Discovery of Dual Targeting GSK-3β / HIV-1 Reverse Transcriptase Inhibitors as Neuroprotective Antiviral Agents

Abstract

Glycogen Synthase Kinase 3 beta (GSK-3β or GSK-3B) is a serine-threonine kinase involved in various pathways and cellular processes. It has been shown that GSK-3β is associated with several neurological diseases including Alzheimer’s disease (AD), bipolar disorder, as well as rare diseases like Rett syndrome. GSK-3β has also been implicated in HIV-associated dementia (HAD), as it is upregulated in HIV-1-infected cells and plays a role in neuronal dysfunction. Therefore, a small molecule that can both inhibit GSK-3β and HIV-1 reverse transcriptase could offer neuroprotective therapy for patients suffering from HIV-1. Despite this, there are no known GSK-3β inhibitors currently approved, thus prompting us to screen our panel of various antiviral compounds against this kinase to better understand its structure–activity relationship. We show for the first time that the approved drugs etravirine and rilpivirine possess GSK-3β activity (IC₅₀ 619 nM and 502 nM, respectively). We have also identified 3 lead molecules exhibiting IC₅₀ < 1 μM (11726169, 12326205 and 12326207), with compound 11726169 being the most potent in vitro GSK-3β inhibitor (IC₅₀ = 12.1 nM). We also describe the generation of machine learning models for GSK-3β inhibition and their validation with our data as an external test set and propose their use for future optimization of such inhibitors.

Graphical Abstract

INTRODUCTION

Glycogen Synthase Kinase 3 (GSK-3) is a serine-threonine kinase with broad specificity which plays a crucial role in various processes including inflammation, cell growth, cell differentiation, and energy metabolism. It is associated with age-related diseases like diabetes, Alzheimer's disease (AD), and cancer ^{1, 2} as well as rare diseases such Rett syndrome ³ and cyclin-dependent kinase-like 5 (CDKL5) disorder ⁴. GSK-3 exists as two isoforms, GSK-3α and GSK-3β, which share an identity of approximately 98% within their kinase domains and 100% similarity, being able to phosphorylate the same substrates ⁵. Both isoforms are ubiquitously expressed throughout the body, but their functions exhibit some degree of distinction ². GSK-3β is found in higher abundance compared to GSK-3α within the central nervous system (CNS) of rats and levels of GSK-3β were found to increase with age ⁶. GSK-3β is involved in both peripheral and central inflammation, where it is crucial in mediating inflammatory responses in astrocytes and microglia ^7-9. In addition to controlling the innate immune system, GSK-3β is an important regulator of adaptive immunity, which contributes to the overall pro-inflammatory effect ¹⁰.

GSK-3β is a negative regulator of the canonical Wnt signaling pathway, which plays a crucial role in neuronal development and in adult CNS physiology ¹¹. This pathway regulates various processes, including neurite outgrowth in adults, synapse formation, plasticity, and neurogenesis ¹². Additionally, activation of the pathway involving β-catenin ¹³ in microglia and other immune cells has been shown to alleviate neuroinflammation ⁹. GSK-3 antagonizes the canonical Wnt signaling pathway and is therefore a potential therapeutic target in neuroinflammation and neurodegeneration. Pharmacological inhibition of the enzyme activates the Wnt pathway and promotes neurogenesis¹¹, such that GSK-3β inhibitors have potential as therapeutics for CNS diseases including AD and others. While numerous GSK-3 inhibitors described in the literature are under development, there are none that are FDA approved and only a handful have reached clinical stages.¹⁴

GSK-3β has also emerged as an important target for treating HIV infections as it has been demonstrated that GSK-3β is implicated in HIV-1 replication, where it is is upregulated in HIV-1-infected cells ¹⁵. Furthermore, inhibitors of GSK-3β have been shown to reactivate HIV-1 from latency by promoting transcription of HIV-1 proviral DNA ^{16, 17}. Reactivation of HIV-1 allows immune cells to target latent HIV-1 proviruses and could help to eradicate viral reservoirs. Moreover, a common result of chronic HIV infections is HIV-1-associated dementia (HAD) which results in part from metabolic dysfunction of neuronal pathways caused by HIV neurotoxins. The use of GSK-3β inhibitors has been shown to reverse HIV-1-induced damage to synaptic pathways in the CNS ¹⁸, highlighting the need for neuroprotective antiviral agents. These findings also support the concept that a small molecule demonstrating dual GSK-3β / HIV-1 inhibition could serve as a novel HIV antiviral which additionally protects against HAD. This prompted us to screen our existing panel of antiviral compounds (including HIV non nucleoside reverse transcriptase inhibitors, (NNRTI)) for GSK-3β inhibition as part of our ongoing antiviral research program ¹⁹. We have also previously described curating 2368 GSK-3β inhibitors from the literature and building a Bayesian machine learning model that was used for virtual screening which identified ruboxistaurin as a GSK3β inhibitor ²⁰. We have now updated these machine learning models and used them to predict the in vitro GSK-3β inhibition data from this study as an external test set.

RESULTS AND DISCUSSION

Given that GSK-3β is a potentially important target to address in HIV infections, we were initially curious if existing approved HIV NNRTIs were also already targeting GSK-3β. We therefore tested five FDA-approved NNRTIs for their inhibition of GSK-3β and compared the resulting IC₅₀ to their activities against HIV reverse transcriptase (RT) and whole cell activity shown below in Figure 1 and Table 1, demonstrating a range in activity. None of these 5 molecules had previously been described as GSK-3β inhibitors to our knowledge.

Figure 1.

Dose-response curve of FDA approved HIV NNRTIs against GSK-3β.

Table 1.

FDA approved HIV NNRTIs tested against GSK-3β and HIV reverse transcriptase (RT). * previously reported data ¹⁹

Compound	Structure	GSK-3β IC50	HIV RT IC50	*HIV, WT EC50
Etravirine		619 nM	573 nM*	1 nM 21
Rilpivirine		502 nM	160 nM	0.67 nM 21
Nevirapine		1.67 μM	400 nM*	110 nM 21
Doravirine		>20 μM	20.7 nM*	13 nM 22
Efavirenz		>20 μM	30 nM*	2 nM 21

The most potent FDA approved NNRTIs which inhibited GSK-3β were rilpivirine (IC₅₀ = 502 nM) and etravirine (IC₅₀ = 619 nM). Nevirapine was a weaker inhibitor (IC₅₀ = 1.67 μM), while doravirine and efavirenz were not inhibitors of GSK-3β in vitro. Interestingly the 3 compounds displaying GSK-3β inhibition (etravirine, rilpivirine and nevirapine) are also significantly less potent RT inhibitors (Table 1). We also noted significant differences in activity against HIV RT and whole HIV-infected cells, which we have previously described in further detail ²³. These results suggest that GSK-3β inhibition may be involved with the activity of some existing NNRTIs against HIV. Despite this, the tested NNRTIs have only been tested against viruses in the literature and none have been screened against kinases, to our knowledge.

To further investigate more structurally diverse compounds that demonstrate inhibition of GSK-3β activity in HIV as well as other viruses such as yellow fever virus ²⁴ we tested a series of our previously described antiviral pyrazolamines and triazolamines ¹⁹ against GSK-3β (Figure 2, Figure S2). In the process we discovered a potent GSK-3β inhibitor 11726169 with an IC₅₀ of 12.1 nM which we had not previously tested against HIV. We also tested the activity of 11726169 against GSK-3α due to its high sequence identity with GSK-3β in the catalytic domain, and discovered a relatively similar IC₅₀ value of 18.8 nM (Figure S3). Additionally we also identified compounds 12326205 (IC₅₀ = 344 nM), 12326207 (IC₅₀ = 183 nM), 11626004 (IC₅₀ = 539 nM) and 11726157 (IC₅₀ = 815 nM) as promising GSK-3β inhibitors (Figure 2, Table 2) which are comparable to the activity for several of the FDA approved NNRTI (Figure 1 and Table 1). We used the known GSK-3β inhibitor β-laduviglusib as a positive control for the GSK-3β inhibition assay (IC₅₀ = 6.86 nM, Figure 2)

Figure 2.

Dose-response curves of pyrazolamine and triazolamine analogs tested against GSK-3β.

Table 2:

Pyrazolamine and triazolamine analogs with antiviral activity. ND = not determined; YFV = Yellow Fever Virus

Compound	Structure	GSK-3β IC50	HIV RT IC50 (nM)	*HIV, WT IC50 (nM)	YFV, EC50 (CC50)(μM) 24
11626003		5.7 μM	58.64± 48.59	ND	6.1 (24)
11626004		539 nM	65.72 ± 28.61	ND	>36 (36)
11626007		1.84 μM	280*	9.33 ± 2.27	6.1 (28)
11726154		1.30 μM	34.88 ± 1.6	ND	4.1 (28)
11726157		815 nM	224.8 ± 8.4	ND	>19 (19)
11726159		7.4 μM	15.53 ± 7.81	4.68 ± 0.57	5.6 (16)
11726162		Not active	21.85 ± 5.06	ND	8.1 (70)
11726169		12.1 nM	1315 ± 553.45	ND	>40 (40)
12126065		2.68 μM	230*	0.26 ± 0.05	>17 (17)
12326205		344 nM	3673#	ND	ND
12326207		183 nM	21020#	ND	ND

^#For 12326205 and 12326207 the experiment was performed once using technical triplicates.

Structure-Activity Relationship (SAR) of GSK-3β Inhibitors

With respect to the triazolamines, compounds bearing cyano substituents on the aniline fragment were among the least active, while those bearing halogens were the most active against GSK-3β. Similarly, compounds bearing two halogens were the most active and the substitution pattern seems to be critical for GSK-3β inhibition. For example, the 3,4- vs 3,5-dichloro moiety of analogs 11726169 and 11726159 results in a 500-fold difference in activity, and compound 12326205 retains good activity when the 3,5-dichloro moiety is connected to the other amine at the triazole 5-position. The presence of an acrylonitrile substituent on the sulfonylnaphthalene ring (12126065) did not have a significant effect on GSK-3β inhibition activity relative to other sulfonylnapthalene rings. Similarly, alkyl substituents on the aniline moiety did not have a significant effect on activity, given the range of activies for compounds 11726154, 11726162, and 12326207. Lastly, compound 12326207, which lacks the sulfonyl group, was the second most potent inhibitor of GSK-3β and supports the further investigation of acylnaphthalene substituents. Future analogs in this series would include compounds bearing polyhalogenated aniline moieties and electron-deficient sulfonylnaphthalene or acylnaphthalene moieties to further probe the SAR of the triazole scaffold against GSK-3β. In addition to GSK-3β inhibition, in the process of this work we also identified additional potent HIV RT inhibitors 11726159 (IC₅₀ = 15.53 ± 7.81 nM) and 11726162 (IC₅₀ = 21.85 ± 5.06 nM) with significant activity which may warrant further investigation. Several of these molecules had also previously been reported as possessing YFV activity ²⁴ but in these cases we saw no relationship with GSK-3β, and none has been previously reported either.

Machine Learning

As our search for antiviral compounds which inhibit GSK-3β is ongoing, we hope to utilize machine learning models for GSK-3β inhibition to help us more quickly identify potential lead molecules. We have generated regression models for GSK-3β inhibition utilizing IC₅₀ data from the ChEMBL database. The training set contained 2618 unique structures. We have generated multiple models and performed 5-fold cross validation (Table 3), then used the models to predict the above activity of known HIV NNRTIs and our panel of antivirals (Figure S1). Cross-validation statistics were acceptable for support vector regression (SVR) performing the best for nested 5-fold cross validation (R² = 0.66, Table 3).

Table 3.

Regression models for GSK-3β inhibition built using 2618 molecules with ECFP6 (2048 bits) nested 5-fold cross validation statistics shown for the training dataset, (rfr = random forest regression, knnr = k-nearest neighbors’ regression, svr = support vector machine regression, br = Naïve Bayesian regression, adar = AdaBoosted decision trees regression, xgbr = xgboost regression, enr = ElasticNet), mae = mean average error; rmse = root mean square error; mpd = mean Poisson deviance; mgd = mean gamma deviance

	mae	rmse	r2	mpd	mgd
adar	0.85	1.06	0.28	0.18	0.03
svr	0.52	0.72	0.66	0.08	0.01
br	0.64	0.86	0.53	0.12	0.02
enr	0.77	0.98	0.38	0.15	0.03
knnr	0.55	0.76	0.62	0.09	0.02
xgbr	0.69	0.98	0.38	0.16	0.03
rfr	0.56	0.78	0.61	0.10	0.02

Using known HIV NNRTIs and our panel of triazolamines as external validation sets, we found the GSK-3β inhibition svr model to be predictive of both the tested NNRTIs (MAE = 0.56) and our panel of compounds (MAE = 0.92), although the correlations for these datasets are poor (Figure S1). The structural similarity of the compounds used for this test set is shown in a t-SNE plot (Figure 3). t-SNE plots compress molecular descriptors (ECFP6) to a lower dimensional space, allowing for visualization of chemical property space covered by the test set. The tested HIV NNRTIs and triazolamines, while distinct from each other, are both well-represented by the ChEMBL training set and support the validity of our models for predicting GSK-3β inhibition. As we continue to improve upon our machine learning models for GSK-3β inhibition, we plan to utilize these models as a an approach for screening additional antiviral compounds for GSK-3β inhibition.

Figure 3.

t-SNE visualization showing the ChEMBL GSK-3β inhibition training set, the tested HIV-1 NNRTIs, and the tested panel of triazoles.

Molecular Descriptors

We utilized ChemAxon’s molecular descriptor generator to enumerate logP, logD[pH=7], H-bond acceptors, H-bond donors, molecular weight, ring count, polar surface area, number of aromatic rings, and number of rotatable bonds for our panel of antivirals as well as the known HIV-1 NNRTIs (Table 4). Of the tested triazolamines, GSK-3β inhibition activity was generally correlated with logP and logD [pH = 7], while more lipophilic molecules displayed higher activity. This trend was largely consistent with the HIV-1 NNRTIs, in which etravirine and rilpivirine were most active and least lipophilic. The other ChemAxon descriptors did not display a significant trend with respect to GSK-3β inhibition.

Table 4.

ChemAxon molecular descriptors and GSK-3β inhibition for HIV-1 NNRTIs and triazolamines and pyrazolamines. logD [pH=7] = distribution coefficient between n-oxtanol and buffer at pH = 7.

Compound	logP	logD [pH = 7]	Number of H-bond acceptors	Number of H-bond donors	MW (g/mol)	Ring count	Polar surface area (Å2)	Rotatable bonds	Activity (nM)
11626003	3.73	2.95	5	3	407.42	4	110.67	4	5700
11626004	3.9	3.09	5	3	425.41	4	110.67	4	539
11626007	4.09	4.09	6	2	399.85	4	102.9	4	1840
11726154	4.6	4.6	6	2	413.88	4	102.9	4	1300
11726157	4.39	3.97	7	3	450.29	4	123.13	4	815
11726159	4.69	4.69	6	2	434.3	4	102.9	4	7400
11726162	3.85	3.83	7	2	404.45	4	126.69	4	>20000
11726169	4.69	4.69	6	2	434.3	4	102.9	4	15
12126065	4.33	4.3	8	2	475.91	4	150.48	4	2680
12326205	3.26	3.26	7	2	426.27	3	119.97	4	344
12326207	5.41	5.41	5	2	357.417	4	85.83	3	183
Doravirine	2.23	2.23	5	1	425.75	3	98.03	3	>20000
Efavirenz	4.46	4.46	2	1	315.68	3	38.33	1	>20000
Nevirapine	2.49	2.49	4	1	266.304	4	58.12	2	1670
Etravirine	5.54	5.54	6	2	435.285	3	120.64	3	619
Rilpivirine	5.47	5.47	6	2	366.428	3	97.42	3	502

In vitro ADME profiling

Since 11726169 was the most active compound identified against GSK-3β in this study, we performed in vitro absorption, distribution, metabolism and excretion (ADME) profiling to evaluate kinetic solubility ²⁵, metabolic stability ²⁶, Caco-2 permeability ²⁷, and plasma protein binding ²⁸ (Table 4). These in vitro ADME characteristics suggest that 11726169 is predicted to be metabolically stable, but has low solubility and Caco-2 permeability which would predict low oral bioavailability.

CONCLUSIONS

Several previous studies have linked HIV-1 neurotoxicity to abnormal activation of GSK-3β ^{18, 29-31}. Inhibition of GSK-3β with lithium and sodium valproate (VPA) was shown to restore long term potentiation in the hippocampal side of HIV-1 encaphalitis mice ³⁰. Moreover, VPA treatment of rat cortical neurons exposed to HIV-1 gp120 induced neurite outgrowth, microtubule-associated protein 2 (MAP-2) and neuron-specific nuclear protein (NeuN) antigens ²⁹. Importantly, HIV-encoded Tat activates GSK-3β to antagonize nuclear factor-kappaB survival pathway in neurons ³¹. In light of this evidence, the need for neuroprotective antiviral agents is important. Hence, we screened our existing panel of antiviral compounds (including HIV NNRTI inhibitors) for GSK-3β inhibition to identify small molecules with dual GSK-3β / HIV-1 inhibition which could protect against HAD. We first showed that inhibition of GSK-3β may be associated with several existing FDA approved HIV-1 NNRTIs, and our data also points towards the pyrazolamine scaffold for further exploration of antiviral GSK-3β inhibitors. We further identified 5 low nanomolar inhibitors (11626004, 11726157, 12326205, 12326207 and 11726169), with 11726169 being the most active against GSK-3β (IC₅₀ = 12.1 nM) and GSK-3α (IC₅₀ = 18.8 nM) indicating that the compound likely mainly targets the catalytic domain of GSK-3 which is highly conserved between the 2 isoforms and shares several common substrates ^{2, 32}. In vitro ADME analysis for 11726169 predicts it to be metabolically stable, but with low solubility and Caco-2 permeability predicting low oral bioavailability. Nonetheless, several known FDA approved NNTRI’s are known to be notoriously insoluble in aqueous solvents ^{33, 34}, but still have reasonable bioavailablity.

Although we observed compounds like 11626004 and 11726157 that showed a similar extent of HIV NNRTI and GSK-3β inhibition in vitro, we also saw an inverse correlation between the extent of HIV NNRTI and GSK-3β inhibition in higly potent compounds in general. Utilizing ChemAxon chemical descriptors, we found that GSK-3β inhibition was positively associated with logP and logD [pH = 7], suggesting more lipophilic molecules are more liklely potent inhibitors. Lastly, we have generated new machine learning models for the inhibition of GSK-3β curated from ChEMBL and used our in vitro data as an external test set for validation. Our results suggest the use of these models may be used as a tool to assist in future compound design and identify the next-generation of dual targeting HIV-1 NNRTI /GSK-3β inhibitors as potentially neuroprotective antiviral therapeutics for patients with HAD. Our study also warrants further mechanistic investigation into the newly identified HIV-1 NNRTI /GSK-3β inhibitors in future. As the GSK-3β inhibition by the existing FDA approved NNRTI described herein had not been previously reported, this off target effect may (or may not) be desirable. These compounds could be repurposed as GSK-3β inhibitors while their activity against other kinases will also need to be more broadly profiled. Certainly one could also evaluate the GSK-3β inhibition of these identified compounds in cell based models as well as determine the potential for neuroprotection in disease / animal models. In summary, this work has identified new scaffolds as GSK-3β inhibitors with potential antiviral and CNS applications.

METHODS

GSK-3β enzyme activity assay

GSK-3β and GSK-3α assay was performed using Z’LYTE kinase assay using ThermoFisher Scientific SelectScreen Biochemical Kinase Profiling Service. The compounds were screened in 1% DMSO (final) in the well. For 10 point titrations, 3-fold serial dilutions are conducted starting at 20 μM. The 2X GSK-3β / Ser/Thr 09 mixture is prepared in 50 mM HEPES pH 7.5, 0.01% BRIJ-35, 10 mM MgCl₂, 1 mM EGTA. The final 10 μL Kinase Reaction consists of 0.12 - 0.8 ng GSK-3β and 2 μM Ser/Thr 09 in 50 mM HEPES pH 7.5, 0.01% BRIJ-35, 10 mM MgCl₂, 1 mM EGTA. After the 1 h Kinase Reaction incubation, 5 μL of a 1:512 dilution of Development Reagent A is added.

NNRTI assay methods.

The fluorescent method was done using an EnzChek^® reverse transcriptase assay kit purchased from Molecular Probes (Eugene, OR) with a modified manufacture’s protocol. In short, compounds were serially diluted 10-fold starting with 10 mM stocks in DMSO. DMSO concentrations were kept the same for all dilutions used within an assay (2%). The final compound concentrations ranged from 10,000-20,000 nM to 0.1-0.2 nM respectively with six, 10-fold serial dilutions. 96-well plates were set up with 20 μL reaction mixture (poly(A) ribonucleotide template/ oligo d(T)16 primers), 4 μL diluted HIV RT enzyme purchased from Millipore Sigma at 0.08u/μl and 0.5 μL of compound or 2% DMSO control. The reaction was incubated at room temperature for 45 min followed by the addition of 125μL of PicoGreen dsDNA quantitation reagent. This was incubated for 5 minutes then RTase activity was quantified based on the formation of dsDNA (excitation 480 nm, emission 520 nm) using a Spectramax ID5 plate reader from Molecular Devices (San Jose, CA). Etravirine was used as a positive inhibitor control, while no enzyme control was used as a negative control for the reverse transcriptase assay. A minimum of 2 biological replicates are performed for each compound.

ADME testing

Caco-2 cell permeability testing:

For the Caco-2 cell permeability assay the Caco-2 cell monolayer was plated in a 96 well format (0.4 μm pore-sizes, 0.143 cm² surface area). The Caco-2 cell line was procured from Evotec and cells were cultured in 5% CO₂ in a humidified cell culture chamber at 37 °C. The monolayers in the apical chamber were pre-incubated with pre-warmed HBSS (Hank’s balanced salt solution) containing 10mM HEPES buffer (pH 7.4) and cells in the basolateral chamber received pre-warmed HBSS containing 10mM HEPES buffer (pH 7.4) and 1% BSA. Digoxin, Atenolol and Propranolol at a final concentration of 10μM were used as assay controls. A final DMSO concentration of 0.1% was used for the assay. Compounds were tested at a final concentration of 2μM and 10μM in the assay.

On the day of assay, Caco-2 cell monolayers were washed with transport buffer (HBSS, pH 7.4) and preincubated for 30 min (37 °C, 5% CO2, 95% RH). Transport experiments were initiated by addition of HBSS buffer (pH 7.4) solution spiked with test compounds and permeability markers (10 μM) to the donor compartment in duplicate (n=2) wells and drug free HBSS buffer with 1% BSA (pH 7.4) to the receiver compartment. The volumes of apical and basolateral compartments are 0.075 and 0.25 mL, respectively. Plates are incubated in incubator for 120 min at 37 °C. Samples were collected both from acceptor and receiver chambers at 120 min post assay initiation. Transport experiments were conducted in both apical to basolateral (A-B) and in the basolateral to apical (B-A) directions. Concentration of test compounds in the samples are analyzed by liquid chromatography tandem mass spectrometry (LC-MS/MS) method using discovery grade bioanalytical method.

The integrity of Caco-2 cell monolayers were examined post experiment. Caco-2 cell monolayers were washed with pre-warmed HBSS buffer. Apical insert was dosed with 75 μL of 50 μM Lucifer Yellow (LY) and basolateral wells with 250 μL of HBSS buffer. Plates were incubated in incubator for 60 min at 37 °C. After 60 min, samples (200 μL) were transferred from each of the basolateral receiver plate into 96 well black plates. The samples were read on spectrophotometer at an excitation wavelength of 428 nm and emission wavelength at 540 nm.

Apparent Permeability (${\mathsf{P}}_{\text{app}}$) was calculated using the following formula: Papp is expressed in units of 10⁻⁶ cm/sec.

$$\text{Papp}=\frac{\text{dq}}{\text{dt}}\mathrm{X}\frac{1}{\text{Co}}\mathrm{X}\frac{1}{\mathrm{A}}$$

Where $\mathsf{dq}/\mathsf{dt}=\mathsf{rate}\mathsf{of}\mathsf{transport}$ (rate of transport of compound in the receiver compartment), $\mathsf{Co}=\mathsf{initial}\mathsf{concentration}\mathsf{in}\mathsf{the}\mathsf{donor}\mathsf{compartment}$, $\mathsf{A}=\mathsf{surface}\mathsf{area}\mathsf{of}\mathsf{the}\mathsf{effective}\mathsf{filter}\mathsf{membrane}$.

Efflux Ratio was calculated using following equation:

$$\text{Efflux ratio}=\frac{\text{Papp values in B}-\text{A direction}}{\text{Papp values in A}-\text{B direction}}$$

% Recovery was calculated using following equation:

$$\%\text{Recovery}=\frac{\text{Total compound in donor and receiver chamber at the end of transport experiment}}{\text{The initial amount of compound loaded in the donor chamber}}\mathrm{x}100$$

Acceptance Criteria:

Permeability Markers	Papp x 10−6 cm/sec (A-B)	Papp x 10−6 cm/sec (B-A)
Propranolol	8.2-29.3	6.6-27.02
Atenolol	0.04-0.96	0.06-1.24
Digoxin	0.25-1.08	8.7-30.42

Kinetic solubility testing:

4 μL of master stock (10 mM) was added to 396 μL of 100 mM potassium phosphate buffer (pH 7.4) and incubated for 2 h at room temperature on a thermomixer at a shaking speed of 1100 rpm. Final DMSO concentration in the incubations was 1% v/v. After incubation, samples were filtered at 4000 rpm for 5 min in filter plate and supernatant was separated. Collected supernatant was subjected to HPLC analysis. Test samples were analysed against 8 point calibration curve with concentrations ranging from 1 μM to 150 μM. Carvedilol and diethylstilbesterol were used as assay controls with acceptance criteria being Carvedilol solubility > 55 μM and diethylstilbesterol solubility being ≤ 10 μM.

Mouse and human liver microsome stability:

Mouse and human liver microsomes (MLMs and HLMs respectively) were purchased from Xenotech, (MLMs- catalog no: M5000, lot no: 2210070), (HLMs- catalog no: H2620, lot no: 1910096). Liver microsome (LMs) at a final concentration of 0.56 mg/ml were prepared mixing 5ul of LM stock solution 20mg/ml to 173 μL of 100mM potassium phosphate buffer , pH 7.4 to prepapre the microsomal mix. 2 μL of test compound working stock (100 μM) solution was added to 178 μL of buffer microsomal mix and preincubated for 15 min at 37 ± 1 °C (600 RPM) into the wells of a 96 well plate. After preincubation, 45 μL (0 min sample) of preincubation mixture was precipitated with 400 μL of ice cold quenching solution containing internal standard (50 ng/mL of tolbutamide and 50 ng/mL of Telmisartan) and 5 μL of 10 mM NADPH was added. 45 μL of preincubation mixture was mixed with 5 μL of buffer and incubated for 30 min (-NADPH sample). The reaction was initiated by addition of 10 μL of cofactor (10 mM) to the preincubation mixture and incubated on a thermomixer maintaining 37 ± 1 °C (600 RPM). 50 μL of incubation mixture was taken at 30 min time point and stopped by addition of 400 μL of ice-cold quenching solution containing internal standard (50 ng/mL of tolbutamide and 50 ng/mL of telmisartan). The samples were vortexed for 10 min at 1000 RPM and centrifuged at 4000 RPM for 30 min at ~6 °C. The supernatants were collected into 96 well plate as per the cassette combination and samples were analyzed by LC-MS/MS. For -NADPH samples, the reaction was stopped at 30 min with addition of 400 μL of ice-cold quenching solution containing internal standard (50 ng/mL of tolbutamide and 50 ng/mL of telmisartan) and processed in the above manner. Verapamil and diclofenac were used as positive controls for the assay. The percentage of parent compound remaining (%PCR) at a specific time point was calculated using LCMS as -

$$\%\text{PCR}=\frac{\text{Area ratio at X min}}{\text{Area ratio at 0 min}}\mathrm{x}100$$

where X is the final time point.

Human and mouse plasma stability assay:

Intermediate stock solution (100 μM) of test compound (4 μL) was spiked to 396 μL of plasma and incubated for 5 h at 37°C. At each time point (0, 15, 30, 60, 120 and 300 min), 50 μL of incubation mixture was precipitated with 300 μL of ice cold acetonitrile containing internal standard (250 ng/mL). Samples were vortexed for 10 min at 1000 rpm and centrifuged at 4000 rpm for 10 min. After centrifugation, 100 μL of supernatant was diluted with 100 μL of water and submitted for LC-MS/MS analysis. Propantheline was used as a positive control for the assay. The percentage of parent compound remaining (%PCR) at a specific time point and half-life was calculated using LCMS.

Human/mouse plasma protein binding assay:

Human/mouse plasma protein binding was determined using the Rapid Equilibrium Dialysis method (RED) ( 200 μL Plasma/350 μL Buffer ). Human plasma was thawed to 37°C and centrifuged at 4000 rpm for 10 min. Clear supernatant was separated and used for the assay. 1 μL of working stock (2 mM) was spiked to 999 μL of preincubated plasma, to give a final concentration of 2 μM (DMSO ~0.1%). 200 μL of compound spiked plasma was added to the RED chamber and 350 μL of PBS buffer was added to the buffer chamber. Teflcon block consisting of RED inserts was incubated for 5 h on a thermomixer at a shaking speed of 450 rpm. T0 control plasma samples were matrix matched with equal volume (1:1, v/v) of PBS buffer and precipitated with 300 μL of acetonitrile containing internal standard (telmisartan). At 5 h, 50 μL of sample was aliquoted from both red and buffer chamber. These samples were matrix matched and precipitated with 300 μL of acetonitrile containing internal standard. Samples were vortexed at 1050 rpm for 3 minutes and centrifuged at 4000 rpm for 14 min. After centrifugation, 100 μL of supernatant was diluted with 100 μL of Milli Q water and submitted for LC-MS/MS analysis.

Machine learning

Machine learning models were generated using our proprietary software, Assay Central, which we have described previously ³⁵. The machine learning algorithms used to generate models included Bernoulli naïve Bayes, Linear Logistic Regression, AdaBoost Decision Tree, Random Forest, Support Vector Machine, Deep Neural Networks, and XGBoost. 5-fold cross-validation was performed to validate machine learning models. Nested 5-fold cross-validation selects and removes a random, stratified 20% hold-out set prior to model building. The model is then built with the remaining 80% of the training data, and the hyperparameters are optimized using a grid search using 5-fold data set splits. This optimized model is then used to predict the 20% hold-out set and this process is repeated until all molecules have been present in a hold-out set. Finally, the metrics from each hold-out set are averaged to give a final, nested 5-fold cross-validation score. Deep learning (DL) uses a 20% leave-out set instead, due to its high computational requirement. Models were built using ECFP6 descriptors and metrics were generated as described ³⁵.

Molecular descriptor analysis and t-SNE visualization.

We generated molecular descriptors (molecular weight, logP, molecular fraction polar surface area, logD (pH 7.4), the number of aromatic rings, hydrogen bond acceptor and donor, rings and rotatable bonds) with ChemAxon software (Budapest, Hungary).

t-SNE ³⁶ embeds data into a lower-dimensional space. 1024 ECFP6 fingerprints were generated for all compounds. The ECFP6 fingerprints were then embedded into a 2-dimensional vector using t-SNE. All t-SNE values were generated using the scikit-learn library in python with default hyperparameters (n_components = 2, perplexity = 30, early exaggeration = 12.0, learning rate = 200, n_iter = 1000).

Table 5.

ADME properties for 11726169.

11726169
ADME property	Results
Solubility	0 μM at pH 7.4
Mouse liver microsomes	t1/2 = 26.07 min, CLint = 53.15 μL/min/mg protein
Human liver microsomes	t1/2 = 65.59 min, CLint = 54.85 μL/min/mg protein
Mouse plasma protein binding	% bound=100, % stability = 85.11
Human plasma protein binding	% bound= 100, % stability = 82.92
In vitro stability in human plasma	Half-life (min) = 780.64 (stable)
In vitro stability in CD-1 mouse plasma	Half-life (min) = 4393.75 (stable)
Caco-2	Papp A-B =0.00 ; B-A = 0.03 (x10−6 cm/s) Efflux ratio = N/A

ABBREVIATIONS

AD

Alzheimer’s disease

GSK-3β, CDKL5

cyclin-dependent kinase-like 5

CNS

central nervous system

Glycogen

Synthase Kinase 3 beta

HAD

HIV-associated dementia

NNRTI

non-nucleoside reverse transcriptase inhibitors