Multiscale Explainable Machine Learning Reveals Descriptor-Invariant Molecular Determinants of Small-Molecule PD-1/PD-L1 Inhibition

Abstract

The PD-1/PD-L1 immune checkpoint pathway is a major target in cancer immunotherapy; however, small-molecule inhibitor development remains challenging due to the hydrophobic, structurally shallow PD-L1 interface. We developed an explainable artificial intelligence (XAI)-based QSAR framework to identify determinants governing PD-1/PD-L1 inhibition. A data set of 844 compounds, represented using MACCS, PubChem, and Mordred descriptors, was modeled using multiple machine learning algorithms with eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) models achieving the highest predictive performance. SHapley Additive exPlanations (SHAP) analysis and scaffold enrichment revealed convergence across descriptors, highlighting nitrogen-rich heteroaromatic systems, sulfur-containing motifs, and fused aromatic scaffolds as key determinants of activity. Active compounds occupied a distinct physicochemical space characterized by low molecular weight (MW) and topological polar surface area (TPSA) < 85 Ų. Docking identified conserved interactions within the PD-L1 dimer interface, and molecular dynamics confirmed their stability under-near physiological conditions. Molecular Mechanics/Poisson–Boltzmann Surface Area (MM/PBSA) calculations indicated that binding is predominantly driven by van der Waals interactions and hydrophobic stabilization, providing mechanistic insight and design guidance for next-generation PD-L1 inhibitors.

The PD-1/PD-L1 immune checkpoint is a central regulator of immune homeostasis that limits excessive immune activation by suppressing T-cell function following PD-1 engagement. Although this mechanism maintains self-tolerance, many tumors exploit it by overexpressing PD-L1 through oncogenic and microenvironmental pathways, including MAPK, PI3K/AKT, HIF-1α, STAT3, and NF-κB, thereby promoting T-cell exhaustion, reducing cytokine production, and facilitating immune evasion. −

Immune checkpoint blockade with monoclonal antibodies, including nivolumab, pembrolizumab, atezolizumab, and durvalumab, has revolutionized cancer therapy. However, antibody-based treatments are limited by poor oral bioavailability, restricted tumor penetration, prolonged systemic exposure, high manufacturing costs, and immune-related adverse events. These limitations have stimulated the development of small-molecule PD-1/PD-L1 inhibitors, which offer improved tissue penetration, oral administration, tunable pharmacokinetics, and lower production costs. ,

Designing small-molecule inhibitors for PD-1/PD-L1 remains challenging because the interaction occurs across a flat, hydrophobic, and conformationally flexible protein–protein interface (PPI) with few well-defined binding pockets. , Ligand binding frequently induces PD-L1 dimerization and conformational rearrangements that are essential for inhibitory activity, complicating rational drug design despite the identification of transient druggable hotspots. Machine learning (ML), QSAR, docking, and molecular dynamics (MD) simulations have become indispensable tools for drug discovery, yet they are typically applied as independent or sequential approaches. ML-QSAR models often provide accurate predictions without mechanistic interpretation, whereas docking and MD alone cannot fully capture the dynamic nature of protein–protein interactions (PPI). Consequently, a gap remains between high-throughput prediction and atomistic understanding of inhibitors. A central, largely untested question in ML-driven drug discovery is whether molecular determinants of activity identified by one modeling approach are genuine chemical signals or artifacts of a particular descriptor choice or algorithm. Here, we present a unified explainable artificial intelligence (XAI) framework that integrates chemical space analysis, QSAR-ML modeling, SHAP interpretation, scaffold enrichment, docking, MD simulations, and MM/PBSA free-energy calculations within a single data-centric workflow. We hypothesized that PD-1/PD-L1 inhibitors share convergent physicochemical, structural, and energetic determinants that can be independently identified across distinct molecular descriptors and validated using complementary structure-based approaches; such convergence would provide strong evidence that the identified features reflect genuine structure–activity relationships rather than representation-specific artifacts. Unlike prior QSAR-docking-MD studies of PD-L1 inhibitors, which have typically applied these methods sequentially, using QSAR for activity prediction followed by docking or MD on a small subset of top hits, without a systematic framework linking ML-derived feature importance to structural mechanisms, our approach uses SHAP interpretation across independent descriptor spaces to generate specific, falsifiable structural hypotheses, which are then directly tested against docking, MD, and MM/PBSA energetics. To our knowledge, this is among the first studies demonstrating that XAI can bridge ligand-based and structure-based drug discovery for a challenging PPI target, enabling bidirectional validation between ligand-based and structure-based methodologies that links ML-derived hypotheses with atomistic simulations and binding energetics. PD-1/PD-L1 inhibitors were curated from ChEMBL based on IC₅₀ values. After data preprocessing, 844 unique molecules were analyzed. Compounds with pIC₅₀ ≥ 5.8 were classified as active (397) and those below this threshold, as inactive (447). Physicochemical profiling revealed distinct activity-associated chemical space (Table ). Active compounds exhibited significantly lower MW (370.11 vs 411.24 Da; p = 6.63 × 10^–15), TPSA (76.34 vs 90.72 Ų; p = 2.29 × 10^–18), HBA (4.67 vs 5.71; p = 4.49 × 10^–17), and HBD (1.34 vs 1.66; p = 1.81 × 10^–7) than inactive compounds (Figure A). Although slightly less lipophilic (LogP 3.14 vs 3.33; p = 1.58 × 10^–2), active molecules retained sufficient hydrophobicity to interact with the predominantly nonpolar PD-L1 interface (Figure B). They also possessed marginally more rotatable bonds (4.74 vs 4.44; p = 2.83 × 10^–2), suggesting that moderate conformational flexibility may facilitate binding within the induced PD-L1 pocket.

1. Exploratory Data Analysis (EDA) of Drug-Likeness Descriptors between Active (A1) and Inactive (A2) PD-1/PD-L1 Inhibitors .

	MW in g/mol		cLogP*		HBA		HBD		nRB		TPSA in Å
Descriptor	A1	A2	A1	A2	A1	A2	A1	A2	A1	A2	A1	A2
Min	126.11	133.17	–1.30	–0.06	1.00	1.00	0.00	0.00	0.00	0.00	12.47	41.13
Max	700.77	610.70	7.33	6.86	11.00	10.00	4.00	4.00	11.00	10.00	195.18	155.76
Median	371.41	408.50	3.03	3.21	4.00	6.00	1.00	2.00	4.00	4.00	77.25	90.41
Mean	370.11	411.24	3.14	3.33	4.67	5.71	1.34	1.66	4.74	4.44	76.34	90.72
Skew	0.36	–0.20	0.07	0.25	0.63	0.10	0.32	–0.08	0.60	0.54	0.69	0.18
Kurtosis	0.57	1.20	0.92	0.03	0.48	–0.37	–0.05	–0.21	0.06	0.70	2.18	–0.37
p-value	6.63 × 10–15		1.58 × 10–2		4.49 × 10–17		1.81 × 10–7		2.83 × 10–2		2.29 × 10–18

^aWhere *cLogP is calculated LogP, MW: molecular weight, LogP: octanol–water partition coefficient, HBA: no. of hydrogen bond acceptors, HBD: no. of hydrogen bond donors, nRB: no. of rotatable bonds, and TPSA: topological polar surface area.

1.

Chemical space analysis of PD-1/PD-L1 inhibitors. (A) PCA score plot showing substantial overlap between active and inactive compounds. (B) LogP density distribution. (C) LogP–TPSA density contours highlighting active compound enrichment within an optimal hydrophobicity-polarity region. (D) MW–LogP scatter plot showing a positive correlation. (E) TPSA violin plots comparing polarity distributions. (F) TPSA–HBA correlation showing clustering within a moderate-polarity chemical space.

The LogP–TPSA density map (Figure C) showed substantial overlap between active and inactive compounds, primarily within LogP 2–4 and TPSA 60–100 Ų, although active molecules were enriched at lower TPSA values. Likewise, the MW–LogP plot (Figure D) demonstrated a positive relationship between molecular size and lipophilicity, consistent with the importance of hydrophobic surface complementarity in this PD-L1 interaction interface. TPSA strongly correlated with HBA (Figure F), while inactive compounds clustered at higher TPSA and HBA values, indicating increased polarity that may reduce membrane permeability and PD-L1 binding.

The Mann–Whitney U test showed significant differences (p < 0.05) for all descriptors with MW, HBA, and TPSA exhibiting the greatest discrimination between active and inactive compounds. Skewness and kurtosis indicated differences in distribution shape between active and inactive compounds, particularly for TPSA and hydrogen-bonding descriptors (Table ).

To further resolve the global chemical space, PCA using six physicochemical descriptors was performed. The first three principal components explained 82% of the total variance, indicating that these descriptors capture most of the data set’s structural diversity (Table ). PC1 was primarily driven by molecular size and polarity with the highest positive loadings for MW (0.555), TPSA (0.548), and HBD (0.508), defining a size–polarity axis. PC2 was dominated by LogP (0.658) and negatively associated with HBA (−0.540), representing a hydrophobicity–hydrogen bonding trade-off. PC3 was mainly influenced by HBA (0.724) and HBD (−0.507), reflecting variation in hydrogen-bonding architecture independent of molecular size.

2. PCA Analysis of the Six Properties and Their Cumulative Variance.

Descriptors	PC1	PC2	PC3
MW	0.555	0.249	0.105
LogP	0.065	0.658	0.350
HBD	0.508	–0.108	–0.507
HBA	0.087	–0.540	0.724
RB	0.349	0.300	0.291
TPSA	0.548	–0.335	0.022
Cumulative Variance (%)	0.41	0.67	0.82

The PCA score plot (Figure ) showed substantial overlap between active and inactive compounds, indicating a shared chemical space. However, active compounds formed a more compact cluster toward positive PC1 values, whereas inactive compounds were more dispersed, reflecting a greater structural and physicochemical diversity.

2.

PCA score plot of active and inactive PD-1/PD-L1 inhibitors. The first two principal components explain 67% of the total variance. Active compound (blue) cluster toward positive PC1 values, whereas inactive compounds (orange) show broader dispersion with partial overlap, indicating shared chemical space but class-specific structural trends.

To predict the bioactivity beyond physicochemical descriptors alone, we used MACCS, PubChem, and Mordred descriptors for QSAR modeling. Figure Ten classification algorithms were optimized through hyperparameter tuning (Supplementary Table 1) using a fixed random state (42). Ensemble and boosting methods consistently outperformed linear and probabilistic models with Mordred descriptors showing the strongest generalization performance, followed by MACCS and PubChem (Figure ). XGBoost achieved the strongest performance, reaching a test accuracy of 0.834 and MCC of 0.676 with Mordred descriptors and an accuracy of 0.852 and MCC of 0.700 with MACCS fingerprints. LightGBM also performed strongly with MACCS (MCC = 0.697) and Mordred (MCC = 0.637). Random forest and extra trees showed consistent performance across descriptor sets with random forest achieving MCC values of 0.662 (MACCS), 0.603 (PubChem), and 0.647 (Mordred). Extensive overlap between the training and test sets across descriptor spaces confirmed evaluation within a chemically representative applicability domain (Figure ).

4.

Applicability domain analysis of training and test sets. Principal component projections of MACCS, PubChem, and Mordred descriptors showing substantial overlap between training and test compounds, confirming reliable chemical space coverage and model validation.

3.

Comparative ML performance across three molecular descriptors. Performance of ten algorithms trained using (A) MACCS, (B) PubChem, and (C) Mordred descriptors. Ensemble tree-based models consistently outperformed other algorithms with XGBoost and LightGBM trained on MACCS fingerprints achieving the highest external validation accuracy and MCC.

Gaussian Process classifier performed well, with MCC scores of 0.647 (Mordred) and 0.676 (MACCS), but showed reduced performance with PubChem (0.561). Descriptor selection strongly influenced predictive performance. MACCS fingerprints provided the most consistent MCC scores across top-performing models, including XGBoost (0.700), LightGBM (0.697), Gaussian Process (0.676), and Random Forest (0.662). PubChem fingerprints showed potential overfitting with high training performance (accuracy ∼ 0.985; MCC ∼ 0.970) but lower test performance.

To interpret the molecular basis of these predictions, SHAP analysis using MACCS, PubChem, and Mordred descriptors identified the molecular features driving PD-1/PD-L1 inhibition (Figure ). Despite different molecular representations, all three descriptor sets converged on similar activity-associated characteristics (Supplementary Tables S2–S4). Nitrogen-containing Mordred descriptors (nN, NaaN, NaasN, NaaNH, NssNH) and PubChem (16, 300, 375, 418, 422, 515, 559, 638) showed the strongest positive SHAP contributions, highlighting nitrogen-rich heteroaromatic systems as key determinants of PD-L1 recognition. Molecular complexity (fragCpx) and charge-distributed surface descriptors (PEOE-VSA4, PEOE-VSA7, PEOE-VSA10) positively influenced activity, whereas highly substituted carbon environments (e.g., NdssC) contributed negatively (Supplementary Table S5). Sulfur-containing motifs were consistently identified across all descriptors, including Mordred descriptors (NssS, NdS, NaaS, NddssS, SaasC), PubChem fingerprints (33, 305, 414), and MACCS keys (37, 38, 52, 57), indicating privileged features of PD-1/PD-L1 inhibitors. Likewise, aromatic scaffold descriptors from Mordred (nARing, nFARing, nFRing, n10FARing, n12FHRing), PubChem (357, 358, 697, 712, 734), and MACCS (109, 113, 128, 144) showed positive SHAP contributions, consistent with benzoxazine-, chromene-, and biphenyl-based scaffolds that stabilize the PD-L1 interface.

5.

Global SHAP summary plot for PD-1/PD-L1 inhibitor prediction. Features are ranked by mean absolute SHAP values with each point representing a compound and colored by descriptor value (blue, low; red, high). Positive SHAP values increase the probability of classifying a compound as active, whereas negative values decrease it, highlighting the key molecular determinants of PD-1/PD-L1 inhibitory activity.

Consistent with these descriptor-level trends, scaffold decomposition identified dominant chemotypes accounting for a substantial proportion of active PD-1/PD-L1 inhibitors (Table ). Active scaffolds were enriched in fused oxygen- and nitrogen-containing polyaromatic systems, consistent with SHAP and fingerprint analyses. The most prevalent scaffold (108 compounds) was a fused tricyclic benzoxazine/dibenzoxazepine-like core with tert-butyl and methyl substitutions, resembling the Bristol-Myers-Squibb PD-L1 inhibitor series. The next most abundant scaffolds were fused aromatic amines (93 compounds) and secondary amine-containing benzoxazines (92 compounds). Additional enriched chemotypes included chromene-based fused scaffolds (64, 50, and 50 compounds), biphenyl-expanded fused systems (40 compounds), reduced benzoxazine analogues (40 compounds), diphenyl-fused tricyclic cores (39 compounds), and aromatic ether amine pharmacophores (39 compounds), underscoring the importance of extended aromatic surfaces, heteroatoms, and controlled molecular flexibility for PD-1/PD-L1 inhibition.

3. Top 10 Enriched Scaffolds Identified from PD-1/PD-L1 Inhibitors.

To validate these QSAR relationships at the structural level, the five highest-ranked compounds were docked into the PD-L1 interface using the cocrystallized inhibitor ChEMBL5835997 as the reference. The reference achieved a docking score of −12.349 kcal/mol (Figure ), forming alkyl interactions with Met115 and Ala121, π-anion interactions with Asp122, π–π stacking with Tyr123, and hydrogen bonding with Arg125 (chain A), while maintaining π–π stacking with Tyr56 and hydrophobic contacts with Met115 and Ala121 (chain B) (Supplementary Table S6). Conserved interaction hotspots included Tyr56, Met115, Ala121, Asp122, Tyr123, Lys124, and Arg125 (Table ). ChEMBL4529967 showed a docking score of −11.002 kcal/mol despite exhibiting a very high experimental potency (pIC₅₀ = 13.09), maintaining hydrogen bonding (Arg125), alkyl interactions (Met115, Ala121), π–π stacking (Tyr123), and additional contacts (Asp61, Asn63, Gln66, and Val68) on the opposite monomer. ChEMBL5202283 achieved a docking score of −11.165 kcal/mol, maintaining alkyl interactions with Met115 and Ala121, hydrogen bonding and π-anion interactions with Asp122, aromatic contacts with Tyr56 and Tyr123, and a unique halogen-mediated interaction that may further stabilize binding. ChEMBL5423533 exhibited one of the strongest docking scores (−11.914 kcal/mol) through extensive π–π interactions with Tyr56 and Tyr123, together with alkyl contacts involving Met115 and Ala121 across both chains, resembling the binding mode of biphenyl-like PD-L1 inhibitors.

6.

Binding mode analysis of PD-L1 inhibitors. Top: Three-dimensional binding poses of the reference ligand (PDB: 6R3K) and lead compounds within the PD-L1 interface (Chain A, green; Chain B, orange). Bottom: 2D-interaction maps showing hydrogen bonds, π-stacking, and hydrophobic contacts with key PD-L1 residues.

4. Summary of Key Protein–Ligand Interactions at the PD-L1 Dimer Interface.

Compounds	Interacting Residues (Chain A)	Interacting Residues (Chain B)	Interactions
ChEMBL5835997 (Reference) (S = −12.349 kcal/mol; pIC50 = 8.73)	Phe19, Thr20, Gln66, Met115, Ala121, Asp122, Tyr123, Lys124, Arg125	Tyr56, Gln66, Met115, Ala121	π–π stacking, alkyl, π-anion, Hydrogen bonding
ChEMBL4529967 (S = −11.002 kcal/mol; pIC50 = 13.09)	Arg125, Met115, Ala121, Asp122, Tyr123	Asp61, Asn63, Gln66, Val68, Met115, Ile54, Ser117, Ala121	Hydrogen bonding, π–π stacking, alkyl, π-anion
ChEMBL5202283 (S = −11.165 kcal/mol; pIC50 = 13.09)	Ala18, Met115, Ala121, Asp122, Tyr123, Arg125	Tyr56, Asp61, Gln66, Met115, Ala121	Hydrogen bonding, π–π stacking, alkyl, halogen, π-anion
ChEMBL5423533 (S = −11.914 kcal/mol; pIC50 = 12)	Ile54, Tyr56, Gln66, Val68, Met115, Ala121	Ala18, Tyr56, Ala121, Asp122, Tyr123	π–π stacking, alkyl, CH interactions, hydrogen bonding
ChEMBL5427257 (S = −8.517 kcal/mol; pIC50 = 13.09)	Ala121, Asp122, Tyr123, Lys124	Tyr56, Asp61, Asn63, Gln66, Val76, His78, Met115	π–π stacking, π-anion, alkyl, CH interactions, unfavorable acceptor–acceptor

Although ChEMBL5427257 retained potency (pIC₅₀ = 13.09), it produced the weakest docking score (−8.517 kcal/mol). The ligand preserved interactions with Tyr56, Lys124, Met115, Ala121, Asp122, and Tyr123, but an unfavorable acceptor–acceptor interaction likely reduced its predicted binding affinity.

To assess the temporal stability of these docked poses, backbone RMSD analysis over 200 ns demonstrated ligand-dependent PD-L1 stabilization (Figure A,B). Complex 1 rapidly equilibrated (∼25 ns) and remained highly stable (3.2 Å), whereas Complex 2 showed moderate adaptation (4.05 Å). Complex 3 sampled multiple conformational states (2–6 Å), Complex 4 exhibited persistent instability (5–7 Å), and Complex 5 showed intermediate fluctuations (∼4 Å). Radius of gyration (Rg) analysis supported these trends with Complex 1 maintaining compactness (20.79 Å), Complex 2/3 showing moderate expansion (∼21 Å), and Complex 4/5 displaying increased expansion (∼23 Å), indicating reduced structural compactness.

7.

Molecular dynamics analysis of PD-L1–ligand complexes over 200 ns showing (A) RMSD, (B) radius of gyration, (C) minimum protein–ligand distance, and (D) free-energy landscapes. Complex 1: ChEMBL5835997; Complex 2: ChEMBL4529967; Complex 3: ChEMBL5202283; Complex 4: ChEMBL5423533; Complex 5: ChEMBL5427257. The changes in RMSD, Rg, and free energy landscape suggested the protein–ligand complex stability, compactness, and the energetic states of the system.

Protein–ligand distance analysis (Figure C) revealed stable contacts for Complex 1 and Complex 2 (∼3.1 Å), whereas Complex 3 remained within a stable interaction range despite fluctuations. Complex 4 showed variable contacts, and Complex 5 displayed distances up to 4.8 Å, suggesting intermittent weakening of protein–ligand contacts. Free-energy landscapes (Figure D) showed that Complex 1 occupied a single deep energy basin, indicating a stable low-entropy conformation. Complex 2 displayed moderate flexibility, Complex 3 showed multiple metastable states, Complex 4 exhibited a heterogeneous landscape, and Complex 5 exhibited a dominant basin with minor substates (Supplementary Figures S1–S5).

Finally, to quantify the energetic basis of these binding differences, MM/PBSA calculations indicated that van der Waals interactions were major favorable contributors to binding, whereas electrostatic contributions varied across complexes (Table ). The reference compound ChEMBL5835997 exhibited the most favorable binding free energy (ΔG–bind = −50.24 kcal/mol), driven primarily by vdW (−80.84 kcal/mol) and electrostatic (−113.56 kcal/mol) interactions that partially compensated for unfavorable polar solvation (150.61 kcal/mol). The favorable vdW contribution is consistent with extensive close-range packing involving key residues, including Tyr56, Met115, Ala121, and Tyr123. Among the compounds, ChEMBL5423533 showed the strongest binding free energy (−34.98 kcal/mol), followed by ChEMBL5202283 (−34.39 kcal/mol), ChEMBL4529967 (−28.74 kcal/mol), and ChEMBL5427257 (−19.11 kcal/mol), reproducing the docking-score ranking and broadly supporting the MD-derived stability profiles.

5. MM/PBSA Energy Decomposition of PD-L1-Ligand Complexes Showing vdW (VDWAALS), Electrostatic (EEL), Polar Solvation (EPB), Nonpolar Solvation (ENPOLAR), Gas-Phase (GGAS), Total Solvation (GSOLV), and Overall Binding Free (ΔG–Bind) Energy.

ChEMBL ID	VDWAALS	EEL	EPB	ENPOLAR	GGAS	GSOLV	TOTAL
5835997	–80.84 ± 3.14	–113.56 ± 8.74	150.61 ± 14.33	–6.45 ± 0.21	–194.4 ± 18.56	144.16 ± 14.35	–50.24 ± 7.64
4529967	–65.26 ± 3.86	–12.64 ± 6.39	54.96 ± 7.5	–5.81 ± 0.26	–77.9 ± 7.55	49.15 ± 7.37	–28.74 ± 4.94
5202283	–62.81 ± 3.8	–13.29 ± 6.39	47.43 ± 5.96	–5.71 ± 0.33	–76.1 ± 8.02	41.72 ± 5.85	–34.39 ± 6.21
5423533	–72.76 ± 4.26	–15.23 ± 7.51	59.44 ± 9.35	–6.43 ± 0.28	–87.99 ± 8.64	53.01 ± 9.26	–34.98 ± 4.75
5427257	–36.38 ± 6.11	–6.77 ± 5.2	28.3 ± 8.28	–4.26 ± 0.46	–43.15 ± 8.6	24.04 ± 7.99	–19.11 ± 3.94

Energy decomposition demonstrated that vdW interactions were the principal favorable energetic contributors to binding across all complexes, whereas electrostatic contributions were more variable. Polar solvation (EPB) consistently opposed complex formation, reflecting the energetic cost of desolvating polar groups before ligand binding. In contrast, nonpolar solvation remained uniformly favorable (−4.26 to –6.45 kcal/mol).

The principal finding of this study is that independent molecular representations converge on the same set of molecular determinants of PD-1/PD-L1 inhibition. This is observed across EDA, PCA, QSAR, SHAP, and scaffold enrichment analyses that rely on fundamentally different mathematical representations of molecular structure and indicates that these features are unlikely to be artifacts of any single modeling choice. These independent analyses do not merely report convergent conclusions, in parallel; they validate one another through a chain of increasingly stringent, methodologically orthogonal tests. Ligand-based analyses (EDA, PCA, QSAR, SHAP, scaffold enrichment) operate purely on molecular structure and activity labels without reference to the PD-L1 protein and identify the features as statistical correlates of activity; at this stage, these remain correlative hypotheses that could in principle reflect data set bias rather than genuine molecular recognition. Structure-based analyses (Docking, MD, MM/PBSA) test these hypotheses independently. Because each method uses a distinct computational basis and none is derived from the others’ input features, agreement across this chain substantially reduces the likelihood that the identified determinants are artifacts of any single method, data set, or descriptor choice.

Small-molecule inhibition of the PD-1/PD-L1 immune checkpoint has been largely shaped by early Bristol Myers Squibb (BMS)-derived chemotypes in which biphenyl and extended polyaromatic scaffolds were optimized to stabilize PD-L1 homodimerization through extensive hydrophobic contacts within an induced protein–protein interface. , While this framework established PD-L1 as a druggable PPI target, it has also led to a strong bias toward large, highly aromatic, and lipophilic molecules with suboptimal physicochemical profiles, motivating renewed efforts to refine the structure–activity landscape of this chemotype space. ,

Consistent with this need for refinement, EDA in the present study reveals that active PD-L1 inhibitors occupy a more restricted and optimized physicochemical region than previously assumed. Active compounds are characterized by lower MW, reduced polarity, and a more balanced hydrophobic–hydrophilic distribution compared with inactive molecules. A clear separation in TPSA threshold (∼80–85 Ų) separates active from inactive compounds, above which inhibitory potency sharply decreases, suggesting that excessive polarity imposes a desolvation penalty poorly compensated within the hydrophobic PD-L1 interface. These findings indicate that ligand efficiency, rather than hydrophobic surface maximization alone, is a key determinant of PD-L1 inhibitory activity, consistent with recent optimization studies emphasizing physicochemical efficiency over molecular expansion.

Alongside this physicochemical refinement, sulfur-containing motifs were independently enriched across multiple descriptors. Given sulfur’s high polarizability, this substitution likely enhances dispersion interactions and hydrophobic complementarity and may subtly modulate π–π stacking with residues such as Tyr56 and Tyr123, representing a meaningful but underexplored contributor to activity, consistent with recent scaffold-diversification efforts extending beyond the classical biphenyl-dominated space. , The sulfur-associated features across independent descriptors strongly suggests that sulfur chemistry represents a meaningful but underexplored contributor to PD-L1 inhibitory activity.

ML models trained on MACCS, PubChem, and Mordred descriptors consistently indicated that PD-L1 inhibition arises from nonlinear structure–activity relationships with ensemble learning methods significantly outperforming linear classifiers, reflecting the importance of interacting features such as aromaticity, heteroatom distribution, and topology.

XAI using SHAP provided interpretability across all descriptors on five characteristics associated with PD-1/PD-L1 inhibitors: (i) nitrogen-rich heteroaromatic systems, (ii) sulfur-containing motifs, (iii) extended fused aromatic scaffolds, (iv) balanced electrostatic surface properties combined with moderate lipophilicity, and (v) structurally organized molecular architectures with controlled three-dimensionality.

Scaffold enrichment translated these SHAP-identified descriptors and fingerprints into explicit chemical architectures: the nitrogen-rich heteroaromatic and fused-aromatic features highlighted by SHAP correspond directly to the enriched benzoxazine, dibenzoxazepine, chromene, and diphenyl-fused systems observed in the active set, consistent with known BMS scaffolds, alongside partially saturated, more flexible analogues, together suggesting that PD-L1 inhibitors require rigid aromatic cores for π–π stacking combined with localized flexibility for induced-fit adaptation.

Structural and energetic analyses provided convergent support for these descriptor-level findings. Docking identified a conserved hotspot network (Tyr56, Gln66, Met115, Ala121, Asp122, Tyr123, and Arg125) matching residues previously reported in BMS-inhibitor crystal structures, dominated by π–π stacking between ligand cores and the Tyr56–Tyr123 pair, a structural rationale for the aromatic descriptors’ predictive importance. Sulfur atoms frequently localized near Met115’s polarizable thioether side chain, consistent with dispersion-driven rather than directional binding, while nitrogen-rich heterocycles preferentially engaged entrance residues (Gln66, Asp61, Asn63, Lys124, Arg125), providing electrostatic orientation. MD simulations confirmed these interactions remained stable under near-physiological conditions with limited peripheral flexibility enabling adaptive optimization, reinforcing the rigid-core/localized-flexibility model. MM/PBSA decomposition showed binding dominated by van der Waals interactions with electrostatic contributions limited by polar desolvation penalties, mechanistically explaining both the TPSA activity threshold and the stabilizing role of sulfur.

This convergence across QSAR, SHAP, scaffold, docking, MD, and MM/PBSA analyses provides cross-validated support for the identified structure–activity relationships, reducing the likelihood of spurious correlations and demonstrating the value of coupling XAI with physics-based modeling. The resulting framework of bidirectional validation between ligand-based and structure-based methods is not inherently specific to PD-1/PD-L1 and may extend to other poorly druggable PPI targets (e.g., MDM2–p53, Bcl-2 family, TIGIT/PVR), whereas the specific chemical determinants identified here (nitrogen/sulfur-rich heteroaromatic, fused-scaffold, compact-lipophilic profiles) remain PD-L1-specific findings that should be regarded as hypothesis-generating guidance pending experimental validation across structurally diverse chemical series, rather than definitive design rules. More broadly, this work illustrates how coupling XAI with multiscale computational modeling can help bridge statistical pattern recognition and mechanistic understanding in drug discovery, a strategy that may generalize to other protein–protein interaction therapeutics that remain difficult for conventional structure-based design due to their flat, dynamic, and often poorly druggable interfaces.

Safety

No unexpected or unusually high safety hazards were encountered.

Glossary

Abbreviations

PD-1

Programmed cell death protein 1

PD-L1

Programmed death-ligand 1

PPI

Protein–Protein Interaction

XAI

Explainable Artificial Intelligence

SAR

Quantitative Structure–Activity Relationship

ML

Machine Learning

MD

Molecular Dynamics

MM/PBSA

Molecular Mechanics/Poisson–Boltzmann Surface Area

SHAP

SHapley Additive exPlanations

MACCS

Molecular ACCess System

MCC

Matthews Correlation Coefficient

PCA

Principal Component Analysis

Rg

Radius of Gyration

RMSD

Root Mean Square Deviation

LogP

Logarithm of the Octanol–Water Partition Coefficient

IC₅₀

Half-Maximal Inhibitory Concentration

MAPK

Mitogen-Activated Protein Kinase

PI3K/AKT

Phosphoinositide 3-Kinase/Protein Kinase B

HIF-1α

Hypoxia-Inducible Factor 1-alpha

STAT3

Signal Transducer and Activator of Transcription 3

NF-κB

Nuclear Factor Kappa-Light-Chain-Enhancer of Activated B Cells

XGBoost

eXtreme Gradient Boosting

LightGBM

Light Gradient Boosting Machine

vdW

van der Waals

EPB

Polar Solvation Energy

ENPOLAR

Nonpolar Solvation Energy

GGAS

Gas-Phase Energy

GSOLV

Total Solvation Free Energy

ΔG–bind

Binding Free Energy

VDWAALS

van der Waals Energy

EEL

Electrostatic Energy

ns

Nanoseconds

The data set used in this study is available at https://github.com/SI319.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsmedchemlett.6c00382.

• Experimental Procedures (DOCX)

• RMSD profiles; RMSF profiles; FEL analysis; minimum protein–ligand distance profiles; radius of gyration profiles (PDF)

• Hyperparameter tuning; PubChem-SHAP; MACCS-SHAP; Mordred-SHAP; convergence of SHAP; PPI profile (XLSX)

A.M.: conceptualization, methodology, validation, formal analysis, investigation, resources, visualization, writing-review and editing, project administration. S.I.: conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing-original draft preparation, writing-review and editing, visualization, project administration.

The authors received no funding for this work.

The authors declare no competing financial interest.

#.

mananriaz012@gmail.com; abdulmanan@gachon.ac.kr