Combining QSAR and Molecular Docking for the Methodological Design of Novel Radiotracers Targeting Parkinson’s Disease

Abstract

Parkinson’s disease (PD) is a neurodegenerative disorder marked by the progressive loss of dopaminergic neurons in the nigrostriatal pathway. The dopamine active transporter (DAT), a key protein involved in dopamine reuptake, serves as a selective biomarker for dopaminergic terminals in the striatum. DAT binding has been extensively studied using in vivo imaging techniques such as Single-Photon Emission Computed Tomography (SPECT) and Positron Emission Tomography (PET). To support the design of new radiotracers targeting DAT, we employ Quantitative Structure–Activity Relationship (QSAR) analysis on a structurally diverse dataset composed of 57 compounds with known affinity constants for DAT. The best-performing QSAR model includes four molecular descriptors and demonstrates robust statistical performance: R² = 0.7554, Q²_LOO = 0.6800, and external R² = 0.7090. These values indicate strong predictive capability and model stability. The predicted compounds are evaluated using a docking methodology to check the correct coupling and interactions with the DAT. The proposed approach—combining QSAR modeling and docking—offers a valuable strategy for screening and optimizing potential PET/SPECT radiotracers, ultimately aiding in the neuroimaging and early diagnosis of Parkinson’s disease.

1. Introduction

Parkinson’s disease (PD) is the second most common neurodegenerative disorder after Alzheimer’s disease, affecting approximately 1% of the global population over the age of 60 [1,2,3]. It is caused by the degeneration of dopaminergic neurons located in the substantia nigra pars compacta [4]. This neuronal loss leads to a decrease in dopamine release in the striatum, which results in the hallmark symptoms of PD, which are bradykinesia, rigidity, postural instability, tremors, and a significant decline in patients’ quality of life [5].

The diagnosis of PD still relies on clinical criteria and remains challenging [6], particularly in its early stages when symptoms may resemble those of other disorders [7] such as multisystem atrophy, progressive supranuclear palsy, essential tremor, and vascular or drug Parkinsonism [5,6,8]. Therefore, the use of tools that allow for an accurate and early diagnosis of PD is crucial, given their confirmed impact on both prognosis and treatment outcomes [9]. One such approach involves evaluating the integrity of the nigrostriatal dopaminergic pathway using dopamine active transporters (DATs) [10]. The DATs are a group of transmembrane proteins associated with chloride channels, located at the presynaptic terminals of dopaminergic neurons [11]. They are responsible for the reuptake of dopamine after its release into the synaptic cleft [11,12]. In patients with PD, because of the degeneration of these neurons, the density of DATs is significantly reduced [13].

Non-invasive molecular neuroimaging techniques, such as SPECT (Single-Photon Emission Computed Tomography) and PET (Positron Emission Tomography), enable in vivo visualization of various tissues and organs using radiopharmaceuticals [12,14]. Several PET and SPECT radiotracers have been developed to target DATs [15]. Among them are radio-iodinated cocaine analogues, such as ¹²³I-Ioflupane (¹²³I-N-ω-fluoropropyl-2β-carbomethoxy-3β-(4-iodophenyl)nortropane) and ¹²³I-β-CIT (¹²³I-2β-carbomethoxy-3β-(4-iodophenyl)tropane) [16]. These SPECT radiopharmaceuticals are widely used to assess the integrity of the nigrostriatal dopaminergic pathway and have demonstrated good diagnostic sensitivity and specificity for suspected PD cases [15]. However, both tracers require pre-treatment with a thyroid-blocking agent to protect the thyroid gland from radioactive iodine uptake [17]. Additionally, imaging with ¹²³I-β-CIT must be conducted 24 h post-injection, which can be inconvenient for outpatient settings [18]. There are also PET radiotracers with affinity for DAT, such as ¹¹C-N-(3-iodoprop-2(E)-enyl)-2β-carbomethoxy-3β-(4′-methylphenyl)nortropane (¹¹C-PE2I) [16]. In spite of its high specificity, the short half-life of ¹¹C (approximately 20 min) limits its clinical utility, as it requires on-site cyclotron production [19]. These limitations underscore the need for the development of new DAT-targeting radiopharmaceuticals that offer improved pharmacokinetic properties and logistical advantages, enabling better assessment of the nigrostriatal pathway in the early diagnosis of PD.

The estimated cost and time required to bring a new drug to market are approximately USD 2.8 billion and 7 to 12 years [20]. Moreover, of the roughly 40,000 compounds tested in animals, only five typically proceed to human trials, and just one in five of those that reach clinical studies ultimately receives approval [21]. This reflects a significant expenditure of financial, human, and time resources. In response, virtual (computational) screening has emerged as a promising alternative to the traditional “trial-and-error” approach in identifying new chemical entities with biological activity. As a result, the pharmaceutical industry has shown growing interest in computational methodologies [22]. Among these approaches, QSAR (Quantitative Structure–Activity Relationship) modeling stands out as a valuable in silico technique connecting chemical structure with pharmacological activity. QSAR offers clear advantages in terms of cost- and time-efficiency, especially during the early stages of drug discovery, by enabling the prioritization of the most promising compounds for further evaluation [23,24,25,26].

In recent years, numerous computational studies have been undertaken to identify novel compounds with potential radiopharmaceutical applications (references). Notably, some of these investigations have targeted specific biological systems, such as the dopamine (D₂) receptor and the adenosine A_2A receptor, the latter with an emphasis on antagonist identification. A significant number of studies have also concentrated on modeling interactions with the dopamine transporter. However, many of these efforts have relied on relatively limited compound databases or congeneric series, which restricts the generalizability and predictive power of their findings.

In recent years, a growing number of computational studies have aimed to identify novel compounds with potential radiopharmaceutical applications [27,28,29,30]. These efforts have usually focused on specific molecular targets, including the dopamine (D₂) receptor [27] and the adenosine A_2A receptor (particularly in the context of antagonist discovery) [28]. In parallel, several studies have focused on modeling interactions with the dopamine transporter [29,30]. In spite of their contributions, many of these studies are based on relatively small datasets or congeneric compound series, which limits the extrapolation, robustness, and external validation of the resulting models. Given this context, the main aim of the present work is to propose a well-validated computational model based on a structurally diverse dataset, capable of identifying the key structural features required for DAT affinity, and to integrate it with molecular docking studies. This combined approach aims to facilitate the prediction of novel radiopharmaceutical candidates with potential utility in the diagnosis of Parkinson’s disease.

2. Materials and Methods

2.1. Dataset

The database used in the QSAR study is composed of 57 compounds that were taken from the previously published literature [31,32]. These compounds cover different pharmacological groups, are structurally diverse, and belong to the Binding Database. All of them have affinity reports for the DAT, evaluated by determining the value of the dissociation equilibrium constant (K_d) at DAT binding, using this parameter to define the affinity of the molecules to be bound to the transporter. The K_d value was determined by measuring the activity of the ³H-labeled radioligand bound to the DAT using an adjustment of the method proposed by Pristupa et al. [33]. The K_d values ranged from 70 to 52,000 nM (

Table 1. Reported affinity for DAT expressed as dissociation equilibrium constant (K_d) obtained from the Binding Database.

No	Ligand	Kd (nM)	Log Kd	No	Ligand	Kd (nM)	Log Kd
1	BDBM82071	9300	3.968	30	BDBM50014247	2050	3.312
2	BDBM50069447	360	2.556	31	BDBM50028091	9200	3.964
3	BDBM50069446	3500	3.544	32	BDBM50010859	8500	3.929
4	BDBM50069452	42,000	4.623	33	BDBM81448	6530	3.815
5	BDBM50069449	340	2.531	34	BDBM82437	18,000	4.255
6	BDBM50069453	70	1.845	35	BDBM50101973	1000	3.000
7	BDBM50069450	21,400	4.330	36	BDBM31005	9400	3.973
8	BDBM50069445	12,000	4.079	37	BDBM50180661	420	2.623
9	BDBM50366567	1080	3.033	38	BDBM112777	1140	3.057
10	BDBM50069451	7260	3.861	39	BDBM85218	4340	3.637
11	BDBM30130	3600	3.556	40	BDBM50105417	8400	3.924
12	BDBM50069442	27,000	4.431	41	BDBM50176062	2100	3.322
13	BDBM50069443	24,000	4.380	42	BDBM50022784	1080	3.033
14	BDBM50069444	340	2.531	43	BDBM50113851	5100	3.708
15	BDBM50069454	4310	3.634	44	BDBM50073444	7400	3.869
16	BDBM50069448	140	2.146	45	BDBM50240410	3780	3.577
17	BDBM22416	490	2.690	46	DB00655	13,500	4.130
18	BDBM50020712	3250	3.512	47	BDBM18860	8100	3.908
19	BDBM22870	4310	3.634	48	DB00661	20,000	4.301
20	BDBM50048392	520	2.716	49	BDBM50017712	41,000	4.613
21	BDBM85220	3940	3.595	50	BDBM50022723	2900	3.462
22	BDBM25870	28,100	4.449	51	BDBM86207	220	2.342
23	BDBM77970	2190	3.340	52	BDBM50359499	2800	3.447
24	BDBM35229	3190	3.504	53	BDBM81466	1060	3.025
25	BDBM85219	18,300	4.262	54	BDBM50017674	2200	3.342
26	BDBM50028066	129	2.111	55	DB00988	2400	3.380
27	BDBM82548	5310	3.725	56	BDBM35234	16,200	4.210
28	BDBM50079527	12,100	4.083	57	BDBM50021952	207	2.316
29	BDBM82438	52,000	4.716

); because of the large affinity span of the compounds, we decided to apply a transformation and use Log K_d as the dependent variable. All standard parametric assumptions (including linearity, normality of residuals, homoscedasticity, and absence of multicollinearity) were evaluated using appropriate statistical diagnostics. The complete set of assumption checks, including plots and test results, is available in the Supplementary Materials.

2.2. Calculation of Molecular Descriptors

To calculate the molecular descriptors (MDs) used in this study, we employed Dragon software version 7.0.10 [34], developed by the Milano Chemometrics and QSAR Research Group. Dragon is one of the most widely recognized tools for molecular descriptor calculation. The study focused exclusively on 0D to 2D descriptor families, including constitutional indices, topological indices, functional group counts, molecular properties, and 2D-autocorrelations, among others. A total of 3839 descriptors were initially generated. After that, the variables with constant or near-constant values were removed and, consequently, not used in further analysis. In addition, variables with a standard deviation (SD) lower than 5% and a pair correlation over 85% were also discarded from subsequent analysis.

2.3. Multiple Linear Regression Model

MLR (multiple linear regression) is the process that obtains a linear relationship between DAT affinity and molecular descriptors using the OLS (ordinary least squares) method [35]. We used the QSARINS (QSAR-Insubria) software (version 2.2.4 year 2019) [36] developed by the QSAR Research Unit, University of Insubria, Italy. It allows the creation of MLR models using the GA (genetic algorithm) method for variable selection [35,37]. This software permits dividing the database into training and prediction sets by either modeled endpoint responses or structure; in our case, we used the first option. Figure 1 shows the structures of the molecules included in the training (learning) set, and Figure 2 shows the structures of the compounds used in the prediction set. The best model was selected based on several criteria such as a high value of R² (good fitting) and Q²_LOO (robustness), a low value for R²–Q²_LOO (stability) and K_XX (low correlation of descriptors), and the greatest R²_ext (predictability), among others, so that it has a slight difference between fitting, cross-validation, and external validation parameters.

2.4. Validation of the QSAR-MLR Model

The obtained model must be carefully evaluated and rigorously validated to ensure its reliability. QSARINS provides several tools to assess whether the model complies with the standards established by the Organization for Economic Cooperation and Development (OECD) for the development, validation, acceptance, and use of QSAR models. Adhering to these guidelines enhances confidence in the predictive reliability of the model. According to OECD principles, a valid QSAR-MLR model must satisfy the following criteria [38,39]:

1

A defined endpoint;

2

An unambiguous algorithm;

3

A defined domain of applicability;

4

Appropriate measures of goodness-of-fit, robustness, and predictability;

5

A mechanistic interpretation, if possible.

To comply with the 1st principle, the dissociation equilibrium constant (K_d) between the molecules and the DAT was used as the endpoint. For compliance with the 2nd principle, molecular descriptors were calculated using Dragon software, while the QSARINS platform was employed to develop the MLR model as we explained in the previous section. This algorithm allows us to correlate the chemical structure descriptors and their binding affinity for DAT.

The definition of the applicability domain (AD) of the model is considered a critical aspect in chemometric studies and is the 3rd principle. Only predictions for compounds that fall within the AD may be considered reliable, but not the extrapolations of the models, even if they show robustness and significance. In order to define the AD of the model, the Williams graph/plot was used, so that reliable predictions of the model have leverage values lower than the critical leverage, ranging within ±3 standard deviations [40]. Compounds falling outside these ranges were considered outliers. In addition, the Insubria graph (generated by QSARINS) was utilized to assess further the reliability of predictions. This graphical tool is particularly useful for visualizing the position of molecules with unknown experimental responses and comparing them with those in the prediction set, helping to evaluate the confidence in their predicted values.

Following the 4th principle, the goodness of the fit was evaluated using the coefficient of determination R² and its modified form R²_adj; the latter accounts for the number of descriptors in the model and penalizes the inclusion of non-informative variables. To evaluate the robustness of the model, several internal validation strategies were employed. First, the Q²_LOO metric was used, applying Leave-One-Out (LOO) cross-validation, where each compound in the dataset is sequentially excluded, and the model is recalculated to predict the excluded compound. Additionally, QSARINS provides a more rigorous validation approach, Leave-Many-Out (LMO) cross-validation, which was also applied. In this method, between 5% and 30% of the dataset was randomly excluded in each iteration, allowing us to examine the model’s stability and predictive behavior under more challenging conditions. To ensure that the model is not the result of a casual correlation, a randomization test (Y-scrambling) was performed [41]. In this procedure, the response values (K_d) are randomly shuffled, breaking any real association with the molecular descriptors. A valid model is expected to show a significant drop in performance under these conditions, confirming that the original model’s predictive power is not because of chance.

To further confirm the predictive capability of the model, an external validation was carried out. The external set included several widely used reference radiopharmaceuticals, which were excluded from model training and used solely to assess predictability. The affinity values of these compounds (see Figure 2) were predicted using the final MLR model equation, providing an independent evaluation of the model’s generalization ability.

2.5. Virtual Screening and Docking Experiment

With the aim of proving the possibilities of our approach to identifying new compounds with potential affinity for DAT, we performed a virtual screening of several compounds that have different modes of action in the central nervous system (CNS). The structure of these compounds is shown in Figure 3. We performed virtual screening with eleven compounds to evaluate the performance of our model with respect to those compounds and identified new compounds with possible affinity for DAT. The compounds were selected based on the following criteria: (1) chemical diversity (representing different chemical scaffolds to explore diverse interactions within the binding site of the target); (2) drug-likeness and physicochemical properties (the compounds comply with Lipinski’s Rule of Five and exhibit acceptable ADME profiles); (3) commercial availability (compounds readily purchasable from chemical suppliers for potential future in vitro validation). After that, docking experiments were performed to evaluate the affinity of the predicted compound with the receptor. The optimized structure of each molecule was used for docking experiments using Autodock Vina software v1.2.3 [42]. The dopamine transporter was the protein used in this study, which was downloaded from the Protein Data Bank with code 8Y2D, resolved at 2.80 Å [43]. The dopamine was prepared in vitro by the addition of hydrogen atoms at the physiological pH of 7.4 and by the elimination of the water molecules around the protein. Ligands were prepared to account for rotatable bonds using Autodock Tool v1.5.7. This same program was used for the treatment of the dopamine transporter protein before the docking experiments, setting the grid size to 40 × 40 × 40 ų located at the center of mass of L-dopamine in the crystal structure of the dopamine transporter.

The coordinates of the grid box were x = 112.241 Å, y = 105.521 Å, and z = 106.058 Å, with a grid spacing of 0.375 Å, covering a volume of 30 × 30 × 30 ų and located in the center of mass of L-dopamine in the crystal structure of the dopamine transporter to completely cover the binding site and adjacent regions, allowing flexibility in the exploration of poses [44,45], which is consistent with similar studies in G-protein-coupled receptors (GPCRs) [46]. The mode number was 50, and the energy rank was set to 1 kcal/mol. The best docking poses were selected following two criteria: the relative total binding energy score and the positional root-mean-square deviation (RMSD) [47,48], comparing each docking ligand result with the position of L-dopamine in the crystallographic structure of the dopamine transporter from the Protein Data Bank [49,50,51]. Each RMSD value was calculated using the Pymol program v2.5.0 [52,53].

2.6. Molecular Dynamic Simulation Protocol

The starting point for preparing the complexes for molecular dynamics simulations was to obtain the best poses derived from docking experiments. These complexes were those formed by the compounds Diethylpropion, Procaine, and Benztropine (see Figure 3) with the dopamine transporter, whose PDB ID is 8Y2D, resolved at 2.80 Å [43]. The protein was prepared by adding hydrogen atoms at physiological pH (pH = 7.4) using the UCSF Chimera software version 1.14 [54,55,56]. The force field used was CHARMM36 for proteins [57], and parameters for organic molecules were obtained from the SwissParam web server [58].

The ligand–protein complexes were placed in a lipid bilayer formed by the POPC lipid model, which was constructed using the CHARMM-GUI server [59,60,61]. This lipid bilayer contains 55,160 lipid molecules. These complexes, along with the membrane, were placed into a cubic water box of 15.0 × 15.0 × 15.0 ų centered on the center of mass of each ligand using the TIP3P water model [62,63,64]. All the complexes were submitted to 2000 steps for energy minimization using the conjugate gradient methodology [65,66] at a temperature of 298.15 K using the weak coupling algorithm [67,68,69]. The Van der Waals cutoff was fixed to 12 Å, under the NPT ensemble (constant number of particles, pressure, and temperature). The complexes studied were submitted to a 1.0 fs time step under the velocity Verlet algorithm [70,71] and 2.0 ns of equilibration and 100 ns of molecular dynamics simulation using the NAMD 2.13 software package [72,73,74,75].

3. Results

3.1. Development of the QSAR-MLR Model

Large numbers of models were obtained while exploring the best combinations of molecular descriptors that show a high correlation with the response variable (LogK_d). Therefore, an analysis of the model’s parameters was performed considering the principle of parsimony. Based on this analysis, a QSAR-MLR model was developed to evaluate the binding affinity for DAT (so that it can describe the maximum information with the least number of descriptors). The equation and the statistical parameters of the best model are as follows:

Log K_d = 3.538 (±0.204) − 0.403 (±0.204) × O-058 + 0.162 (±0.039) × CATS2D_03_AL
− 0.133 (±0.039) * CATS2D_08_LL − 0.537 (±0.291) * B08[C-N]

R2 = 0.755	R2adj = 0.729	R2 − R2adj = 0.026	LOF = 0.195
Kxx = 0.236	ΔK = 0.076	RMSEtr = 0.358	MAEtr = 0.276
RSStr = 5.377	CCCtr = 0.861	s = 0.381	F = 28.566

where R² is the coefficient of determination, R²_adj is adjusted R², s is the standard error of estimate, F is the variance ratio, and LOF is Friedman lack of fit. The parameter K_xx is the correlation among descriptors, and ΔK is the difference of the correlations between the descriptors (K_x) and the descriptors plus the responses (K_x). Moreover, RMSE_tr is the root-mean-square error of fit (for the training set), RSS_tr is the residual sum of squares in the fit, and CCC_tr is the concordance correlation coefficient calculated over the training set [76].

Table 2. Experimental and predicted values (expressed as log K_d) of DAT affinity for the compounds in the training set.

Compound	Exp. a Endpoint	Predicted	Residual	Pred. LOO b	Pred. LOO Res.	Leverage Value c
1	3.968	3.974	0.006	3.975	0.007	0.066
2	2.556	3.066	0.510	3.102	0.546	0.065
4	4.623	-outlier-
5	2.531	2.878	0.347	2.909	0.378	0.080
6	1.845	2.183	0.338	2.272	0.427	0.208
7	4.330	4.349	0.019	4.351	0.021	0.122
8	4.079	3.653	−0.426	3.594	−0.485	0.122
9	3.033	3.145	0.112	3.151	0.118	0.049
10	3.861	4.187	0.326	4.227	0.366	0.110
12	4.431	-outlier-
13	4.380	3.491	−0.889	3.372	−1.008	0.118
15	3.634	3.545	−0.089	3.542	−0.093	0.041
16	2.146	2.021	−0.125	1.985	−0.161	0.221
17	2.690	3.012	0.322	3.033	0.343	0.062
18	3.512	3.164	−0.349	3.130	−0.382	0.087
20	2.716	2.821	0.105	2.843	0.127	0.171
21	3.595	3.326	−0.269	3.305	−0.290	0.072
22	4.449	4.298	−0.151	4.283	−0.166	0.094
23	3.340	3.679	0.339	3.696	0.356	0.048
25	4.262	4.136	−0.126	4.126	−0.136	0.077
26	2.111	2.601	0.490	2.654	0.543	0.097
27	3.725	3.164	−0.562	3.110	−0.615	0.087
28	4.083	3.974	−0.109	3.966	−0.117	0.066
29	4.716	-outlier-
30	3.312	2.907	−0.405	2.861	−0.451	0.103
31	3.964	3.974	0.010	3.975	0.011	0.066
32	3.929	3.974	0.045	3.977	0.048	0.066
34	4.255	4.058	−0.198	4.034	−0.221	0.106
35	3.000	3.001	0.001	3.002	0.002	0.106
37	2.623	3.116	0.493	3.139	0.516	0.043
38	3.057	3.001	−0.056	2.995	−0.062	0.106
40	3.924	3.538	−0.386	3.488	−0.436	0.115
42	3.033	3.145	0.112	3.151	0.118	0.049
43	3.708	3.538	−0.170	3.516	−0.192	0.115
44	3.869	3.276	−0.593	3.225	−0.645	0.079
45	3.577	4.298	0.721	4.373	0.796	0.094
46	4.130	4.108	−0.022	4.105	−0.026	0.129
47	3.908	3.953	0.045	3.958	0.050	0.093
48	4.301	3.909	−0.392	3.764	−0.538	0.270
49	4.613	4.313	−0.300	4.114	−0.499	0.400
51	2.342	2.310	−0.032	2.299	−0.043	0.260
52	3.447	3.538	0.091	3.550	0.103	0.115
53	3.025	3.571	0.546	3.754	0.729	0.250
55	3.380	4.187	0.807	4.286	0.906	0.110
57	2.316	2.177	−0.139	2.127	−0.189	0.266

^a Experimental values; ^b LOO: Leave-One-Out; ^c the critical Leverage value is h* = 0.3571.

shows the results of the training set, and Figure 4 shows the scatter plot of predicted vs. observed values.

3.2. Applicability Domain of the Model

The leverage (h) and standardized residual approach, described in the literature as a Williams plot (Figure 5A), was used for the analysis of the applicability domain. The study shows the graph for the training and prediction sets. In addition, the Insubria plot allows easy visual inspection of the diagonal values of the leverage values (HAT) against the predicted responses for each compound. This second approach was introduced in the QSARINS software (Version 2.2.4 year 2019) and is shown in Figure 5B.

3.3. Validation of the Developed Model

The results of the internal validation experiments demonstrated both fitness and stability of the model. The most important techniques for internal validation are LOO, LMO, and Y-scrambling. The values achieved in these experiments showed a Q²_LOO value of 0.680, an RMSE_cv value of 0.409, and MAE_cv = 0.316. Figure 6A shows the values predicted by LOO vs. the experimental LogK_d values. The model performance for internal validation using LMO was Q²_LMO = 0.648; Figure 6B shows the correlation between descriptors and DAT binding affinity (Q²_LMO vs. K_XY values).

The results of the Y-scrambling procedure are shown in Figure 7: the values of R² and Q² of every iteration (yellow and red circles, respectively) and their averages were R²_Y-scr = 0.098 and Q²_Y-scr = −0.174, as well as the respective values of these parameters for the model (light and dark blue, for R² and Q², respectively).

The predictability of the model was tested using the set of external predictions of 12 compounds. Its parameters were R²_ext = 0.709, RMSE_ext = 0.339, MAE_ext = 0.302, PRESS_ext = 1.381, Q²-F1 = 0.298, Q²-F2 = 0.250, Q²-F3 = 0.780, CCC_ext = 0.772, r²m_aver = 0.424, and r²m_ delta = 0.331, where R²_ext is the external determination coefficient; RMSE_ext is the root-mean-square error in the external prediction; MAE_ext is the mean absolute error in the external prediction; PRESS_ext is the predictive residual sum of squares (external validation); Q²-F1, Q²-F2, and Q²-F3 are the explained variances in the external prediction; CCC_ext is the concordance correlation coefficient; and r²m-aver and r²m-delta are the Roy criteria, average and delta, respectively [76]. The predictions of compounds in the external set are shown in Figure 4, Figure 5 and Figure 6 (blue circles) and

Table 3. Experimental and predicted values (expressed as LogK_d) of DAT affinity for the compounds in the prediction set.

Compound	Exp. a Endpoint	Predicted	Residual	Leverage Value b
3	3.544	3.707	0.163	0.045
11	3.556	3.116	−0.440	0.043
14	2.531	2.421	−0.110	0.118
19	3.634	4.032	0.398	0.070
24	3.504	3.974	0.470	0.066
33	3.815	3.650	−0.165	0.059
36	3.973	4.511	0.538	0.139
39	3.637	3.164	−0.473	0.087
41	3.322	3.001	−0.321	0.106
50	3.462	3.538	0.076	0.115
54	3.342	3.517	0.175	0.047
56	4.210	4.511	0.301	0.139

^a Experimental value; ^b the critical leverage value is h* = 0.3571.

.

3.4. Comparison with Other Approaches

In this study, our MLR model is compared with other approaches from the literature in the task of predicting DAT inhibitors. The results obtained by using the MLR-based model are shown in

Table 4. Comparison between computational models developed to identify novel compounds with potential radiopharmaceutical applications.

Target	Technique	Structures	No Var.	N (Train)	N (Test)	R2	R2adj	Q2LOO	Q2LMO	RMSEcv	R2ext	Ref.
D2	PLS	Congeneric	5	27	7	0.731	0.696	0.623	nr	nr	0.742	[27]
A2A	ICP-MLR	Congeneric	5	25	10	0.901	0.875	0.850	nr	nr	0.540	[28]
DAT	PLS	Congeneric	2	26	10	0.842	nr	0.761	0.746	0.546	0.997	[30]
DAT	MLR	Congeneric	10	69	np	0.814	0.692	nr	nr	nr	nr	[29]
DAT	PLS	Congeneric	10	69	np	0.807	0.732	nr	nr	nr	nr	[29]
DAT	GFA	Congeneric	10	69	np	0.718	0.67	nr	nr	nr	nr	[29]
DAT	MLR	Diverse	4	45	12	0.755	0.729	0.680	0.648	0.409	0.709	TW

PLS: partial least squares; nr: not reported; ICP-MLR: Intelligent Consensus Predictor Multiple Linear Regression; MLR: Multiple Linear Regression; GFA: Genetic Function Approximation; np: not performed; TW: this work.

, where the models obtained with other approaches are also shown.

A comparative analysis of the presented models reveals key differences in performance, dataset composition, and validation rigor. As shown in Table 4, the previously developed studies to identify novel compounds with potential radiopharmaceutical applications were all based on congeneric series, which limits their applicability to chemical structures that differ from those used during model training. In contrast, our model was developed using a dataset with greater structural diversity, thereby enhancing its ability to predict compounds outside the original chemical space.

We would like to highlight that the model targeting the adenosine A_2A receptor inhibitors [28] achieved the highest internal coefficient of determination (R² = 0.901) among all the models; however, it also exhibited the lowest external R² value (0.540), indicating limited predictive performance on the test set. Similarly, the PLS model for dopamine D₂ receptor [27] showed acceptable internal performance (R² = 0.731; Q²_LOO = 0.623) as well as external performance (R²_ext = 0.742), but also relied on a small, structurally homogeneous dataset.

Among the models targeting the dopamine transporter (DAT), the PLS model reported by de Oliveira et al. [30] achieved the highest external predictivity (R²_ext = 0.997), along with strong internal validation metrics (Q²_LOO = 0.761; Q²_LMO = 0.746). However, such a pronounced discrepancy between internal fit values and external predictability is uncommon in computational modeling and is likely attributable to the specific selection of the external test set. The models reported by Mavel et al. [29] were built using the largest dataset (69 compounds), which is advantageous in terms of statistical power; however, the absence of appropriate cross-validation and external validation metrics weakens confidence in their robustness and predictive reliability. By contrast, the model proposed in the present work stands out because of its use of a structurally diverse dataset and its comprehensive validation strategy. In spite of a slightly lower adjusted R² (0.755), it demonstrated consistent performance across internal (Q²_LOO = 0.680; Q²_LMO = 0.648; RMSEcv = 0.409) and external metrics (R²_ext = 0.709), suggesting a better balance between robustness, generalizability, and practical utility for identifying novel radiopharmaceutical candidates targeting DAT.

3.5. Virtual Screening and Docking Experiment

The selected compounds were evaluated in the model to predict their LogK_d value, and the docking experiment was used to determine the binding mode of the ligand to the dopamine transporter pocket. The predicted LogK_d values, obtained with the computational model, and the binding affinities achieved in the docking experiment are shown in

Table 5. Predicted values of LogK_d, calculated binding free energies (kcal/mol), RMSD (Å), and number of H-bond interactions of the first-rank Autodock Vina poses for the complexes studied.

Molecules	Predicted LogKd	ΔGbinding (kcal/mol)	RMSD (Å) (Rank)	N° H-Bond
Modafinil (001)	2.386	−6.669	2.187 (5)	0
Phentermine (002)	3.538	−5.278	2.449 (2)	1
Procaine (003)	2.713	−6.452	2.227 (1)	1
Phenmetrazine (004)	4.024	−5.336	3.063 (2)	0
Diethylpropion (005)	3.517	−5.610	2.814 (3)	1
Benztropine (006)	2.449	−6.417	1.568 (2)	2
Fencamfamine (007)	3.405	−5.993	3.070 (1)	1
Sibutramine (008)	4.053	−6.012	2.795 (2)	1
Ioflupane (009)	3.088	−7.303	1.127 (1)	2
Armodafinil (010)	2.386	−6.436	2.838 (2)	2
Pseudoephedrine (011)	4.024	−5.742	2.332 (2)	1

In bold are highlighted the best results.

.

Figure 8 shows that the docked structures fit acceptably with the available L-dopamine X-ray crystal structures; all inhibitors were properly oriented at the active center of the dopamine transporter (Figure 8). To validate the docking experiments, we redocked dopamine with its receptor using the same methodology as with the other compounds. The RMSD obtained from the redocking of dopamine (reference ligand) was 0.85 Å, which is within the acceptable range (<2.0 Å) according to the reviewed literature [42]. This result confirms that the docking parameters we used in this work accurately reproduce the crystallographic pose found in the Protein Data Bank (PDB: 8Y2D) [43]. The best poses were conserved and were those that had a ∆G ≤ −6.0 kcal/mol, which is the common threshold for strong protein–ligand interactions [77].

4. Discussion

The development of new computational models demands a suitable goodness-of-fit, robustness, applicability domain, and predictability. The model developed in this study showed an R² value of 0.755, indicating a suitable fit for modeling DAT binding affinity. This means that the model explains more than 75% of the experimental variance. There was no overfitting in the model, thus representing a good fit with a minimum number of descriptors, because of the low value of the LOF parameter (0.195) and an R²_adj = 0.729; the LOF is used to penalize the addition of descriptors in the model equation and should be as low as possible, while the R²_adj value should be as similar as possible to the R² value, which the developed model complied with. The correlation between the descriptors of the model was low since the K_XX value (0.236) was small, indicating that there was no redundant information in the selected descriptors. In addition, the correlation between the descriptors and the response variable was appropriate, according to the ΔK parameter (0.076), with a small error in the calculations of both training and prediction estimates (RMSE_tr = 0.358; MAE_tr = 0.276; s = 0.381). Figure 4 shows the values predicted by the model equation against the experimental LogK_d values for both training and prediction series. As can be seen, most of the points are close to the line, and three compounds with atypical behavior in comparison to the rest of the molecules in the database were detected as outliers and not used to develop the model. Table 2 shows the predicted values for the compounds in the training set and the predictions in the LOO experiments as well as the residuals with respect to experimental values and leverage values (which were used to establish the applicability domain). The predictions are quite similar to the experimental values for most compounds. The molecular descriptors included in the model can be seen in

Table 6. Molecular descriptors selected in the obtained model.

Descriptor	Structural Feature Described	Descriptor Block
O-058	Hydrogen bonding capacity	(Atom-centered fragments) = O
CATS2D_03_AL	Lipophilicity	CATS2D Acceptor-Lipophilic at lag 03
CATS2D_08_LL	Lipophilicity	CATS2D Lipophilic-Lipophilic at lag 08
B08[C-N]	Presence/absence of C-N at topological distance 8	2D Atom pairs

with the general interpretation of the molecular feature that describes each descriptor and the descriptor block to which they belong in the software.

A critical aspect in QSAR studies is the definition of the AD of the models. Next, we used two approaches to establish the chemical space described by our model. First, we used a Williams graph (Figure 5A): the AD is defined as the area at the left of a leverage threshold h* = 0.3571 and within ±3 standard deviations. As shown in the figure, most compounds are within the applicability domain of the model. There is only one compound (49, belonging to the training set) with a leverage value (h = 0.2327) greater than the critical leverage (h*), although it shows the SD value within the limits; therefore, the prediction should be considered with caution. In addition, we used another approach available in QSARINS software, the Insubria graph (Figure 5B), which plots the leverage values against the predicted responses for each compound. The results agree with the previous approach, identifying the same compound outside the AD with a leverage value greater than the cutoff value, but in this approach two compounds in the prediction set (36 and 56) present LogK_d values that were slightly overestimated; see Table 3 for details.

To evaluate the robustness and stability of the model, several internal validation strategies were performed. First, an LOO cross-validation technique, which iteratively excludes a compound of the dataset: The variance explained in the prediction by LOO was Q²_LOO = 0.680, with a small error in the predictions (RMSE_cv = 0.409 and MAE_cv = 0.316); these results are shown in Figure 6A. Second, an internal validation by LMO was performed, which was developed by leaving out 30% of the dataset to study the behavior of our model. The value obtained for this experiment (Q²_LMO = 0.648) was like the value previously achieved in LOO, although it is important to notice that this technique gives better results with larger databases. The Q²_LMO values (red circles) were similar to each other and distributed around the Q² of the model (blue circle), corroborating its good fit and stability (see Figure 6B). In the Y-scrambling experiment, the absence of change in correlation for the model was checked. The expected behavior is that both values of R²_Y-scr and Q²_Y-scr (for each iteration) should differ appreciably from the R² and Q² values of the model. Both values of R² and Q² for every iteration are shown in Figure 6 (yellow and red circles, respectively), and their averages were R²_Y-scr = 0.098 and Q²_Y-scr = −0.174, quite far from the respective values of these parameters for the model (light and dark blue for R² and Q², respectively), confirming its validity.

To evaluate the model’s ability to predict new compounds, an external validation was also performed. The procedure was carried out by applying the model equation developed with the training set to the set of external predictions. Based on the results with the external set, we can say that the model has a good predictive power (R²_ext = 0.709), with lower value of 0.339 for RMSE_ext. As can be seen in Figure 5A, all compounds in the prediction set are within the applicability domain of the model. Table 3 shows the predicted values for the compounds in the prediction set and the residuals with respect to experimental values and the leverage values. The predicted values are quite similar to the experimental values, demonstrating the predictive power of the developed model.

The validated model was then used to predict the affinity for DAT of new molecules of interest. For this, we selected a set of eleven compounds with different actions over the CNS; among the activities of these compounds, we can find psychostimulant, centrally acting sympathomimetic agent, anorectic psychostimulant, centrally acting anticholinergic, a radiopharmaceutical used to diagnose Parkinsonism, nasal decongestant for systemic use, and psychedelic. This feature ensures that they can reach the DAT and makes the evaluation of these compounds in virtual screening more interesting. This group of compounds was evaluated with the computational model and assessed by docking to estimate the binding affinities for the DAT.

Among the compounds selected for virtual screening, ¹²³I-Ioflupane (one of the most widely used DAT radiotracers) was chosen, which achieved a good prediction for the LogK_d value of 3.088; taking this as a point of reference, we highlighted the compounds with affinity for DAT potentially better than ¹²³I-Ioflupane. We found that Modafinil and Armodafinil seem to be the better compounds showing the lowest LogK_d value of 2.386, followed by Benztropine with 2.449 and Procaine with a value of 2.713, slightly lower than the 3.088 shown by ¹²³I-Ioflupane. The results obtained with the QSAR model fully agree with those obtained in the docking experiments, which are described below.

The results of the docking experiments are shown in Table 5. As can be seen, all the ligands present binding energies (ΔG_binding) between −5.278 and −7.303 kcal/mol, with the most negative energy corresponding to the complex formed between the ligand Ioflupane (009) and the dopamine transporter, with a binding free energy of −7.303 kcal/mol. The stability of this complex is based on the two H-bond interactions with Asp68 and Ser72, in addition to the π-π stacking interaction with Tyr355 (Sandwich Type) (Figure 9C).

The second binding energy corresponded to the complex formed between Modafinil (001) and the dopamine receptor, with a binding free energy of −6.669 kcal/mol. The stability of this complex is based on two π-π stacking interactions shared with Phe332 (T-Shaped and Sandwich Type) (Figure 9A). The third and fourth most negative binding free energies were found between the complexes formed between Procaine (003) (−6.452 kcal/mol) and Armodafinil (010) (−6.436 kcal/mol). The stability of the Procaine–dopamine transporter complex is based on one H-bond interaction with Asp79, and two π-π stacking interactions with Phe320 (Sandwich Type) and Phe326 (T-Shaped Type) (Figure 9B). The stability of the Armodafinil–dopamine transporter complex is based on two hydrogen bond (H-Bond) interactions shared with Leu459 (Figure 9D), in addition to the π-π stacking interaction with Phe462 (T-Shaped Type) (Figure 9D). The least stable complex of all those subjected to docking experiments was the one formed between the ligand Phenmetrazine (004) and the dopamine transporter, as well as the least stable complex, corresponding to Phentermine (002) and the dopamine transporter. These two ligands were oriented away from the active center of the protein with loss of non-covalent interactions (Figure 10).

Below, we present a detailed comparison of the results of our docking experiments (ΔGbinding and RMSD) with recently published data from computational and experimental studies. This comparison serves to validate the consistency and reliability of our findings. For example, a recent study developed robust QSAR models on approximately 61 DAT inhibitors, including drugs similar to ours, and obtained ΔGbinding values very similar to ours. Phenmetrazine (L-004) in this study behaved very similarly to ours, with ΔGbinding ≈ −6.8 kJ/mol [78].

To support the validity of our data, we compared the docking results with recent experimental affinities in the scientific literature. Modafinil has been estimated to have a Ki ≈ 2600 nM (≈2.6 µM) per human DAT, which is in agreement with our finding of ΔGbinding ≈ −5.99 kJ/mol, reflecting a low affinity for the dopamine receptor, a result that is in agreement with our study. Furthermore, an analogous relationship between experimental affinity in this study and the binding energies observed in our set of molecules (with several ligands in the range −7.0 to −7.5 kJ/mol with RMSD < 2 Å) has been found, corroborating the predictive power of our docking experiments [79].

Molecular dynamics simulations provide us with information about the behavior of ligand–protein complexes over time at the atomic level [70,80,81,82]. In our case, they also allow us to determine the stability of the systems and whether the interactions found in the docking experiments are maintained over time. To this end, we analyzed several parameters obtained from the trajectory, which will be discussed below. One of the important parameters when analyzing the stability of ligand–protein complexes over time is the RMSD [47,83,84]. The behavior of this variable over time for all the complexes studied is shown in Figure 11. As can be seen in Figure 2, all complexes subjected to molecular dynamics simulations maintained low RMSD values (<2 Å), indicating that these complexes remained stable throughout the simulation period. This result is consistent with that obtained from the docking experiments. The lowest RMSD values were found in the complex formed by Benztropine and the dopamine receptor, meaning that it was the most stable complex over time. This complex had the most negative energy obtained in the docking experiments, coinciding with the results obtained from the docking experiments and the molecular dynamics experiments; therefore, we can expect that this molecule could be a good candidate as a radiotracer for Parkinson’s disease diagnosis.

One of the parameters that could explain the stability of compounds over time is hydrogen bond interactions (Figure 12). As can be seen in Figure 12, the number of hydrogen bond interactions during the trajectory was low. The highest number of hydrogen bond interactions was found in the complex formed by Procaine and the dopamine transporter. Given the structure of this molecule (Figure 3), there is only one possibility of hydrogen bond interaction as the hydrogen acceptor: the carbonyl oxygen of this ligand, which interacts strongly and stably with Asp79 of the dopamine transporter with an occupancy of 88.81% (

Table 7. H-bond occupancies analysis during 100 ns of simulation time.

Complexes	H-Bond Interaction	Occupancy (%)
Diethylpropion-8Y2D	ASP421-Side	60.98%
	PHE320-Main	15.76%
Procaine-8Y2D	ASP79-Side	88.81%
	VAL78-Main	15.16%
Benztropine-8Y2D	LEU459-Main	0.08%
	ASP231-Side	0.04%

). This interaction could give the complex strong stability during the simulation time.

The second complex with stable hydrogen bond interactions over time was found in the complex formed by diethylpropion with the dopamine transporter. This complex maintained strong interactions with Asp421 with an occupancy of 60.98% (Table 7), which is considered a stable interaction over time. This behavior could explain the low RMSD values found during the 100 ns trajectory.

In the case of the complex formed by Benztropine and the dopamine transporter, no stable hydrogen bond interactions were found over time. Given the structure of this molecule (Figure 11), it has little chance of interacting by hydrogen bond with the amino acids of the dopamine transporter, because it only has one hydrogen acceptor group and it is an oxygen of the ether group within Benztropine. The stability of this complex found in the RMSD parameter could be because of another type of interaction such as π-π staking, Van del Waals interactions, or solvation type.

Another parameter usually used to analyze ligand–protein complexes is the radius of gyration. This variable provides information about the degree of compaction of a complex during the simulation time. As can be seen in Figure 12, the two most compact complexes were those formed by Procaine and Benztropine, with radii of gyration values below 4 Å. This result is consistent with the RMSD results and with the results obtained from the docking experiments, so we can conclude that these two compounds could be good markers for Parkinson’s disease.

5. Conclusions

In the present study, a QSAR-MLR model was developed by using molecular descriptors calculated with the Dragon software, which adequately predicts the binding affinity for DAT. The model developed with the QSARINS software fulfilled all regulatory principles established by the OECD. The robustness of the model was tested through several internal validation techniques (LOO, LMO, and Y-scrambling), and its predictability was tested using a set of external predictions (external validation). The developed model was used to perform a virtual screening of compounds with different actions on the CNS. The model predicts four compounds with LogK_d values lower than the ¹²³I-Ioflupane value; these compounds should present good binding affinities for DAT. The docking experiments corroborate the results of the model prediction and give information about the main interactions of these compounds with the active site. The combination of both strategies strengthens the probabilities of identifying new compounds with potential affinity for DAT. In conclusion, the proposed computational tool is efficient in the identification of novel DAT radiotracers.