used to build the machine learning model and this molecule was not in the model; therefore, our work demonstrates the power of using machine learning techniques to discover promising active compounds.

Ruboxistaurin was dissimilar to the other active compounds identified with Tanimoto similarities of <0.4 (Table S4). Athough ruboxistaurin was absent from the training set, substructure features were represented in the model based on its applicability score of 0.82. This relatively high score indicates that the molecular fingerprints of ruboxistaurin are found in the compounds used to make the model training set when compared to previously tested inhibitors for GSK3β such as the ATP mimetic staurosporine,⁵³ which has a Tanimoto similarity of 0.6 (Table S5), which still suggests considerable differences. Ruboxistaurin is also dissimilar to other known GSK3β inhibitors such as tideglusib and SB216763 (Tanimoto similarities of 0.45 and 0.57, respectively; Table S5). Interestingly, both nithiamide and thiabendazole have high model applicability scores of 0.77 and 0.94, respectively, although their similarity to staurosporine is <0.4 (Table S5). As the applicability reflects how much is known about a specific molecule with regard to the model, it is indicative of the “trustworthiness” of the prediction score. Perhaps, this may explain why nithiamide (prediction score = 0.52, applicability = 0.77) performed better in the inhibition assay than dantrolene (prediction score = 0.65, applicability = 0.67) despite having a lower prediction score.

GSK3β is a proline-dependent serine–threonine signaling kinase that is involved in several cellular pathways and has been shown to have important roles in embryonic development, glucose regulation, gene transcription, and apoptosis.^22,54−56 However, although important for development and homeostasis in healthy individuals, aberrant expression of this kinase has been observed in the brains of AD patients, and both the α and β isoforms of GSK3 have been shown to have roles in AD pathology.^15,57,58 The GSK3α and GSK3β isoforms share 95% identity in the ATP-binding domain and 97% identity in the catalytic domain, which presents a potential problem in finding an isoform-specific inhibitor.^59,60 However, since both GSK3 isoforms are implicated in the disease process, this may not necessarily be a deterrent for an AD therapeutic. Pan GSK3 inhibitors have been shown to decrease AD pathology in mice, and due to the strong similarity between the α and β isoforms, it is doubtful that these compounds would inhibit one isoform and not the other.⁶⁰⁻⁶² We have also used our Bayesian machine learning model for GSK3α to score the hits identified for GSKβ, and these predictions were further validated in vitro. We showed that ruboxistaurin inhibits GSK3α with sub-micromolar affinity (IC₅₀ = 695.9 nM), though not as potently as it inhibits GSK3β (average IC₅₀ = 97.3 nM). We also showed that nithiamide and thiabendazole also inhibit GSK3α (Figure S4), on a par with the affinities shown for GSK3β. It is worth mentioning that these three compounds were not predicted to inhibit GSK3α and that this discrepancy between prediction and in vitro observed activity may be related to the much smaller training dataset used to build the GSK3α model. Although structural analysis and docking were outside the scope of this machine learning study, we have shown using our Assay Central machine learning models that we can highlight the atoms contributing to activity (green is positive and red is negative) on ruboxistaurin. This may explain how the “top of the molecule” (maleimide) is important for GSK3β activity (colored green), but this is possibly unfavorable for GSK3α activity as it is colored red (Figure S5). These may represent important interactions with the protein.

Given the active role in human diseases, GSK3 has long been an attractive drug target. The first identified inhibitor of GSK3 was lithium, which was a drug approved for psychiatric disorders.⁶³ Lithium ions inhibit GSK3 activity through competition with the cofactor magnesium and through induction of inhibitory phosphorylation of both isoforms; however, several studies looking at lithium as a treatment for AD did not show efficacy for the disease.^64,65 A more recent study has shown promising results for people with early-stage mild cognitive impairment who may be at risk of developing AD, at doses lower than those used for mood disorders.⁶⁶ Additional small-molecule GSK3 inhibitors have been identified or developed that have varying potency and selectivity toward GSK3α or GSKβ. Some of these are ATP-competitive inhibitors, including indirubins, maleimides, and paullones, while some are non-ATP competitive inhibitors, like thiadiazolidinones, and others are peptides.^28,67−72

The severe side effects of the inhibitors identified and tested to date have resulted in very few being investigated in clinical trials, and so far, none have been FDA-approved.⁶⁸ AstraZeneca’s AZ1090 failed due to nephrotoxicity, and the GSK3β-isoform-specific, non-ATP competitor tideglusib was well tolerated but failed due to lack of efficacy.^27,32,73 One of the advantages of repurposing molecules is that the toxicity (or other issues) of the molecules is already well understood. Our goal was to use machine learning to identify pre-existing molecules that we identified as effective against GSK3β and, by their prior usage, should be deemed safe. Ruboxistaurin was shown to be safe up to 32 mg/day for 4 years in a cohort of diabetic individuals.⁷⁴ Considering the nature of AD as a neurodegenerative disease, it is likely that new treatments will have to be tolerated over very long-term usage. Both the sub-micromolar affinity of ruboxistaurin for GSK3β and the clinical safety in humans warrant further examination as a possible treatment in AD, and future experiments will be carried out to examine the ability of this small molecule to attenuate AD pathology in animal models. Potential opportunities to improve upon this drug would be to identify analogs without the potential for drug–drug interactions, as ruboxistaurin has demonstrated potent interactions with CYP2D6 in vitro,⁷⁵ and to remove (or reduce) some of the other kinase inhibition activities, which may be undesirable.⁷⁶ The former improvements should be prioritized, as those with AD are likely to be older and prescribed multiple medications, which would increase the potential for drug–drug interactions. Previously published kinase screening against over 400 kinases showed several low nanomolar activities but interestingly had micromolar activity for GSK3α and GSKβ.^77,78 In contrast, our data suggests that ruboxistaurin is more potent against GSKβ, with activity in line with some of the other kinases, which it potently inhibits like protein kinase C β (PKC β).^79,80

Conclusions

We have demonstrated that Bayesian machine learning techniques using Assay Central can be applied to build a GSK3β model and successfully predict molecules with this activity in silico. This approach increases the chance to find inhibitors for the target of interest and could decrease both the time and cost of lead discovery. This technology can be applied alongside other methods and applied to other drug targets for AD that have sufficient available data with which to build such models, expediting the discovery of molecules for these targets.

Materials and Methods

Chemicals and Reagents

Ruboxistaurin hydrochloride, dantrolene sodium hemiheptahydrate, darifenacin hydrobromide, nefopam hydrochloride, nifedipine, nithiamide, and thiabendazole were purchased from MedChemExpress, and staurosporine was purchased from Sigma-Aldrich.

Machine Learning

The GSK3α and GSK3β datasets were used to generate machine learning models with Assay Central, which is internal software developed from open source descriptors and algorithms that have been previously described^{35−37,39−44,81,88} combined with additional proprietary scripts. Structure–activity datasets were collated in Molecular Notebook (Molecular Materials Informatics, Inc. in Montreal, Canada) and were thoroughly curated to generate a Bayesian machine learning model with multiple scripts. We employed a series of rules to detect problematic data; corrections were implemented by a combination of automated and human re-curation for structure standardization. This approach produces a high-quality dataset and a Bayesian model to predict activities for proposed compounds. These Bayesian models utilize extended-connectivity fingerprints of maximum diameter 6 (ECFP6) descriptors generated from the Chemistry Development Kit library.⁸² These descriptors have widely been noted for their ability to map structure–activity relationships.⁸³ From all training set molecules, the Assay Central software enumerates all fingerprints from the training set and determines a given fingerprint’s contribution to a binary activity classification from the ratio of its presence in active and inactive molecules. Assay Central uses the summation of these contributions for a given molecule to produce a probability-like score. Metrics such as receiver operator characteristic (ROC), recall, precision, F1 score, Cohen’s kappa, and Matthews correlation coefficient were generated from internal five-fold cross-validation of the model. To maximize these internal performance statistics, the software can select a reasonable activity threshold and generate predictions as well as applicability scores for any desired compound. Higher prediction scores are desirable as scores higher than 0.5 are assigned to active compounds (inhibitors). Higher applicability scores are also desirable as they ensure the representation of the drug in the training set.⁸³ Assay Central has been used in various drug discovery projects and the applicability of the model statistics have also been previously described.³⁵⁻⁴⁴

Specifically, the GSK3β model was built with all the inhibitors reported in ChEMBL version 25⁸⁴ for target 262. Assay Central was used to build the model using IC₅₀ values, and compounds such as Zn²⁺ (CHEMBL1201279), Li⁺ Cl^– (CHEMBL69710), and Li⁺ (CHEMBL1234004) were removed to increase the performance of the model. The activity of repeated compounds was averaged by the software, and the final model consisted of 2368 compounds. The GSK3α model was also based on all the inhibitors reported in ChEMBL version 25.⁸⁵ The ChEMBL inhibitors for target 2850 consisted of 430 compounds with IC₅₀ values.

The performance of the GSK3β model was also assessed using 41 molecules that were found to be inhibitors of this target from several other published studies.⁴⁶⁻⁴⁸ The activity threshold for this external dataset was set at 400 nM (in order to provide sufficient actives in the set), and Assay Central was used to generate the model performance metrics.

Additional Machine Learning Algorithms

The extended-connectivity fingerprint (ECPF6) molecular descriptor that Assay Central utilizes was also exploited by multiple machine learning algorithms. The algorithms included random forest, k-nearest neighbors, support vector classification, naïve Bayesian, AdaBoosted decision trees, and deep learning architecture. Further details of these machine learning methods have been reported previously by us and others.^41,42,81

Similarity Analysis

Two-dimensional similarity analysis was performed using the “find similar molecules” protocol in Discovery Studio version 2019 (Biovia San Diego, CA, USA). MACCS keys for the drugs of interest were used as the descriptor and the Tanimoto similarity coefficients calculated.⁸⁶ A similarity of 1 implies maximum similarity (e.g., a molecule compared to itself), whereas 0 indicates maximal dissimilarity.

Kinase Assays

Five compounds were purchased from MedChemExpress and 48 compounds were obtained from The MicroSource Spectrum Collection of MicroSource Discovery Systems, Inc. (Gaylordsville, CT, USA) (Table S1) and were tested for GSK3β inhibition using the GSK3β kinase enzyme system (Promega, V1991) with the ADP-Glo kinase assay (Promega, V6930) at an initial concentration of 100 μM. Compounds that showed inhibition were then tested to generate a dose–response curve and to determine IC₅₀ values. Single-point enzymatic reactions were performed with 1.2 ng/μL GSK3β, 200 ng/μL peptide substrate, 25 μM ATP, and 50 μM DTT with a final DMSO concentration of 1% and incubated for 60 min at 25 °C in a 384-well white-walled microplate (Greiner Bio-One, 784075). ADP-Glo reagent was then added to the reaction and incubated for 40 min at room temperature, and the reaction was stopped by adding the kinase detection reagent and incubating for 30 min at room temperature. Luminescence was read on a SpectraMax M3 (Molecular Devices, San Jose, CA) with an integration time of 1 s per well. The dose–curve reactions were performed the same as above, with the addition of 0.01% Triton X-100, conducted in LUMITRAC 200 plates (Greiner Bio-One, 781075), and read on a SpectraMax iD5 with a read-height of 5.02 mm and an integration time of 1 s per well. All ADP-Glo dose–response curves were performed in duplicate, with a maximum inhibitor concentration of 100 μM with a 10-point, three-fold serial dilution.

Confirmatory assays for GSK3β were performed by ThermoFisher Scientific SelectScreen Biochemical Kinase Profiling Service (Life Technologies Corporation, Bank of America Lockbox Services, 12088 Collections Center Drive, Chicago, IL 60693). Ten-point titration dose–response curves starting at 100 μM for ruboxistaurin and 1 mM for nithiamide and thiabendazole were performed in duplicate using the Z’LYTE kinase assay. The same assay was employed to test the activity of these compounds against GSK3α. IC₅₀ values were then calculated based on percent inhibition data from two separate runs. Statistical analysis was performed using GraphPad Prism 8 software (GraphPad Software, 2365 Northside Dr., Suite 560, San Diego, CA 92108). A non-linear regression log(inhibitor) vs. response equation with four parameters was used, and the Hill slopes for each of the graphs are reported.

Glossary

Abbreviations Used

AD

Alzheimer’s disease

AC

Assay Central

MCC

Matthews correlation coefficient

AUC

area under the receiver operating characteristic curve

CK

Cohen’s kappa

rf

random forest

knn

k-nearest neighbors

svc

support vector classification

bnb

Bernoulli naïve Bayes

ada

AdaBoost decision trees

DL

deep neural networks

GSK3β

glycogen synthase kinase 3 beta

BDNF

brain-derived neurotrophic factor

APP

amyloid precursor protein

pTau

hyperphosphorylated tau

ROC

receiver operator characteristic

PKC

protein kinase C.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.0c03302.

• Supporting further details on the models, structures of public molecules, and computational models (PDF)

The authors declare the following competing financial interest(s): SE is founder and owner of Collaborations Pharmaceuticals, Inc. PAV, ACP, DHF and EM are employees of Collaborations Pharmaceuticals, Inc.