Synthesis, Biological Evaluation and Machine Learning Prediction Model for Fluorinated Cinchona Alkaloid-Based Derivatives as Cholinesterase Inhibitors

A series of 46 Cinchona alkaloid derivatives that differ in positions of fluorine atom(s) in the molecule were synthesized and tested as human acetylcholinesterase (AChE) and butyrylcholinesterase (BChE) inhibitors. All tested compounds reversibly inhibited AChE and BChE in the nanomolar to micromolar range; for AChE, the determined enzyme-inhibitor dissociation constants (Ki) ranged from 3.9–80 µM, and 0.075–19 µM for BChE. The most potent AChE inhibitor was N-(para-fluorobenzyl)cinchoninium bromide, while N-(meta-fluorobenzyl)cinchonidinium bromide was the most potent BChE inhibitor with Ki constant in the nanomolar range. Generally, compounds were non-selective or BChE selective cholinesterase inhibitors, where N-(meta-fluorobenzyl)cinchonidinium bromide was the most selective showing 533 times higher preference for BChE. In silico study revealed that twenty-six compounds should be able to cross the blood-brain barrier by passive transport. An extensive machine learning procedure was utilized for the creation of multivariate linear regression models of AChE and BChE inhibition. The best possible models with predicted R2 (CD-derivatives) of 0.9932 and R2(CN-derivatives) of 0.9879 were calculated and cross-validated. From these data, a smart guided search for new potential leads can be performed. These results pointed out that quaternary Cinchona alkaloids are the promising structural base for further development as selective BChE inhibitors which can be used in the central nervous system.


Introduction
The neurodegenerative disease generally implies a range of conditions that primarily affect the neurons in the human brain and cause problems with movement and/or mental functioning. The most common disease characterized by progressive non-reversible mental deterioration is Alzheimer's disease (AD), which is caused mainly by a loss of cholinergic innervation in the cerebral cortex and characterized by decreased levels of the neurotransmitter acetylcholine in neurons [1,2]. This has led to the development of AD drugs that inhibit the activity of acetylcholine hydrolyzing enzymes, acetylcholinesterase (AChE), and butyrylcholinesterase (BChE). Today, four out of six currently in use drugs for the treatment of AD are AChE or non-selective cholinesterase inhibitors: donepezil, rivastigmine, galantamine, and pyridostigmine. All four drugs exhibit their pharmacological effect by increasing the amount of ACh in the brain consequently leading to temporally improved cognitive abilities of patients and alleviation of the AD symptoms. The problem in the development of AD drugs that can slow or stop the progression of disease lies in AD's highly complex etiology associated with diverse clinical hallmarks which do not occur at the same time and whose intensities of occurrence vary which makes it difficult to establish Because of their unique structure with a few available places for derivatization, Cinchona alkaloids and their derivatives are one of the most versatile groups of natural compounds [10]. Major applications of Cinchona alkaloid derivatives are in numerous organic reactions as chiral organocatalysts and chromatography for the separation of various racemic analytes as a chiral stationary phase Besides diverse chemical applications, they are also bioactive compounds. The most known bioactive alkaloid is quinine, which was used as a powerful antimalarial drug for centuries [11]. Derivatives bearing Cinchona scaffold were identified as autophagy inhibitors, potential antibiotics, anticancer or insecticidal agents [12][13][14][15].
In our previous study, we have identified several potent BChE inhibitors based on quaternary CD and CN scaffolds with high selectivity toward BChE [16]. Those results have encouraged us to continue research on Cinchona based alkaloids as potential cholinesterase inhibitors with selectivity toward BChE. We synthesized additional 46 CN and CD derivatives that differ in the structure of the quinuclidine quaternary nitrogen atom. In the benzyl moiety fluorine atom(s) were introduced to explore the influence on the compound's bioactivity. Fluorine atoms are highly electronegative and relatively small in size, and they can be used in medicinal chemistry as bioisosteres of functional groups (e.g., hydrogen atom and methyl group), and serve as functional mimetics of the carbonyl group. Trifluoromethyl group can be bioisostere of iso-propyl and tert-butyl groups, replacing also the iodine or bromine atom. Furthermore, fluorine atoms and groups are important because of the ability to tune up more than one molecular property such as change of the conformation constraints, change of bioactivity of the molecule, membrane permeability (lipophilic properties), metabolic pathways, etc. Thus, derivatization of bioactive compounds with fluorine atoms is a well-known strategy in drug discovery [17,18]. Also, we have prepared a small series of corresponding 10,11-dihydroderivatives by reduction of the vinyl group in CD and CN to investigate whether this structural modification has influence on the inhibitory activity toward AChE and BChE. The in vitro-determined affinity of the studied compounds toward cholinesterases, their inhibition selectivity as well as the stereoselectivity of the enzymes, were analysed and defined by molecular modelling and extensive machine learning procedures. In addition, the ability of compounds to cross the blood-brain barrier by passive transport was evaluated as one of the desired features that CNS active compounds should have.

Synthesis
Since our pilot study on CD and CN derivatives of Cinchona alkaloids gave promising results regarding the inhibition potency of those compounds towards human cholinesterases [18], we have synthesized a series of 46 quaternary derivatives of Cinchona derivatives of which 18 were not previously described and/or mentioned in the chemical literature as stated in the experimental section.

Inhibition of Cholinesterases
In this paper we broadened our previous research on the binding affinity of CD's and CN's derivatives toward BChE and AChE [18] and all synthesized compounds were tested as human BChE and AChE inhibitors. Since the effect of fluorine carrying groups on the binding affinity of the parent compound is difficult to predict, we synthesized a series of fluorine-containing substituents on benzyl group of CD's and CN's to enable us to carry out structure-activity relation studies concerning the inhibition of human cholinesterases. In line with that, we synthesized a series of fluorine and fluorine-containing CD's and CN's where the position of fluorine atom(s) and fluorine-containing groups was changed systematically (Figure 2). Moreover, according to our previous study [18], bioactive conformers of the Cinchona derivatives had different positioning of the vinyl group in the AChE and BChE active sites and are positioned to allow favorable orientation of other moieties in the molecule. Structures of prepared compounds: CD Bzl (R=H); CD 2F; CD 3F; CD 4F; CD 3CF 3 ; CD 4CF 3 ; CD 3,5F; CD 3,4F; CD 2,3F; CD 2,4F; CD 2,6F; CD 3OCF 3 ; CD 4OCF 3 ; CD 2F-6CF 3 ; CD 4Br-2F; CD 2Cl-6F; DHCD; DHCD Bzl (R=H); DHCD 3F; DHCD 4F; DHCD 3CF 3 ; DHCD 4CF 3  "*" denotes two chiral centers at positions 8 and 9 in pseudo-enantiomers.
All synthesized compounds reversibly inhibited BChE and AchE, and as a measure of their inhibition potency the dissociation constants (± standard errors) of the enzymeinhibitor complex (K i ) were determined. The impact of the introduction of fluorine atoms in the structure of inhibitors and the impact of reduction of the vinyl group on inhibition potency were analyzed.

Inhibition by Cinchonine and Cinchonidine Derivatives
The impact of changes in substituents on the benzene ring on inhibition potency can be analyzed separately for CD's and CN's. The K i constants (± standard errors) for BChE and AChE are presented in Table 1. The activity of BChE was inhibited by all tested compounds with K i constants ranging from 0.075-9.9 µM. The most potent inhibitors were CD 3F and CD 3,5F, both with K i constants around 76 nM which classified them as highly potent BChE inhibitors [27] and suggested the possible importance of meta-positioned fluorine atom (position 3) on inhibition potency. This is additionally supported by the fact that the ortho (CD 2F) and para (CD 4F) monofluorine substituted compounds didn't reach such high potency of inhibition, being up to 19 times less potent inhibitors. Substitution of meta-fluorine with other electronwithdrawing groups-trifluoromethyl group in CD 3CF 3 or trifluoromethoxy group (OCF 3 ) in CD 3OCF 3 led to the decrease of inhibition potency in order CD 3F > CD 3OCF 3 > CD 3CF 3 implying that inhibition potency of compounds with a substituent in meta position on benzene ring depend also on its volume due to the steric constraints of enzyme's active site. In the case of CF 3 or OCF 3 as substituents, inhibition potency was similar whether they were in meta or para position. Five compounds with two substituents were synthesized as well, and from the values of their K i constants it seems that for achieving optimal interactions with BChE active site, meta positions are preferred. The most potent disubstituted inhibitor is the one with fluorine atoms on positions 3 and 5. The introduction of other halogen or CF 3 beside ortho positioned fluorine atom, revealed that para position for additional halogen is preferred, e.g., CD 2F-4Br. Inhibition potency of CD 3F and CD 3,5F is comparable with one determined for CD's with unsubstituted benzyl moiety, and benzyl moiety with meta-positioned methyl group as determined previously [18]. Moreover, if the inhibition potential of compounds studied in this work was compared to CD's with other meta-positioned substituents on benzyl moiety, it can be concluded that for achieving optimal position and binding interactions, halogens larger than fluorine are not desirable.
Contrary to CD's, different positions of fluorine atom alone or in combination with additional fluorine atoms, fluorine carrying groups or other halogens have minor or no effect on the inhibition potency of CN compounds. Inhibition constants of CN compounds were in the range of 1.2-9.6 µM where the most potent was the compound CN 2Cl-6F followed by CN 2F and CN 4F (Table 1). Disubstituted orthoand para-positioned fluorine atoms on CN derivatives or a combination of fluorine atoms with another halogen on position 6, have slightly better inhibition potency than corresponding CD derivatives. Generally, the affinity of BChE (1/K i ) toward fluorinated CN derivatives is in the range of that determined for non-fluorinated CN's [10].
All compounds inhibited AChE activity with K i constants ranging from 3.9-69 µM ( Table 1) with affinities about 1000 times lower than that of donepezil, a selective AChE inhibitor in use for the treatment of AD [28,29]. The most potent inhibitor was CN 4F with fluorine atom in para-position on the benzene ring. Inhibition potency of monosubstituted CD's was generally 2.5-8 times lower compared to inhibition potency of disubstituted compounds and compounds with the trifluoromethoxy group. The affinity of fluorine derivatives of CN has changed very little, and no pattern can be deduced related to the determined differences in inhibition potency of CN's and AChE. When inhibition potencies of compounds with meta (CN 3F) or para (CN 4F) positioned fluorine atom were compared to the inhibition potencies of CN's with bromine or chlorine atoms in para positions, it can be concluded that substituents with the bulkier and less electronegative halogens are more favorable for binding into AChE active site [18].

Inhibition by 10, 11-Dihydrocinchonine and 10, 11-Dihydrocinchonidine and Their Derivatives
The reduction of the 3-vinyl group to the 3-ethyl group affected the inhibition potency of compounds (Table 2).
For BChE, inhibition potencies of CD's with the reduced 3-vinyl group were not changed or lowered compared to their non-reduced analogues (Tables 1 and 2). In the case of DHCD, inhibition potency was similar to that of CD [18] while the inhibition potency of quaternary DHCD Bzl was 5 times lower than CD (Tables 1 and 2). For CN's with the 3-ethyl group, the effect was the opposite-reduction of the vinyl group increased the inhibition potency 4 or 3 times in the case of DHCN and DHCN Bzl, respectively, compared to their 3-vinyl analogues. The most potent CD-based inhibitors of BChE with the 3-ethyl group were: DHCD Bzl (a compound with no substituents on a benzene ring), DHCD 3F (a compound with the fluorine atom as meta-substituent), DHCD 3CF 3 (a compound with meta-substituted trimethyl group). DHCD 3CF 3 is a 5 times weaker inhibitor compared with DHCD and DHCD 3F. Introduction of other substituents on benzene ring decreased the inhibition compared to DHCD Bzl and DHCD 3F, which can be related to the increase of their lipophilicity. Inhibition potency of CN's with the reduced 3-vinyl group was in the range of 0.9-2.2 µM, which indicates that introduction of the benzyl group with or without fluorine atoms and fluorine-containing groups did not contribute to the formation of new favorable binding interactions within the BChE active site.
All compounds with the reduced 3-vinyl group inhibited AChE with K i constants in the 4.8-206 µM range. Interestingly, the lowest inhibition potency is determined for non-quaternary DHCD, while the highest potency was determined for quaternary nonsubstituted DHCD Bzl. Inhibition potencies of other compounds were in the 15-68 µM range. It seems that for binding of 3-ethyl derivatives in the AChE active site, compounds having CF 3 or OCF 3 substituents in the ortho-position on benzene ring are preferred. The inhibition potency of 3-ethyl derivatives was generally higher compared to their 3-vinyl analogues in AChE active site [18].

Selectivity of Inhibition
The inhibition selectivity denoted as a selective index (SI) of compounds was calculated as the ratio of K i constants determined for inhibition of AChE and BChE, Tables 1 and 2. Generally, all synthesized compounds had a higher preference for BChE over AChE, and the highest selectivity was determined for CD 3F which displayed 533 times higher inhibition selectivity toward BChE over AChE. CD 3,5F followed with 123-times and DHCD 3F with 84-times higher inhibition selectivity towards BChE. As expected, the highest range of selectivity (SI = 0.77-533) was determined for CD's while CN's are generally less selective (SI = 2.1-33). In the case of 3-ethyl compounds, DHCN's are generally more selective towards BChE than DHCD's. The selectivity of here tested fluorinated CD's and CN's corresponds to that earlier determined for similar compounds [18] where the selectivity of N-para-bromobenzyl CD derivative was 510 times more selective to BChE than AChE.

Stereoselectivity of Inhibition by Pseudo-Enantiomers
The ratio of K i constants determined for CN's and their corresponding pseudo-enantiomers CD's (K i (CN's/CD's)) determined the stereoselectivity of inhibition for each cholinesterase, Table 3. When the ratio is > 1 it denotes that the enzyme has a higher affinity toward CD's over CN's. Generally, neither of the cholinesterases showed pronounced stereoselectivity toward one of the pseudo-enantiomers. For fluorinated pseudo-enantiomeric pairs, BChE displayed 1.5-81 times higher affinity toward CD in ten enantiomeric pairs, Table 3, and the highest stereoselectivity was determined for pseudo-enantiomeric pair CN 3F /CD 3F where BChE had 81 times higher affinity toward CD pseudo-enantiomer. BChE had a 78 higher preference for meta disubstituted CD 3,5-F over CN 3,5-F. Those results pointed out the importance of meta-position of substituents in obtaining selectivity of BChE toward pseudo-enantiomeric pairs of tested compounds. It was observed that increasing the size of the meta substituent and changing its electronegativity, resulted in a selectivity decrease: 3F ∼ = 3,5F > 3CF 3 > 3OCF 3 . For the last pair (CN 3 OCF 3 /CD 3 OCF 3 ) the reversal of stereoselectivity was determined.
AChE displayed 1.7-5 times higher affinity toward CD in 12 pseudo-enantiomeric pairs, where the highest stereoselectivity was determined for the trifluoromethoxy group in meta-position on the benzene ring (3OCF 3 ). For pseudo-enantiomeric pairs with 3F and 4F substituents, AChE displayed 3-and 18-times higher affinity for CN pseudo-enantiomers. Selectivity of AChE toward pseudo-enantiomeric pairs of meta-positioned substituents changed from being three times more selective towards CN 3F pseudo-enantiomer and 5 times more selective towards CD pseudo-enantiomers in order: 3F > 3CF 3 > 3OCH 3 . The meta substitution proved to be important for cholinesterases stereoselectivity. These results correspond to that determined earlier for pseudo-enantiomers of analogues with different substituents on benzene [18].
In the case of pseudo-enantiomers with the reduced 3-vinyl group, no pronounced preference for any of pseudo-enantiomers was observed, Table 4. BChE generally slightly preferred binding of CN derivatives of six (6/8) pseudo-enantiomeric pairs. 16 times higher affinity was found for unsubstituted DHCN over DHCD. On the contrary, AChE was up to 4.4 times more selective in the binding of DHCD pseudo-enantiomers (6/8). It is worth noting that both cholinesterases showed the same enantiomeric preference in the case of DHCN/ DHCD and DHCN Bzl/ DHCD Bzl, preferring DHCN derivatives. For all other compounds, the stereoselective preference of BChE and AChE were opposite.

In Silico Modelling
Since the final aim of the study was to explore the possibility of fluorinated CD's and CN's to be used as a drug for the treatment of diseases related to the changes in acetylcholine levels, particularly in CNS, we evaluated the drug-likeness of the compounds and their ability to cross the blood-brain barrier by active or passive transport.
Physico-chemical properties (molecular weight (MW), partition coefficient (logP), hydrogen bonds donors (HBD), hydrogen bond acceptors (HBA), number of rotatable bonds (RB) and polar surface area (PSA)) are presented in Table 5. They were equal for pseudo-enantiomers (CD and CN) and only results for CD derivatives are given. The physicochemical properties of 9 compounds were within the recommended range for orally active drugs in humans indicating their ability to pass the interface between blood and intestine by passive transport. The clogP, HBD, HBA and PSA values were in the desirable range in all tested compounds. In 14 compounds the MW value deviated from the recommended value, but it is relatively close to it. Thus, it can be assumed that those compounds would have a moderate possibility to passively transport through lipid membranes. The ability of compounds to penetrate the blood-brain barrier, the log BBB value, was determined using BIOVIA ADMET Protocol (ADMET_BBB) implemented in Biovia Discovery Studio Client v18.1. According to the BIOVIA ADMET Protocol (ADMET_BBB) all of the tested compounds can penetrate the BBB by passive diffusion after oral administration, Table 6. Eight of the tested compounds are very high BBB penetrants having log BBB values within the range 0.773-0.894 (the concentration of the compound in the brain is at least five times higher than in the blood), while thirty-eight compounds are high BBB penetrants with logBBB within the range 0.256-0.671 (the ratio of compound concentration in the brain and blood is between 1:1 up to 5:1).

PCA and Activity/PES Model Established by Machine Learning
Principal component analysis was performed on the compounds' inhibitory activities for BChE and AChE. The first principal component explained 99.74% of the total variance among the data assuring the proper description of the activities with only one-dimensional reduced space. All investigated compounds were classified according to their score values and the results are presented in Figure 3.  According to the results of PCA, the 1st principal component was the most important in describing the inhibition activity of the compounds. To establish a connection between inhibition of AChE and BChE and calculated potential energy surfaces of the compounds, an activity/PES model was created by using the first three principal components of the reduced PES data. The 1st principal component for inhibition of AChE and BChE was regressed on these three principal components.
An extensive machine learning procedure was utilized to build the best multivariate linear regression model. During the calculation all possible regression models were generated with determined B-matrices of coefficients (Equation (2)). These models were cross-validated using the leave-one-out cross-validation procedure [30]. The best possible regression models were determined based on following parameters: adjusted and predicted R 2 values, LOO-CV mean squared error and the number of variables in the models.
For fluorinated CD-derivatives the best calculated 3D regression model is presented in Figure 4. The value of adjusted R 2 was 0.9936 and predicted R 2 0.9932 (R 2 was 0.9971). In this calculated model the high values of adjusted R 2 and predicted R 2 ensures that there is no overfitting, and the models are thoroughly validated. The best calculated 3D regression model for fluorinated CN-derivatives is presented in Figure 5. Again, the high values of adjusted R 2 (0.9849) and predicted R 2 (0.9879) (R 2 was 0.9899) assure that there was no overfitting in the model and that the model was thoroughly validated. Established activity/PES regression models can be used for the prediction of AChE and BChE inhibition for new similar compounds based solely on the reduced space of compounds' potential energy surfaces obtained from ab initio molecular dynamics simulation.

Chemicals
Reagents and solvents for the preparation of compounds were purchased from Sigma-Aldrich (St. Louis, MO, USA), Fluka. The reactions were monitored by thin-layer chromatography plates coated with silica gel (Sigma-Aldrich, St. Louis, MO, USA). TLC plates were visualized by UV irradiation (254 nM) or by iodine fumes. 1D and 2D 1 H and 13

Enzymes Activity Measurement
Enzyme activities were measured spectrophotometrically by the Ellman method [32] at 412 nM using 0.30 mM DTNB as thiol reagent, ATCh as substrate and water in 0.1 M phosphate buffer, pH 7.4. For the inhibition measurements, the reaction mixture contained tested compounds instead of water. No side interactions of tested compounds with ATCh or DTNB were detected. Measurements were done at 25 • C on a Tecan Infinite M200Pro plate reader (Austria). A detailed description of measurement protocol was described previously [18].

Enzyme-Inhibitor Dissociation Constants
The reversible inhibition of BChE and AChE by tested compounds was measured by determining the decrease of enzyme activity towards ATCh in their presence.
where y-intercept determines the enzyme-inhibitor dissociation constants (K (I) ), while xintercept determines the enzyme-substrate dissociation constant, K (S) . The determination of kinetic constants was carried out using the GraphPadPrism 6.0 program (GraphPad Software, San Diego, CA, USA).

In Silico Prediction of Drug-Likeness
The evaluation of compounds' drug-likeness was carried out in terms of evaluation of their ability to pass through the intestinal after it is consumed orally and penetrate the lipid-based cell membrane from the blood. The drug-likeness evaluation was performed based on the values of physicochemical properties considered important for the ability of the compounds to be orally active drug in humans: molecular weight (MW), partition coefficient (logP), hydrogen bonds donors (HBD), hydrogen bond acceptors (HBA), number of rotatable bonds (RB) and polar surface area (PSA) [33,34]. The calculated values of those properties (Table 5) were compared to generally accepted recommended values: MW from 180 to 500, logP ranged from −0.4 to +5.6, HB < 5, HBA < 10 and RB ≤ 10, and PSA < 140 Å 2 and those that have no more than one violation of those rules are orally active compounds [33][34][35]. The values of all physicochemical properties were calculated using the ChemAxon Chemicalize 2018 platform (accessed on 30 September 2022) [36].

In Silico Prediction of Blood-Brain Barrier Penetration
The ADMET descriptors protocol implemented in BIOVIA Discovery Studio 2018 predicts BBB penetration after the oral administration of a drug. The model contains a quantitative linear regression model for the prediction of blood-brain penetration, as well as 95% and 99% confidence ellipses derived from a correlation of the polar surface area (PSA-2D) and atom-based LogP (AlogP98) parameters derived from over 800 compounds known to enter the CNS after oral administration. Based on this model there are four BBB penetration levels: 0 (very high penetrant, logBBB ≥ 0.7), 1 (high penetrant, 0 ≤ logBBB < 0.7), 2 (medium penetrant, −0.52 < logBBB < 0) and 3 (low penetrant, logBBB ≤ −0.52). Log BBB denotes the base 10 logarithm of the ratio of the concentration of a compound measured in the brain to the concentration of the compound measured in the blood at a steady state [37].

Principal Component Analysis
Multivariate analyses were conducted by a 2nd-order tensor analysis tool known as principal component analysis (PCA) [38,39]. PCA data matrix X, which has rank r, is decomposed in the sum of r matrices t i p τ i . Each matrix t i p τ i has a rank of 1.
t i is a vector of scores and p τ i is a vector of loadings. PCA provides the best linear projection of multidimensional data by minimizing the least squares objective function. Scores are used for classification, while loadings can be used for the variability identification among the data. PCA development goes back to Beltrami [40] and Pearson [41], and the name was introduced by Harold Hotelling [42].
Inhibition data were arranged in the data matrix X and PCA on the covariance matrix was performed by our parallelized code for multi-and univariate analysis [43][44][45]. Extraction of eigenvectors was based on the NIPALS algorithm [46] and obtained 1st principal component was subsequently used as a regressed variable.

Sampling of the Potential Energy Surfaces
Ab initio molecular dynamics simulations with on-the-fly calculations of forces were used as a sampling procedure for potential energy surfaces (PES). Equations were integrated using the velocity Verlet algorithm [47]. PM7 method [48] implemented in MOPAC2016 [49] was used for the calculation of forces at each point of the simulation. Molecular dynamics were conducted by using our in-house developed program qcc [50,51]. Phase space coverage was ensured by setting the initial temperature for Maxwell distribution of velocities to 773.15 K. During the simulation temperature was controlled using the velocity scaling algorithm.
Step size was 0.5 fs and a total of 5,000,000 steps was computed for each compound. PES of compounds spanned in multidimensional space of Cartesian coordinates were decomposed by PCA providing principal components for further regression modelling.

Machine Learning Multivariate Linear Regression
Reduced spaces of measured inhibition data were used as dependent variables for estimation of fluorinated Cinchona alkaloids derivatives inhibition activities. Principal components were extracted by the 2nd-order tensor decomposition method. These principal components were regressed on the theoretically computed energy fingerprints of all compounds by performing extensive machine learning (ML).
ML procedure was applied for the generation of all possible multivariate linear regression models with a linear combination of original variables as well as their higher-order polynomial terms. Multivariate linear regression was performed using the following expression for matrices of coefficients B calculated by singular value decomposition: where X and Y are the matrices of independent and dependent variables, respectively. Every possible regression model of inhibition activity in dependence on molecular dynamics data was built and thoroughly validated by the leave-one-out cross-validation technique (LOO-CV). 3D models were inspected up to the 4 th order and the total number of investigated models was 17,179,869,184. The most optimal representations were selected based on the adjusted and predicted R 2 values, LOO-CV mean squared error as well as the total number of variables in the models.

Conclusions
We have prepared a series of 46 quaternized derivatives of cinchonidine and cinchonine as well as their 10,11-dihydro analogs. Quinuclidine nitrogen atom was quaternized with benzyl group having fluorine atoms in different positions, trifluoromethyl and trifluoromethoxy groups as well as combinations of them with bromine and chlorine atoms. Fluorination of compounds generally affected the acidity and basicity of the parent compound influencing compound's binding affinity to cholinesterases, pharmacokinetic properties and bioavailability. All prepared compounds showed good to excellent inhibitory potential toward BChE with inhibition constants from nano-to micromolar range. CD derivative with a fluorine atom in meta position on the benzene ring was the most selective BChE inhibitor among the tested compounds showing 533 times higher preference for BChE. Also, based on in silico modelling all compounds possess blood-brain barrier penetration ability and inhibition potential toward enzyme CYP2D6. The best possible models of AChE and BChE inhibition with predicted R 2 (CD-derivatives) =0.9932 and R 2 (CN-derivatives) =0.9879 were calculated and cross-validated by utilizing an extensive machine learning protocol. Multivariate linear regression models with a linear combination of original variables and their higher-order polynomial terms were generated and tested. The best models among all generated and tested cases were determined. These activity/PES models can be used for accurate prediction of AChE and BChE inhibition for new similar compounds based solely on their potential energy surfaces calculated from ab initio molecular dynamics simulation enabling a smart guided search for new potential leads. These results strongly encourage further optimization of quaternized cinchonidine and cinchonine structural motifs in further research toward finding selective cholinesterase inhibitors.