A‑iMPO: A Design-Time Prediction Toolbox for Blood Brain Barrier Permeation Using an Integrated and Interpretable Multiparameter Optimization Framework

Abstract

Design time prioritization of central nervous system (CNS) drug candidates remains a challenge due to the restrictive nature of the blood–brain barrier (BBB) governing brain exposure. Although various multiparameter optimization (MPO) strategies have guided CNS medicinal chemistry for over a decade, existing frameworks rely heavily on heuristic cutoffs and offer limited interpretability across chemically diverse scaffolds. Here, we introduce a next generation CNS-MPO frameworkAragen-iMPO, (A-iMPO)which was developed using 5,129 curated compounds through an integrated workflow combining explainable machine learning, rigorous descriptor selection, and low-dimensional discriminant mapping. This yielded six chemically intuitive features capturing polarity, ionization, size, rigidity, and electronic distribution. The resulting score provides a transparent discriminant function enabling direct compound prioritization through a simple threshold rule. Across internal and external validation sets, A-iMPO outperformed established CNS-focused scoring methods while maintaining mechanistic interpretability, providing a practical and design ready tool for CNS drug discovery.

The development of CNS therapeutics is constrained by the blood–brain barrier, a dynamic neurovascular interface that tightly regulates molecular entry into the brain. Despite advances in target biology and ligand design, inadequate brain exposure remains a major contributor to attrition in CNS drug discovery programs. Only a small fraction of drug-like molecules penetrates the BBB effectively. Failure rates exceed 98% for small molecules with even poorer outcomes for biologics and emerging modalities.

Over time, several strategies have been explored to address this limitation. Drug delivery platforms, including carrier-mediated transport, nanoparticle systems, and exosome-based delivery, have shown some promise. However, their clinical translation has been inconsistent, often due to safety concerns or limited efficiency. , Due to these limitations, rational design of small molecules capable of passive BBB permeation continues to be a preferred and practical strategy in CNS drug discovery programs.

Within this paradigm, CNS multiparameter optimization (CNS-MPO) frameworks have been widely used. These approaches typically integrate physicochemical properties, such as lipophilicity, polar surface area, and ionization, to guide compound design. However, most MPO methods rely on heuristic weighting schemes and fixed thresholds, which reduce flexibility and limit generalizability across chemically diverse scaffolds. Parallel to this, machine learning approaches have emerged as powerful tools for predicting BBB permeability. Algorithms such as support vector machines, random forests, and neural networks have been applied using molecular descriptors and fingerprints. − These models although often achieve good predictive accuracy, their practical application remains limited due to issues such as data set bias, overfitting, lack of interpretability, and dependence on descriptor selection. − These limitations highlight an unmet need to develop a framework that combines the interpretability of MPO approaches with the predictive rigor of machine learning.

In this work, we introduce “A-iMPO”, an interpretable scoring model derived from a hybrid workflow integrating explainable ML and linear discriminant analysis (LDA). By coupling statistical robustness with chemical intuition, A-iMPO provides a simple yet mechanistically meaningful scoring function that can be directly applied during medicinal chemistry optimization.

The objective of the present work was to construct a predictive model using a minimal descriptor set that retains interpretability while maintaining strong classification performance. A curated data set of 5,129 compounds balanced between BBB-permeant (Class 1) and nonpermeant (Class 0) molecules were taken for building the foundational models (Figure S1). Multiple machine learning classifiers like logistic regression, random forest, gradient boosting, support vector machines, and others were primarily trained using 1,444 molecular descriptors (Table S1). All models using this set achieved high discriminative performance of AUROC ∼ 0.9 (Figure ). To understand the chemical space coverage between the training, test and the external validation set Bemis-Murcko scaffold analysis was undertaken, and scaffold overlap along with the Tanimoto distribution was analyzed. These results demonstrate that more than half of the scaffolds in both validation sets are absent from the training set, indicating that the evaluation was conducted on compounds with substantial scaffold diversity rather than on closely related analogues (Figure S2).

1.

Heatmap of AUROC value of all the modeling workflow with various descriptor sets.

A structured feature reduction workflow was implemented to derive a compact, yet information-rich descriptor set. Initially, SHAP-based importance ranking across multiple models was applied, which identified 65 unique descriptors (Table S2) from the top 20 features in each model (Figure S3). Models retrained on this 65-descriptor set retained near-optimal AUROC values (Figure ), indicating that these features captured the most influential chemical determinants of BBB permeability.

This set was further refined using the Boruta algorithm, which removed redundant and noninformative features in high-dimensional data sets. This step yielded 60 descriptors (Table S3) without a loss of predictive performance. Finally, to enhance interpretability, L1-regularized LASSO regression was subsequently applied which shrinks the negligible coefficients to zero -producing a final subset of 31 descriptors (Table S4).

Across all reduction stages, the model performance remained consistently high, demonstrating both robustness and stability. Standard metrics, including accuracy, sensitivity, precision, and F1-score, were monitored throughout to ensure that descriptor pruning did not compromise classification quality (Table S5).

The initial set of 31 molecular descriptors selected by L1-LASSO was used to perform iterative LDA models. Combinations of 5, 6, and 7 descriptors were systematically evaluated to identify the most relevant features for the BBB permeation. A total of 169,911 five-descriptor models, 736,281 six-descriptor models, and 2,629,575 seven-descriptor models were generated and analyzed.

Top ten models from each subset were shortlisted based on predictive performance (Table S6). Comparison across descriptor iterations revealed that six-descriptor models consistently provided optimal predictive performance, balancing model parsimony and accuracy.

The best-performing model with six descriptors is being represented in eq and was subjected to further validation using the external cohort to evaluate its robustness and generalizability (Table ).

1. Descriptors’ Impact in BBB Permeation.

Descriptor	Definition in structural terms	Effect on A-iMPO
TopoPSA	Total surface area contributed by N and O atoms representing the hydrogen-bonding pattern.	Negative (strongest term)
MWC7	Number of distinct 7-bond walks through the molecular graph: how much fused, branched, ring-rich structure sits within a 7-bond radius. High for compact polycyclic cores, low for long flexible chains	Positive
GATS 3s	Geary autocorrelation at 3 bonds, weighted by atomic intrinsic state: how dissimilar atoms are to their 3-bond neighbors in electronic character. High = electronically smooth over medium range	Positive
nAcid	Count of acidic functional groupsa proxy for anionic ionization at pH 7.4	Negative
AATSC 5p	Mean centered autocorrelation at 5 bonds, weighted by atomic polarizability: whether polarizable atoms are paired at a 5-bond separation. Reports on how polarizability is distributed, not its total	Positive (weakest term)
AATSC1c	Mean centered autocorrelation at 1 bond, weighted by partial charge: the size of charge separation across directly bonded atom pairslocal bond polarity	Positive

LDA Model:

$$d(x)=-0.01+0.41\times \mathrm{G}\mathrm{A}\mathrm{T}\mathrm{S}3\mathrm{s}-1.62\times \mathrm{T}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{P}\mathrm{S}\mathrm{A}-0.23\times n\mathrm{A}\mathrm{c}\mathrm{i}\mathrm{d}+0.25\times \mathrm{A}\mathrm{A}\mathrm{T}\mathrm{S}\mathrm{C}\mathrm{1}\mathrm{c}+0.07\times \mathrm{A}\mathrm{A}\mathrm{T}\mathrm{S}\mathrm{C}5\mathrm{p}+0.59\times \mathrm{M}\mathrm{W}\mathrm{C}7$$ 1

where the d(x) is the discriminant score or decision threshold.

The above LDA model equation was then transformed to a discriminant function (DF) or probability of identification of class 1 compounds.

$$DF=1/(1+\exp (-d(x))$$ 2

The resultant discriminant function (eq ) derived from a consensus and selected interpretable descriptor set, is hereafter termed as Aragen-iMPO (A-iMPO) score. The derived A-iMPO score is a single, standardized design time metric that reflects the likelihood of BBB permeability. Compounds with an A-iMPO score >0.5 are predicted to be BBB permeant, whereas those scoring ≤ 0.5 are classified as nonpermeant. This interpretation allows the model to be applied as a transparent and auditable multiparameter optimization function, equivalent to traditional CNS-MPO frameworks, while enhancing the interpretability and stability conferred by machine learning driven feature selection.

Our integrated model’s A-iMPO scoring function demonstrates strong statistical significance and robust predictive performance (Table ). Wilks’ lambda (λ) value <0.5 indicates substantial separation between two classes of compounds. reflecting the effectiveness of the selected descriptors in capturing discriminative information. This is further supported by the Mahalanobis distance of 1.47, suggesting meaningful multivariate separation in descriptor space. Model’s overall significance is confirmed by a highly significant chi-square statistic (χ² = 3712.58, df = 6, p < 0.0001), indicating that the discriminant function is not due to random variation.

2. Statistical Evaluation Parameters of LDA Model.

Metric	Value	Metric	Value
Wilks Lambda (λ)	0.40	MCC (Train)	0.65
Mahalanobis Distance	1.47	MCC (Test)	0.65
Chi-square (χ2)	3712.58	AUROC (Train)	0.88
df	6.00	AUROC (Test)	0.86
p-value	0.00	ROCED	0.24
		ROCFIT	0.86

The model showed consistent predictive performance across the training and test sets with a Matthews correlation coefficient (MCC) of 0.65 for both. This consistency indicates good generalizability and a limited amount of overfitting. Ablation analysis further showed that removing individual descriptors reduced model performance, confirming the importance of all six descriptors (Figure S4).

The model also exhibits strong discriminative enrichment ability, with AUROC values of 0.88 for the training set and 0.86 for the test set (Figure A,B, respectively). Additional ROC-based metrics reinforce the quality of classification. The ROC enrichment descriptor (ROCED = 0.24) indicates effective early recognition performance, while the Receiver Operating Characteristic Fit (ROCFIT) value of 0.86 confirms a good overall fit of the ROC curve. Further evaluation of the external cohort yielded an AUROC value of 0.79, indicating acceptable predictive performance of the model (Figure C). Collectively, these results demonstrate that the LDA derived A-iMPO model achieves a favorable balance between statistical significance, discrimination of power, and predictive stability.

2.

ROC curves (A–C for the training, test and external sets) and density distribution diagram (D–F for the training, test and external sets) distinguishing Class 0 and Class 1 compounds. The vertical dashed line denotes the selected decision threshold, which effectively separates the two classes.

The density distributions diagram for the two classes of compounds further supports the robustness of the model. Figure D–F shows that Class 1 and Class 0 compounds occupy distinct probability regions, with limited overlap between the two distributions. The presence of a small overlap region depicts compounds with intermediate prediction confidence. The misclassified compounds are represented as cross signed in red color in Figure .

As additional validation matrices of our model performance, a confusion matrix and its derived features were generated which demonstrated that the model correctly classifies most compounds across all three evaluation settings (Figure S5). The consistent ratio of correct to incorrect classifications suggest a stable model performance despite changes in training, test and external sets (Table S7). Model’s performance values across validation settings such as accuracy, precision, specificity and F1-score, ranged from 0.69 to 0.90 (Figure S6). While the training and test sets show comparably strong performance, modest decrease is observed for the external validation set (0.69–0.78) due to its larger chemical space heterogeneity. Importantly, precision remains stable across all data sets (0.72–0.77), supporting reproducible and robust classification under increasing stringent conditions.

To further evaluate the predictive performance of the selected six descriptors, we developed all of the traditional ML classification models using this optimized descriptor set. While all models exhibited comparable predictive abilities for both the training and test sets, the scoring A-iMPO model demonstrated superior performance in classifying the external cohort (Table S8).

Our A-iMPO model exhibited stable and reproducible performance across stratified 10-fold cross-validation (Table S9), achieving an average training accuracy of 0.82 with high recall (0.92) and an ROC-AUC of 0.87. These results indicate a strong discriminative capability of our model. Performance variability across individual folds was minimal, demonstrating robustness to data partitioning. When evaluated on the external validation set, the model maintained consistent predictive behavior, with an average accuracy of 0.73, recall of 0.77, and ROC-AUC of approximately 0.78–0.79 across all folds. Although a moderate decrease in performance was observed relative to internal validation, the preservation of sensitivity and the limited fold-to-fold variation suggest good generalization and limited overfitting. Overall, these results confirm the reliability of the A-iMPO model and support its effectiveness in the applied cross-validation with good data balance.

After establishing the predictive performance of the model through multiple validation approaches, we next examined the contribution of individual features to the compound classes. To understand this, we generated a Variable Importance Plot (VIP) (Figure ) to rank order the influence of descriptors relevant for distinguishing between BBB permeant and nonpermeant compounds.

3.

VIP plot of the features importance in the model development.

Figure presents the VIP scores, which illustrate the relative contribution of each molecular descriptor toward discriminating between the two classes of compounds. Among all of the evaluated features, TopoPSA exhibits the highest importance, indicating its dominant role in driving class separation. In contrast, AATSC 5p shows a minimal influence relative to the other descriptors. Descriptors such as MWC7, GATS 3s, AATSC1c, and nAcid display moderate contributions, suggesting that they provide complementary support to the classification process. These findings are consistent with previous BBB permeability studies, which identify TPSA as a primary determinant of BBB penetration emphasizing the inherently multivariate nature of CNS drug likeness. Similar machine learning analysis also highlights properties like polar surface area, molecular weight, and lipophilicity as dominant discriminators, supporting the descriptor importance and interrelationships observed in our VIP. Overall, while our model incorporates information from multiple descriptors, the decision boundary is primarily shaped by a small subset of highly informative features, reflecting a focused, yet multivariate structure in the learned classification pattern.

A deeper examination of feature importance at the atomic level was carried out using RDKit’s “GetSimilarityMap-FromWeights” function, yielding insights consistent with attribution-based molecular similarity and visualization approaches. In addition to evaluating the directional contribution of descriptors in the A-iMPO model, atom-level influences were visualized for representative BBB permeant and nonpermeant compounds whose contribution of each descriptor are mentioned in Table .

Among all, topological polar surface area (TopoPSA) was identified as a dominant descriptor. It is well established that high topological polar surface area limits BBB penetration, whereas lower values favor passive diffusion across the lipid bilayer of brain endothelial cells. − In our analysis, compounds such as kanamycin, with elevated TopoPSA, show strong unfavorable contributions and poor permeability, while low-TopoPSA molecules like amitriptyline exhibit enhanced BBB permeability (Figure A,B).

4.

Atomic contribution of the descriptors used for model development on selected compounds.

Molecular walk count of order 7 (MWC7) reflects the structural compactness and rigidity of molecules. BBB permeant compounds typically fall within an optimal complexity window that supports membrane residence and transcellular diffusion. For example, nalmefene displays strong positive contributions consistent with its rigid scaffold, whereas flexible molecules such as hydroxycarbamide show reduced permeability (Figure C,D).

GATS 3s is a 2D autocorrelation descriptor that captures medium-range electronic variation across the molecular graph. Higher values are associated with balanced electronic distribution and limited surface polarity, both favorable for BBB permeation. In contrast, lower values, as seen in peripherally acting compounds like cinacalcet, correlate with reduced BBB penetration in contrast to phentermine (Figure E,F). These observations align with prior findings emphasizing the importance of internal electronic heterogeneity in BBB transport.

nAcid provides a clear mechanistic interpretation of this. Molecules with multiple acidic groups are typically ionized at physiological pH and exhibit poor passive permeability unless active transport is involved. Accordingly, an increase in acidic functionality correlates with diminished BBB penetration in the model.

Average autocorrelation of topological structure at lag 5, weighted by p (AATSC 5p), indicates favorable electronic polarization patterns. Compounds such as dalfampridine display localized positive contributions consistent with passive permeability, whereas molecules with heterogeneous polarizability distributions, like triclofos, exhibit reduced predicted transport (Figure G,H).

Average centered autocorrelation, lag 1, charge weighted (AATSC1c) highlights the importance of balanced charge distribution. Moderate values support BBB permeability by maintaining a balance between localized polarity for binding and overall lipophilicity for membrane passage, whereas excessive charge separation negatively impacts BBB transit (Figure I,J).

Collectively, these atomic-level interpretations reinforce the mechanistic relevance of the selected descriptors in governing the BBB permeability.

After identifying the most influential descriptors through VIP analysis, pairwise relationships were examined using scatterplot matrices (pair plots) for both training (Figure ) and test (Figure ) data sets. These plots display marginal distributions along the diagonal and pairwise correlations in off-diagonal panels, stratified by class 0 (blue) and class 1 (orange). Several descriptors show clear class-dependent shifts, while others exhibit a partial overlap. Although some feature pairs demonstrate separation, no single descriptor achieved complete discrimination. Overall, this analysis highlights the fact that reliable classification arises from combined multivariate effects rather than individual descriptor pairs alone.

5.

Pair plot of the features in training set.

6.

Pair plot of the features in test set.

It is noteworthy that various workers had previously reported different scoring metrics based on MPO format to predict the BBB permeation of small molecules, CNS MPO-PET, and the BBB score. For a comparative assessment, the performance of the A-iMPO discriminant function was evaluated alongside these established metrics by using a common benchmark (Table ). Additionally, to ensure a matched comparison, we recalculated the CNS-MPO and BBB Score using the same data set employed for the development and validation of the A-iMPO model. The resulting performance metrics demonstrate that A-iMPO outperformed both reference methods across most classification metrics (Table and Table S10). Hence, A-iMPO demonstrates improved or comparable discriminatory power relative to these traditional MPO-based scores.

3. Comparison of A-iMPO Scoring Function with Reported Scoring Functions.

Model	Accuracy	Sensitivity	Specificity	Positive predictive value	Negative predictive value	AUC
CNS MPO score	0.57	0.6	0.48	0.76	0.3	0.54
CNS MPO PET score	0.49	0.44	0.63	0.78	0.27	0.54
BBB score	0.66	0.67	0.65	0.84	0.41	0.66
Random forest model	0.83	0.88	0.75	0.87	0.79	0.88
A-iMPO	0.82	0.91	0.72	0.76	0.89	0.88

4. Comparative Analysis of A-iMPO with CNS_MPO and BBB Score with Data Set of A-iMPO.

Metric	CNS_MPO	BBB_Score	A-iMPO
Accuracy	0.68	0.71	0.82
Precision	0.65	0.79	0.77
Sensitivity	0.77	0.57	0.91
Specificity	0.58	0.85	0.72
F1 Score	0.70	0.66	0.83
MCC	0.36	0.44	0.65

Spielvogel et al. (2025) reported a random forest (RF) model exhibiting its superior performance over the other scoring function (Table ). However, our parsimonious, mechanism-aligned model (A-iMPO) provides a direct and fundamentally interpretable BBB permeation score (A-iMPO > 0.5) that matches RF level accuracy while removing the need for any indirect model approximation explanations. Unlike the RF model, which is trained on a small PET-anchored data set (Table S11), our model maintains performance on a larger cohort-emphasizing broader chemical space generalization. Furthermore, when benchmarking A-iMPO against this RF model using an additional independent external data set of 741 compounds, the model achieved excellent performance on its training and test sets, but a marked decline was observed on the external data set, suggesting limited transferability beyond the original chemical space. In contrast, A-iMPO demonstrated more stable performance across data sets and outperformed the RF model on external validation with an accuracy of 0.74 vs 0.67 and AUC of 0.79 vs 0.73 (Tables S8 and S12). Given that external validation is generally considered the most stringent measure of prospective performance, these findings support the superior generalizability of the A-iMPO. In addition, the model provides a continuous optimization score rather than a binary classification output, enhancing its utility as a medicinal chemistry design tool.

The identified descriptor set aligns well with the established principles governing BBB permeation via passive diffusion. This process is driven by a delicate balance of free energy contributions, including the desolvation of polar groups during membrane entry, favorable partitioning into the lipid bilayer, and subsequent diffusion into the cytosol along thermodynamic gradients. Within this framework, the negative contribution of TopoPSA reflects increased desolvation penalties associated with polar hydrogen-bonding functionalities. Similarly, the nAcid descriptor captures the impact of ionization, where strongly acidic compounds remain largely charged at physiological pH and exhibit poor BBB permeability with a potential efflux susceptibility.

In contrast, the positive role of MWC7 highlights the importance of optimal molecular size and rigidity, consistent with CNS-active scaffolds that favor membrane residence without excessive polarity. The autocorrelation descriptors (GATS 3s, AATSC1c, and AATSC 5p) further describe the role of electronic distribution, charge balance, and polarizability, supporting target engagement while minimizing exposed polarity, which is detrimental to membrane transit.

From an applicability and translational perspective, the model is expected to correlate more strongly with passive diffusion-based BBB assays (e.g., PAMPA-BBB) than with transporter-competent systems, particularly for strong efflux substrates. Principal component analysis showed that the test and external validation sets (from literature and Cortellis database) were largely encompassed within the training-set chemical space (Figure S7). Consistent with this, CNS compounds from the Cortellis clinical trial database resided within the applicability domain and were predicted with 90% accuracy (Table S11), while a mixed BBB-penetrant/nonpenetrant external set achieved 74% accuracy. These findings support the model’s applicability across clinically relevant CNS chemical space and a broader chemical class than conventionally reported for CNS drugs. In contrast, performance decreased to 56% on a PET-anchored data set. Evaluation against the data set reported by Spielvogel et al. showed satisfactory classification of 110 CNS/non-CNS compounds but lower performance for 44 efflux substrates (Figure ; Table S12). Therefore, A-iMPO predictions should be interpreted cautiously for underrepresented chemotypes, including zwitterions, macrocycles, and compounds dominated by active transport. Overall, an A-iMPO score of >0.5 provides a practical prioritization metric for early CNS lead optimization.

7.

Comparison of statistical parameters with the existing model.

We also highlight that the model is based on descriptors which were hitherto not reported for CNS drugs and goes beyond the conventional parameters such as MW, logP, and rotatable bonds. Since there are reported molecules with higher molecular weight crossing the BBB, our applicability domain would be encompassing primarily on the unreported descriptors.

Our work reports an integrated and interpretable framework for assessing BBB permeability that bridges medicinal chemistry with neurovascular biology, supporting CNS drug discovery decision-making. We identified six descriptors (TopoPSA, MWC7, GATS 3s, nAcid, AATSC 5p, and AATSC1c) by integrating a high-dimensional ML and low-dimensional LDA aligned with CNS-MPO principles. These descriptors collectively capture the key physicochemical and topological features governing BBB permeation, including polarity, ionization, molecular size, and charge distribution, and provide a chemically intuitive basis for compound prioritization.

The resulting A-iMPO discriminant score translates multivariate descriptor information into a simple metric suitable for application in CNS medicinal-chemistry workflows. As compared to earlier published BBB models, our approach delivers three practical advantages: a) design-time interpretability, enabling rational assessment of permeability effects during structure optimization; b) descriptor stability, supported by consensus selection across heterogeneous classifiers; and c) translational readiness, achieved through a single auditable threshold-based scoring function (A-iMPO > 0.5).

While the framework primarily reflects passive BBB permeability, further integration with transporter-based assays and orthogonal models can enable a better assessment of CNS exposure for chemotypes influenced by active transport or efflux. Nevertheless, the successful real-world implementation of A-iMPO will depend on ensuring the safety, robustness, and accountability of AI models. Future integration into clinical workflows should be guided by established recommendations and emerging clinical AI governance frameworks addressing safety and legal responsibility.

Accordingly, A-iMPO is positioned as a prioritization tool rather than a universal predictor of BBB transport within CNS chemical space. By converting high-dimensional ML features into an MPO-like score, A-iMPO bridges the gap between predictive modeling with neuroscience-based medicinal chemistry actionability.

Glossary

Abbreviations

CNS

Central nervous system

BBB

Blood–brain barrier

A-iMPO

Aragen-iMPO

MPO

Multiparameter optimization

ML

Machine learning

SVM

Support Vector Machine

LR

Logistic Regression

RF

Random Forest

GB

Gradient Boosting

KNN

k-Nearest Neighbors

DT

Decision Tree

NB

Naïve Bayes

SHAP

SHapley Additive exPlanations

AUROC

Area under the receiver operating characteristic curve

ROC

Receiver operating characteristic curve

LDA

Linear discriminant analysis

DF

Discriminant function

d (x)

Discriminant score or decision threshold

VIP

Variable Importance Plot

TopoPSA

Topological polar surface area

MWC7

Molecular Walk Count of order 7

nAcid

Number of acidic functional groups

AATSC 5p

Atomic polarizability is distributed across a molecule’s topology at a graph distance (“lag”) of 5 bonds

AATSC1c

Average Centered Autocorrelation, lag 1, charge weighted

All relevant data described in the manuscript are submitted as .

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.6c00347.

• Workflow diagram, similarity analysis of among training, test and external sets, selected features list using SHAP analysis, ablation analysis of the LDA descriptor, confusion matrix, statistical parameters, PCA analysis and methodology (PDF)

• Table S1: Descriptor’s values of selected 5129 compounds. Table S2: Features obtained from SHAP analysis. Table S3: Features obtained from Boruta Filtering. Table S4: Features obtained from L1-Lasso regression. Table S5: Model validations with features, SHAP analysis, Boruat Filtering, and L1-LASSO regression. Table S6: LDA model validations. Table S7: Observed and predicted class of training, test, and external sets. Table S8: Validation metrics of the models with selected 6 features. Table S9: 10-fold cross validation parameters. Table S10: Calculation of BBB-Score and CNS-MPO Score for total 5129 compounds. Table S11: Prediction of External set from clinical data and PET-anchored data. Table S12: Observed and predicted class of combined sets and compounds for analysis of domain applicability. (ZIP)

No safety hazard was encountered.

The authors declare no competing financial interest.