Computational Screening of AI-Generated Antihypertensive Virtual Leads for Polypharmacological Anticancer Potential

Abstract

Background: The growing recognition of shared molecular pathways and molecular signatures between cardiovascular diseases and cancer has motivated interest in exploring antihypertensive-associated chemical space for oncological applications. Concurrently, artificial intelligence (AI)-driven molecular generation has enabled the rapid creation of virtual lead candidates for specific therapeutic indications, although their broader biological interaction profiles often remain unexplored. Methods: In this paper, we explore the computational screening of a library of AI-generated antihypertensive virtual lead compounds to evaluate their polypharmacological anticancer potential. The compounds were originally designed and prioritized for modulating $\mathsf{\beta }$-adrenergic receptors but are here re-evaluated in a cancer-focused context using a multi-stage in silico approach. We chose five (5) known cancer target proteins and performed compound profiling for drug-likeness, pharmacokinetic suitability, and safety. Docking simulations, binding free energy estimates, molecular interaction mapping, and pharmacophore modeling were used to evaluate the molecules’ interactions with the cancer-linked protein targets. We employed the binding free energy estimates of the ligand–protein complexes to determine compounds with polypharmacological anticancer potential. In addition, molecular dynamics simulations of some of the compounds with polypharmacological anticancer potential were employed to evaluate binding stability and dynamic behavior of selected ligand–target complexes. Results: Several compounds showed good docking scores, physicochemical characteristics, and pharmacokinetic profiles. Also, the results reveal that several AI-generated antihypertensive virtual leads exhibit favorable multi-target binding profiles, with consistent docking affinities and stable interaction networks across multiple cancer-related targets. Conclusions: Our findings suggest that several of the hypothetically evaluated compounds exhibit favorable physicochemical properties, acceptable predicted pharmacokinetic and safety profiles, and consistent predicted binding affinities across multiple cancer-relevant targets.

1. Introduction

Cancer remains one of the leading causes of mortality worldwide, with its progression driven by highly complex, heterogeneous, and adaptive molecular networks [1,2]. Hence, the discovery and development of new cancer therapeutics is vital to continued progress against the disease. Despite significant advances in therapies, the clinical success of many anticancer agents remains limited by drug resistance, pathway redundancy, off-target toxicity, and high attrition rates during the development process [1,2]. These challenges have increasingly shifted attention toward polypharmacological strategies, in which small molecules modulate multiple cancer-relevant targets or pathways simultaneously, thereby enhancing therapeutic robustness and reducing the emergence of resistance [3,4,5].

Currently, computational approaches have emerged as powerful tools for identifying multi-target drug candidates, particularly in the early stages of drug discovery where experimental screening is cost- and time-intensive [6,7]. Recent advances in artificial intelligence (AI)-driven molecular generation have enabled the rapid creation of chemically diverse libraries enriched for specific pharmacological properties [8,9]. AI-based generative models can efficiently explore vast regions of chemical space, surpassing traditional medicinal chemistry approaches, and produce virtual lead candidates optimized for predefined target classes or physiological effects. However, while such AI-generated molecules are often designed for a primary therapeutic indication, their broader biological interaction profiles frequently remain underexplored. Such latent interaction potential may represent an opportunity, especially in therapeutic areas such as cancer, where polypharmacology is often desirable [10]. Therefore, systematic computational evaluation of the AI-generated virtual leads against alternative disease-relevant targets represents a valuable opportunity to uncover previously unrecognized therapeutic potential.

The systematic application of the strategy of evaluating virtual leads against alternative targets requires an understanding of the relationships among the molecular pathways of the diseases of interest. In the case of cancer, there is increasing evidence that suggests significant biological and molecular overlap between cardiovascular diseases and cancer. The evidence includes overlapping signaling pathways, inflammatory mediators, angiogenic processes, metabolic reprogramming, and stress-response mechanisms [11,12]. Clinical epidemiological and mechanistic studies have also highlighted that several cardiovascular drugs, including antihypertensives, exhibit anticancer-associated effects, either through direct modulation of tumor biology or via effects on the tumor microenvironment [13]. This suggests that chemical scaffolds developed for antihypertensive modulation may also engage cancer-associated molecular targets. Indeed, several cardiovascular drugs and lead compounds have demonstrated anticancer-relevant effects through mechanisms extending beyond their original therapeutic scope. For instance, cardiovascular medications such as Losartan, Captopril, Propranolol, Candesartan, and Irbesartan have been considered through the drug repositioning process for cancer treatment [14]. Within this context, computational screening provides a cost-effective, scalable, and hypothesis-driven framework for evaluating the anticancer relevance of virtual lead libraries prior to experimental validation. Structure-based docking, binding free energy estimation, pharmacokinetic and toxicity prediction, and molecular dynamics simulations collectively enable early assessment of binding feasibility, stability, safety, and multi-target engagement. Importantly, when applied to AI-generated compounds, these pipelines enable the rational prioritization of candidates with favorable polypharmacological interaction profiles, without overstating their clinical relevance.

Beyond small-molecule design, recent advances in artificial intelligence have demonstrated the capacity of transformer-based and graph representation learning frameworks to capture complex, non-linear biological relationships across heterogeneous disease networks. Such approaches have been successfully applied to predict circRNA–disease associations and regulatory RNA interaction networks, highlighting the ability of AI models to uncover latent biological connectivity that is difficult to resolve experimentally [15,16]. These developments underscore the broader relevance of AI-driven, network-aware computational frameworks for systematic exploration of under-characterized biological interaction spaces, including those relevant to multi-target drug discovery.

In this study, we conduct a computational screening of AI-generated antihypertensive virtual lead candidates to investigate their potential as polypharmacological anticancer agents. The molecules, originally designed for modulation of antihypertensive targets, the $\beta$-adrenergic receptor, are evaluated against a panel of cancer-associated protein targets representing diverse oncogenic pathways. The drug-likeness and absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles of the molecules are evaluated. Molecules that pass the drug-likeness and ADMET tests are further screened for their docking scores and Molecular Mechanics with Generalized Born Surface Area (MM-GBSA) estimates. We use their MM-GBSA scores across the cancer target proteins to predict their polypharmacological potentials. Molecular interaction mapping and molecular dynamics simulations are performed for the molecules with the most multi-target potentials. Hence, the novelty in this work is the demonstration of the utilization of computational screening as a systematic approach for re-evaluating AI-generated virtual leads in new disease contexts, providing a rational foundation for subsequent experimental investigation.

2. Results and Discussion

Here, following Figure 22, we screen the 60 AI-generated molecules for their potential to modulate cancerous cells. Firstly, we evaluated the drug-likeness and ADMET properties of the 60 molecules to determine their suitability as drug molecules. Secondly, the molecules that passed the ligand profiling tests are subjected to virtual screening, and their docking scores and MM-GBSA estimates are evaluated against each of the cancer target proteins listed in Table 7. Also, we performed molecular interaction mapping and pharmacophore modeling of the two most promising ligands, which exhibit multi-target characteristics. Finally, to observe and analyze the stability, fluctuations, and conformational changes of the molecules over time, MD simulations are carried out on the two most promising ligands that have multi-target characteristics.

2.1. Ligand Profiling of the 60 Molecules

To evaluate the candidacy of the 60 molecules as lead drug molecules, we profiled their drug-likeness and ADMET properties. In all, 25 compounds passed the drug-likeness and ADMET tests as presented in

Table 1. Physicochemical properties of lead compounds.

Compound	Mol. MW	donorHB	accptHB	QPlogPo/w	RuleOfFive
Compound 1	249.309	2	4.95	1.377	0
Compound 3	338.233	2	3.25	3.80	0
Compound 4	290.361	4	5.75	0.912	0
Compound 8	277.363	2	4.95	2.18	0
Compound 9	325.407	4	4.00	2.115	0
Compound 10	262.308	3	6.70	0.227	0
Compound 11	301.344	2	5.20	1.669	0
Compound 14	276.162	2	4.20	2.336	0
Compound 15	290.339	2	6.70	1.204	0
Compound 16	289.374	2	4.95	2.426	0
Compound 17	247.336	2	3.25	2.515	0
Compound 19	303.421	0	4.20	3.97	0
Compound 21	353.438	1	4.95	3.829	0
Compound 22	305.394	1	4.95	2.988	0
Compound 28	360.430	2	9.20	0.187	0
Compound 33	387.883	1	5.45	4.12	0
Compound 35	402.510	1	7.95	2.637	0
Compound 36	364.421	1	6.45	2.809	0
Compound 37	347.431	0	6.90	2.878	0
Compound 38	292.377	3	6.45	0.839	0
Compound 39	255.377	0	4.00	2.709	0
Compound 51	375.442	3	8.95	0.284	0
Compound 52	303.360	3	5.40	1.496	0
Compound 55	219.283	2	4.20	1.641	0
Compound 56	322.360	2	7.45	2.054	0

and

Table 2. Predicted ADMET properties of selected lead compounds.

Compound	QPlogS	CIQPlogS	QPlogHERG	QPPcaco	QPlogBB	QPPMDCK	QPlogKp	QPlogKhsa	#metab	%HOA	CNS
Compound 1	−1.92	−2.103	−2.578	948.714	−0.481	735.410	−2.197	−0.524	4	88.291	0
Compound 3	−4.958	−4.748	−4.194	879.223	−0.423	3140.381	−1.854	0.188	4	100	0
Compound 4	−2.568	−2.409	−3.468	72.634	−1.696	60.985	−4.162	−0.454	5	65.594	−2
Compound 8	−3.431	−2.629	−3.078	1592.470	−0.403	1103.989	−2.057	−0.239	5	100	0
Compound 9	−3.589	−3.961	−4.077	64.354	−1.808	54.001	−3.958	−0.020	6	71.700	−2
Compound 10	−2.124	−1.832	−3.440	110.606	−1.570	74.633	−3.909	−0.750	6	64.852	−2
Compound 11	−2.530	−3.246	−2.438	572.561	−0.430	478.089	−2.779	−0.228	3	86.079	0
Compound 14	−3.187	−2.824	−3.385	1044.133	−0.273	4189.060	−2.041	−0.404	3	94.654	0
Compound 15	−3.261	−3.280	−3.895	354.109	−0.639	343.525	−3.829	−0.255	2	79.622	0
Compound 16	−3.330	−2.858	−3.557	923.778	−0.627	781.690	−2.041	−0.169	6	94.226	0
Compound 17	−3.065	−2.581	−3.435	894.535	−0.702	739.549	−2.034	−0.155	5	94.502	−1
Compound 19	−4.506	−4.328	−4.657	4069.371	0.093	4813.861	−1.378	0.263	4	100	1
Compound 21	−5.083	−5.540	−4.694	1649.289	−0.142	1811.932	−2.123	0.493	3	100	0
Compound 22	−3.657	−4.214	−4.055	1240.352	−0.404	1293.796	−2.292	0.087	4	100	0
Compound 28	−1.842	−2.412	−2.179	45.956	−1.456	140.885	−4.133	−0.943	5	57.792	−2
Compound 33	−5.981	−6.041	−4.922	1628.608	−0.001	3518.573	−2.368	0.577	3	100	0
Compound 35	−4.205	−4.367	−2.868	961.246	−0.386	1389.699	−2.615	−0.122	4	95.774	−1
Compound 36	−5.419	−6.168	−4.863	411.010	−0.776	403.568	−3.099	0.161	2	90.176	−1
Compound 37	−3.143	−4.046	−4.111	6548.547	0.426	7995.951	−1.216	−0.334	3	100	1
Compound 38	−1.174	−1.381	−3.613	140.123	−0.473	104.833	−4.665	−0.517	6	70.275	0
Compound 39	−2.312	−1.928	−4.452	1256.045	0.592	1488.513	−3.752	−0.011	3	100	2
Compound 51	−2.121	−2.768	−1.242	204.951	−0.946	337.351	−3.286	−0.947	6	69.980	−1
Compound 52	−2.008	−3.148	−1.707	1043.423	−0.334	753.731	−2.306	−0.383	4	89.731	−1
Compound 55	−2.631	−2.504	−3.949	787.612	−0.667	382.187	−2.872	−0.224	5	88.390	0
Compound 56	−3.797	−3.419	−4.479	1680.318	−0.374	866.905	−2.394	−0.131	4	96.702	0

, respectively.

Table 3. Predicted ADMET properties of non-selected compounds.

Compound	QPlogS	CIQPlogS	QPlogHERG	QPPcaco	QPlogBB	QPPMDCK	QPlogKp	QPlogKhsa	#metab	%HOA	CNS
Compound 12	−5.901	−5.017	−5.293	1687.39	−0.415	870.849	−2.059	0.813	100.000	4	0
Compound 13	−4.219	−3.688	−7.068	709.908	−0.339	377.917	−2.742	0.352	100.000	7	1
Compound 18	−4.316	−4.124	−5.598	545.750	−0.981	548.280	−2.588	−0.001	90.971	5	−1
Compound 2	−4.651	−4.546	−5.660	469.277	−0.800	1261.270	−2.779	−0.029	91.505	3	−1
Compound 20	−5.422	−5.177	−5.579	663.456	−0.981	317.504	−2.545	0.466	100.000	7	−1
Compound 60	−5.606	−5.481	−5.546	5174.094	0.422	10,000.000	−0.781	0.417	100.000	3	1
Compound 23	−5.969	−6.313	−5.678	1570.738	−0.157	1678.128	−1.866	0.659	100.000	2	0
Compound 24	−1.979	−2.042	−6.537	803.860	−0.182	432.256	−2.849	−0.003	91.795	5	1
Compound 25	−3.476	−2.987	−5.657	868.361	0.408	2651.798	−3.370	0.190	100.000	3	1
Compound 26	−4.620	−4.876	−6.137	373.840	0.179	923.878	−4.041	0.426	94.363	4	1
Compound 27	−4.189	−4.541	−6.316	190.567	−0.354	189.175	−4.387	0.394	86.294	5	1
Compound 29	−6.095	−6.241	−5.045	1647.993	−0.059	3642.759	−2.218	0.646	100.000	3	0
Compound 30	−4.818	−4.896	−5.575	543.818	−0.629	2358.214	−2.674	−0.008	93.543	4	−1
Compound 31	−5.543	−6.659	−4.521	1704.737	0.149	7343.624	−2.132	0.529	100.000	2	1
Compound 32	−3.130	−3.330	−6.365	417.491	−0.121	413.115	−3.422	0.191	92.222	4	0
Compound 34	−1.779	−3.245	−4.151	44.049	−1.348	23.427	−4.650	−0.714	60.123	5	−2
Compound 40	−4.946	−5.025	−5.564	1730.107	−0.154	8487.859	−1.600	0.123	100.000	4	0
Compound 41	−4.931	−3.529	−5.957	1946.157	0.665	2033.705	−2.823	0.862	100.000	2	2
Compound 42	−3.775	−4.094	−5.543	5198.825	0.149	6923.010	−0.545	0.154	100.000	4	1
Compound 43	−4.222	−4.124	−5.545	457.123	−1.060	428.487	−2.732	−0.009	88.783	5	−2
Compound 44	−4.646	−4.994	−5.522	116.245	−1.893	102.278	−3.712	0.207	81.368	5	−2
Compound 45	−2.919	−2.269	−5.442	654.264	0.017	346.004	−3.759	0.238	94.021	4	1
Compound 46	−3.094	−3.234	−5.841	227.431	−0.396	213.118	−4.343	0.105	84.222	4	0
Compound 47	−5.627	−5.017	−5.746	1691.700	−0.537	873.254	−1.549	0.740	100.000	6	0
Compound 49	−5.505	−5.257	−5.628	3609.426	0.247	10,000.000	−1.091	0.319	100.000	6	1
Compound 5	−0.479	−2.206	−2.566	7.193	−2.382	11.777	−4.888	−0.965	39.841	6	−2
Compound 50	−3.703	−3.399	−6.378	422.146	−0.056	506.714	−3.728	0.224	92.112	3	0
Compound 53	−4.318	−4.200	−5.565	542.619	−0.797	953.635	−2.648	−0.064	91.154	5	−1
Compound 54	−3.511	−4.998	−5.428	407.655	0.209	404.667	−3.848	0.424	93.354	3	1
Compound 57	−5.488	−4.903	−6.349	3220.669	−0.275	3168.607	−0.596	0.508	100.000	5	0
Compound 58	−3.397	−3.768	−5.270	616.124	−1.035	293.093	−2.236	−0.216	88.606	10	−2
Compound 59	−5.608	−5.481	−5.547	5181.141	0.423	10,000.000	−0.780	0.417	100.000	4	1
Compound 6	−0.450	−2.206	−2.521	7.600	−2.344	12.431	−4.848	−0.966	40.314	6	−2
Compound 48	−3.527	−3.234	−6.271	374.505	−0.104	445.225	−3.867	0.150	89.779	5	0
Compound 7	−5.221	−4.732	−5.240	1688.248	−0.380	871.328	−1.878	0.642	100.000	3	0

shows the compounds that failed the ADMET test (all the compounds passed the drug-likeness test). The acceptable values for the parameters in Table 1 and Table 2, which were obtained from [17,18,19], are as follows. The acceptable values of molecular weight (Mol MW), hydrogen bond donor (donorHB), hydrogen bond acceptor (accptHB), QPlogPo/w, and Rule-of-Five (RuleOfFive) are 130–725, 0–6, 2–20, (−2.0)–6.5, and ≤4, respectively. The acceptable values of QPlogS, CIQPlogS, QPlogHERG, QPPCaco, QPlogBB, QPPMDCK, QPlogKp, QPlogKhsa, #metab, Percent Human Oral Absorption (PHOA), and CNS are (−6.5)–0.5, (−6.5)–0.5, ≥$-5.0$, <25 (poor) and >500 (good), (−3.0)–1.2, <25 (poor) and >500 (good), (−8)–(−1), (−1.5)–1.5, 1–8, <25 (poor) and >80 (high), and (−2) (inactive) and (+2) (active), respectively.

2.2. Molecular Docking and MM-GBSA Analysis of the Twenty Five (25) Compounds

Here, the molecular docking and MM-GBSA analyses of the twenty five (25) compounds that passed the ligand profiling test are provided. The compounds are docked on the PDB identifiers, namely, 5U1X, 7U9A, 5EW3, 5YA5, and 8XRR. The corresponding molecular docking scores–MM-GBSA scatter plots are shown in Figure 1, Figure 2, Figure 3, Figure 4 and Figure 5. Each plots shows the comparison between the values of the 25 compounds and those of the known cancer drug molecules.

In some of the figures, the novel molecules show a strong positive correlation between docking score and MM-GBSA, indicating that compounds predicted to dock well also tend to have more favorable binding free energies. However, in others, the correlation coefficient is quite low, suggesting that while the initial docking score provided a starting point, the more accurate, energy-intensive MM−GBSA calculation was crucial in ranking the final candidates.

In Figure 1, the marginal histograms show that several novel molecules achieve substantially more favorable MM-GBSA energies (−50 to −60 kcal/mol) than any existing compound, while also maintaining competitive docking scores. This combined profile suggests that the new molecules may have stronger predicted binding interactions, making them promising therapeutic candidates that warrant further molecular dynamics and experimental validation.

In Figure 2, the marginal histograms indicate that while the existing molecules have stronger predicted binding interactions, the novel molecules occupy a wider and more favorable range of predicted binding energies (approximately −40 to −50 kcal/mol). This suggests these new molecules could offer a starting point for further optimization and experimental testing against the 7U9A target.

In Figure 3, critically, the marginal histograms reveal that the novel compounds achieve superior predicted binding affinities with 5EW3, with MM-GBSA values reaching −60 to −70 kcal/mol—far beyond the range of existing compounds. They also maintain strong docking scores. This combination of an outstanding computational correlation and vastly improved predicted binding energy suggests the novel molecules are promising therapeutic candidates.

The marginal histograms in Figure 4 show that, while the existing molecules have stronger predicted binding interactions with 5YA5, the novel molecules occupy a wider and more favorable range of predicted binding energies (approximately −40 to −50 kcal/mol). This implies that the novel compounds might serve as a starting point for future optimization and experimental testing against the 5YA5 target.

In Figure 5, the marginal histograms reveal that the novel compounds achieve significantly more favorable predicted binding energies for 8XRR, with MM-GBSA values ranging from −50 to −55 kcal/mol, which is beyond the range of the existing compounds. They also occupy a wider and more favorable range of docking scores.

To better contextualize the performance of their AI-generated leads, we compare the MM-GBSA of the leads to known standard inhibitors. For the P2X7 standard antagonists GSK1482160 and JNJ-42253432, their interactions with 5U1X are estimated to be −30.17 kcal/mol and −25.01 kcal/mol, respectively, in MM-GBSA. Comparatively, several of the AI-generated leads outperform GSK1482160 and JNJ-42253432 in their interactions with 5U1X in MM-GBSA terms. For instance, compounds 19, 21, 22, and 35 have MM-GBSA scores between −55 kcal/mol and −62 kcal/mol.

The MM-GBSA scores of Tepotinib and Savolitinib with 5YA5 are −95.92 kcal/mol and −64.52 kcal/mol, respectively. Comparatively, the best leads cluster for the AI-generated compounds is around −52 to −57 kcal/mol, which are relatively weaker than Tepotinib and comparable but slightly weaker than Savolitinib. For the VEGFR2 standard inhibitors represented by Pazopanib and Lenvatinib, their interactions with 5EW3 are estimated to be −30.68 kcal/mol and −43.11 kcal/mol, respectively. Comparatively, several AI-generated leads outperform both inhibitors, with compounds 28, 4, and 56 having MM-GBSA scores between −55 kcal/mol and −69 kcal/mol.

Considering Avapritinib and Sunitinib as standard inhibitors of PDGFRA, their MM-GBSA estimates against the 8XRR are −34.91 kcal/mol and −50.79 kcal/mol, respectively. Comparatively, several of the AI-generated leads fall between or beyond the two reference drugs, and the MM-GBSA scores of compounds such as compound 3, 21, 35, 51, and 9 exceed both references. For the standard EGFR inhibitors, represented by Dacomitinib, its MM-GBSA estimate against 7U9A is at −63.68 kcal/mol, which is better than the best AI-generated lead (−45 kcal/mol), indicating moderate affinity.

Table 4. Comparison of MM-GBSA binding affinities between known drugs and virtual leads for selected targets.

Target	PDB Identifier	Known Drug (MM-GBSA, kcal/mol)	Virtual Lead (MM-GBSA, kcal/mol)
P2X7	5U1X	GSK1482160 ($-30.17$) and JNJ-42253432 ($-25.01$)	Compounds 19, 21, 22, and 35 (between $-55$ and $-62$)
c-MET	5YA5	Tepotinib ($-95.92$) and Savolitinib ($-64.52$)
VEGFR2	5EW3	Pazopanib ($-30.68$) and Lenvatinib ($-43.11$)	Compounds 56, 4, and 28 (between $-55$ and $-69$ kcal/mol)
PDGFRA	8XRR	Avapritinib ($-34.91$) and Sunitinib ($-50.79$)	Compound 3, 21, 35, 51, and 9 (>$-51$)
EGFR	7U9A	Dacomitinib ($-63.68$)	All compounds are <$-63.68$

summarizes the comparison between the MM-GBSA scores of known molecules and the presented virtual leads that target specific proteins.

2.3. Polypharmacological Potential Analysis of Novel Compounds

Figure 6 shows the affinity heatmap of the 25 novel compounds against the five different protein receptors. To rank the polypharmacological potentials of the compounds, applying a −40 kcal/mol energy threshold value to Equations (4)–(6) yields the top contenders in

Table 5. Virtual leads with the topmost polyparmacological profile scores.

Compound	SS	APS (kcal mol−1)	Key Observation
9	5	51.75	Hits all five targets with high and consistent binding potency.
19	5	47.17	Hits all five targets with strong multi-target binding affinity.
21	4	54.50	Misses target 5EW3 but exhibits exceptional potency against the remaining four targets.
3	5	50.00	Demonstrates high binding potency across all five cancer-related targets.
14	5	43.55	Hits all five targets, albeit with a slightly reduced average potency.
51	4	50.58	Does not engage target 5UIX but shows strong potency against the four engaged targets.

.

For further analysis, we focus on Compounds 9, 19, 21, and 3. Compounds 9, 19, and 3 are the most promiscuous and potently balanced molecule in the set. Their consistent, strong inhibition across all five diverse targets suggests a high probability of achieving the desired multi-target therapeutic effects. For Compound 21, although it misses target 5EW3 ($\Delta G_{bind}=$−26.88 kcal/mol), it exhibits the highest absolute potency against the four targets it does hit (APS = 54.50, with three scores < −52 and one at −61.63). Hence, going forward, we will conduct a deeper analysis of the interaction between Compounds 9/19/21/3 and their two most predictably potent targets, the 5U1X, 5EW3, and 5YA5.

2.4. Molecular Interaction Analysis of Novel Compounds

In Figure 7, Figure 8, Figure 9 and Figure 10, we present the molecular interaction analysis of Compound 9 and Compound 21 in complex with 5U1X and 5YA5. The ligand–protein interaction diagrams and pharmacophore model of 5U1X–Compound 9, 5U1X–Compound 21, 5YA5–Compound 9, and 5YA5–Compound 21 complexes are presented. The protein is shown as colored ribbons, the ligand in green sticks, and key interacting residues in thin sticks. For the pharmacophore models, hydrogen-bond acceptors are shown in pink spheres/arrows, hydrogen-bond donors are shown in sky blue spheres/arrows, hydrophobic regions are shown in green spheres, and aromatic rings are depicted with orange circles.

Figure 7A shows that Compound 9 forms multiple stabilizing interactions within the binding pocket of 5U1X. Hydrogen bonds are formed with TYR 295, ASP 92, and possibly THR 94 or backbone atoms. The compound also formed aromatic and hydrophobic contacts of $\pi \text{-}\pi$ stacking type with PHE 88, PHE 95, and PHE 102. It forms hydrophobic packing with VAL 312, ILE 310, MET 105. An electrostatic/salt bridge potential also exists between the compound and ARG 316/ASP 92. The predicted binding mode of Compound 9 in the orthosteric site of 5U1X is shown in Figure 7B. The pharmacophore summary for Compound 9–5U1X interaction aligns well with the observed interactions in Figure 7C. The four (4) hydrogen donors and three (3) hydrogen acceptors matches multiple hydrogen bonds seen with polar/charged residues (ASP, TYR, THR). The two (2) aromatic rings corresponds to $\pi –$stacking with PHE 88, 95, 102. The single hydrophobic region fits the hydrophobic cluster VAL 312, ILE 310, MET 105. The high count of four hydrogen donors and three hydrogen acceptors defines the binding motive as highly dependent on polar anchoring within the active site.

Figure 8A shows that the interaction of Compound 21 and 5U1X utilizes a mix of crucial non-covalent interactions. Hydrogen bonds are formed with residues such as TYR 295, TYR 93, THR 94, and possibly backbone atoms of MET 105 or ASP 92. Aromatic contacts are formed utilizing the $\pi \text{-}\pi$ stacking with PHE 293, PHE 95, and possibly PHE 103, and hydrophobic contacts are established using hydrophobic packing with ILE 310, VAL 312, and MET 105. There also exist electrostatic/charged interactions with LYS 297 (positive) and ASP 92 (negative). The predicted binding mode of Compound 9 in the orthosteric site of 5U1X is shown in Figure 8B. The corresponding pharmacophore model in Figure 7C highlights a single hydrogen donor and five (5) hydrogen acceptors, indicating a highly acceptor-driven binding mode, likely formed with tyrosine/hydroxy residues (TYR 295, TYR 93, THR 94). The presence of three (3) aromatic rings matches a strong $\pi \text{-}$stacking network with multiple phenylalanines (PHE 293, 95, 103). The two (2) hydrophobic regions present correspond to aliphatic/aromatic packing with ILE 310, VAL 312, and MET 105. This model mandates a structure that anchors with a hydrogen bond donor in the shallow pocket (the phenolic OH), is stabilized by two stacked aromatic rings ($\mathrm{TYR}/\mathrm{PHE}$ pocket and the thiazole ring), and extends into the solvent-exposed region with a strong acceptor (the morpholine oxygen or nitrogen).

Figure 9 shows that Compound 21 forms multiple stabilizing interactions (Figure 9A), the docking position (Figure 9B), and a pharmacophore model (Figure 9C) within the binding pocket of 5YA5. A diverse bonding network comprising hydrogen bonds, Pi-interactions, hydrophobic interactions, and electrostatic/polar anchors. Multiple hydrogen bonds are mapped with ASP, TYR, ASN, and GLY residues. This aligns perfectly with the pharmacophore model’s three (3) hydrogen bond donors and three (3) hydrogen bond acceptors. The presence of $\pi \text{-}$interactions with TYR 1230 and TYR 1159 suggest engagement with aromatic residues, crucial for specificity and affinity. The compound engages multiple hydrophobic/van der Waals contacts with ILE, LEU, MET, VAL residues. This matches the pharmacophore’s one hydrophobic region requirement, likely a conserved pocket in the target.

In Figure 10, the ligand–protein interaction diagram, docking pose, and the pharmacophore counts of Compound 21 in complex with 5YA5 are shown. The interaction pattern in Figure 10A indicates that Compound 21 engages multiple polar/charged residues, including ASP1164, ASN1167, and backbone carbonyls near GLY1163/GLY1085. This suggests that the binding pocket presents several hydrogen-bond donor sites, which the ligand satisfies via the four acceptor atoms, indicated in the pharmacophore model (Figure 10C). The compound poses well in the pocket, as shown in Figure 10B. Three aromatic rings on the Compound 21 indicated in the pharmacophore model form $\pi \text{-}$interactions/stacking with aromatic residues such as TYR1159 and TYR1230, and provide a hydrophobic/aromatic anchor that positions the polar groups into the donor pockets. The two hydrophobic vectors shown in the pharmacophore model engage aliphatic side chains such as ILE1084, LEU1157, and MET1160 to improve shape complementarity.

The molecular interaction analysis of Compound 3 in complex with 5YA5 is presented in Figure 11. Figure 11A shows that the interaction of Compound 3 and 5YA5 utilizes hydrophobic contacts that are established using hydrophobic packing with LEU1157, VAL1092, ILE1084, ALA1108, PRO1158, and MET1160. Strong hydrogen bonds are formed from the ligand amide NH bond with the backbone carbonyl of MET1160 and between the phenolic OH forms and TYR1159. There are also $\pi \text{-}$cation interaction of the LYS1161 with the aromatic ring system and $\pi \text{-}\pi$ stacking between the ligand’s aromatic rings and with TYR1222 or TYR1159. The predicted binding mode of Compound 3 in the orthosteric site of 5YA5 is shown in Figure 11B. The corresponding pharmacophore model in Figure 11C highlights two (2) hydrogen bond donors and two (2) hydrogen bond acceptors, consistent with interactions involving polar residues. The model identifies two (2) aromatic rings, which align with a strong $\pi \text{-}$stacking network involving multiple phenylalanine residues.

In Figure 12, the molecular interaction analysis of Compound 3 in complex with 5EW3 is presented. Figure 12A indicates the presence of a hydrogen bond from the amide carbonyl oxygen to a backbone NH and the phenolic OH donation of a hydrogen bond to a polar residue (possibly GLU885 or ASP1046). The positively charged LYS868 interacts with one of the aromatic rings for $\pi \text{-}$cation stabilization. The chlorophenyl and terminal phenyl rings engage in $\pi \text{-}\pi$ stacking with PHE1047 and possibly ILE1024/ILE1025. There are also the hydrophobic contacts indicated by the extensive green regions, which indicate tight packing with a large hydrophobic cluster, such as VAL914, VAL916, VAL867, ILE888, LEU889, LEU1035, ILE892, VAL848, and LEU889. The predicted binding mode of Compound 3 in the active site of 5EW3 is shown in Figure 12B. The corresponding pharmacophore model in Figure 12C highlights two (2) hydrogen bond donors and two (2) hydrogen bond acceptors, consistent with interactions involving polar residues. The model identifies two (2) aromatic rings, which align with a strong $\pi \text{-}$stacking network involving multiple phenylalanine residues. The presence of four (4) hydrogen bonds agrees with the strong hydrogen bond interaction observable in the ligand interaction diagram.

Figure 13A shows that the critical interaction of Compound 19 with 5YA5 is a hydrogen bond between a nitrogen atom in the ligand’s heterocyclic core and the MET 1160 residue. The ligand also engages in a $\pi \text{-}$stacking interaction with TYR 1230 through its terminal phenyl ring. This stabilizes the aromatic portion of the molecule within that specific sub-pocket. The ligand is extensively establishing hydrophobic contacts with key residues, including LEU 1157, PRO 1158, VAL 1092, and ILE 1084. The predicted binding mode of Compound 3 in the active site of 5YA5 is shown in Figure 13B. The corresponding pharmacophore model in Figure 13C highlights two (2) hydrogen bond acceptors, which likely drive polar anchoring on the MET 1160 residue. The model identifies two (2) aromatic rings (chlorophenyl and central fused ring), which are positioned for $\pi \text{-}$stacking or $\pi \text{-}$cation interactions with TYR/PHE residues. The presence of three (3) hydrogen bonds correspond to the chlorophenyl ring, central fused heterocycle, and aliphatic linker/morpholine tail.

2.5. Molecular Dynamics Simulation Results

Here, we present the 100 ns all-atom molecular dynamics simulation that evaluates the stability and conformational behavior of Compound 9 in complex with 5U1X and 5UYA5 and Compound 21 in complex with 5U1X and 5YA5. The simulations are conducted on the GROMACS platform for four key structural descriptors, namely, RMSD, RMSF, radius of gyration, and hydrogen-bonding dynamics.

2.5.1. Molecular Dynamic Simulation for 5U1X–Compound 9 Complex

In Figure 14A, the backbone RMSD remained stable between 0.10 and 0.20 nm, with no evidence of global unfolding or major structural drift. This narrow distribution indicates that the protein relative to the starting conformation rapidly equilibrated and maintained conformational integrity a majority of the simulation time. The ligand exhibited a slightly higher RMSD (0.20–0.30 nm) but remained well-confined within the binding pocket throughout the trajectory. The convergent U-shaped profile of both RMSDs suggests initial relaxation followed by a stable production phase. Collectively, these observations confirm a well-anchored ligand with no dissociation-like fluctuations.

The RMSF analysis (Figure 14B) highlights moderate residue flexibility (0.1 nm average), with elevated values in loop regions (residues 50–100 and 150–200), consistent with dynamic allosteric modulation in P2X7 channels. The absence of high-amplitude fluctuations (>0.2 nm) further supports a stable complex and minimal structural perturbation upon ligand binding.

The average radius of gyration (Figure 14C) remains nearly constant at 1.82 nm, indicating structural compactness with no significant unfolding or conformational collapse due to ligand binding. The persistent hydrogen bonding shown in Figure 14D, averaging 24–25 hydrogen bonds, underscores favorable polar interactions, aligning with the ligand’s pharmacophore. The result also correlates with the low ligand RMSD and suggests strong polar anchoring, which likely contributes to binding affinity and specificity.

2.5.2. Molecular Dynamic Simulation for 5YA5–Compound 21 Complex

Figure 15 shows a stable 5YA5–Compound 21 complex with moderate conformational dynamics. The backbone RMSD (Figure 15A) equilibrates around 2.8 nm after approximately 20 ns, indicating a well-defined yet flexible folded state. The ligand RMSD remains consistently low (at about 1.0 nm), demonstrating a stable, well-bound pose without significant positional drift. The RMSF plot (Figure 15B) reveals heightened flexibility in loop regions for certain residues, whereas the binding site exhibits low fluctuation (<0.2 nm), indicating rigidity that favors consistent ligand interactions. The radius of gyration in Figure 15C maintains an average of 2.80 nm with minor fluctuations (2.6–3.0 nm), indicating overall structural compaction without unfolding. In Figure 15D, the hydrogen-bond occupancy is modest, with an average of 1.0 persistent bond maintained throughout the simulation. This suggests the binding is stabilized by a limited but key polar interaction, possibly supplemented by other forms of contacts.

2.5.3. Molecular Dynamic Simulation for 5U1X–Compound 21 Complex

The 5U1X–Compound 21 complex shown in Figure 16 exhibits excellent structural stability, characterized by minimal backbone deviation, a highly rigid binding site, and sustained ligand positioning. In Figure 16A, the backbone RMSD plateaus at approximately 0.15 nm after 20 ns, indicating rapid conformational equilibration and minimal structural drift. The ligand RMSD remains low at <0.1 nm, reflecting a tight, well-defined binding pose with negligible dissociation or positional shift. In Figure 16B, the RMSF analysis reveals very low residual fluctuations, with only slight increases in exposed loop regions around residues 1150–1200. The radius of gyration in Figure 16C is constant and averages around 2.00 nm with minimal fluctuation (1.96–2.04 nm), confirming a compact, stable tertiary structure without significant unfolding or expansion. The hydrogen-bond occupancy shown in Figure 16D shows a moderate and sustained trend, averaging 1.6 persistent bonds.

2.5.4. Molecular Dynamic Simulation for 5YA5–Compound 9 Complex

In the MD simulation of 5YA5–Compound 9 complex, the backbone RMSD (Figure 17A) equilibrates at approximately 0.15 nm, signifying rapid conformational convergence with minimal overall drift. The ligand RMSD remains consistently low (<0.1 nm), confirming a secure, well-defined binding pose without significant positional displacement. In Figure 17B, the RMSF analysis reveals generally low fluctuations (<0.5 nm), with slightly elevated mobility in exposed loop regions (notably around residues 1100–1150). The radius of gyration (Figure 17C) is tightly maintained, averaging 2 nm with negligible fluctuation around 96–2.04 nm, demonstrating a compact, stable tertiary structure with no major expansion or unfolding. In Figure 17D, the hydrogen-bond occupancy is moderate and stable, averaging 1.7 persistent bonds. This consistent polar interaction network, together with the exceptionally low ligand RMSD, suggests that binding is anchored by specific, well-maintained contacts and likely reinforced by additional hydrophobic or van der Waals interactions.

2.5.5. Molecular Dynamic Simulation for 5YA5–Compound 19 Complex

The MD simulation results of 5YA5–Compound 19 complex is shown in (Figure 18). The Backbone RMSD stabilizes at 0.20–0.25 nm after 20 ns, with low fluctuations thereafter, and the Ligand RMSD follows a similar pattern, converging to 0.18–0.22 nm and remaining stable. This implies that the complex reaches equilibrium quickly and maintains structural stability throughout the simulation with no major conformational drift or unfolding occurring. The Residue RMSF (Figure 18B) shows very low flexibility in most regions (<0.15 nm), with only a few localized peaks ( 0.4 nm) at specific residue numbers such as 1150–1200 and 1250–1300. This indicates that the protein backbone and binding pocket are rigid, with minimal local flexibility. The ligand-binding site remains conformationally stable. In Figure 18C, the radius of gyration fluctuates tightly around the average of 2.026 nm, with deviations < 0.04 nm over the entire trajectory indicating that the overall compactness of the protein–ligand complex is highly conserved, confirming no global unfolding or expansion during the simulation. The number of hydrogen bonds shown in Figure 18D between ligand and protein is consistently low, averaging 0.42 per frame, with most frames showing 0 or 1 bond (rarely 2) indicating that the ligand maintains only occasional or weak hydrogen bonding interactions.

2.5.6. Molecular Dynamic Simulation for 5U1X–Compound 19 Complex

In Figure 19A, the Backbone RMSD rises steadily from 0.25 nm to 1.1–1.25 nm over the simulation, with increasing fluctuations after 50 ns, and the Ligand RMSD remains low and stable ( 0.20–0.30 nm) throughout. This trend indicates that the protein backbone shows significant conformational drift and instability over time, suggesting partial unfolding or large-scale flexibility in loop/regions. The ligand, however, remains tightly bound and conformationally stable in its binding pocket. In Figure 19B, the Residue RMSF is generally low (<0.3 nm) in most regions, but shows sharp, high peaks ( 1.0–1.5 nm) at specific residue numbers such as around 300–350, indicating that the protein has well-defined rigid domains, but contains several highly flexible loop regions, which may be responsible for the Backbone RMSD drift. The radius of gyration (Figure 19C) starts at 3.0 nm, shows a gradual downward trend with large fluctuations, and stabilizes around an average of 2.786 nm in the last 20 ns. This implies that the overall protein compactness decreases slightly over time, consistent with partial relaxation or opening of flexible loops. The system does not unfold globally (no dramatic increase) but exhibits moderate structural rearrangement. In Figure 19D, the number of hydrogen bonds between ligand and protein is extremely low, averaging only 0.10 per frame indicating that the ligand maintains almost no persistent hydrogen bonding interactions throughout the simulation. Binding is dominated by hydrophobic contacts, $\pi \text{-}$stacking, and van der Waals interactions, with minimal polar stabilization.

2.5.7. Molecular Dynamic Simulation for 5EW3–Compound 3 Complex

For the MD simulation results of the 5EW3–Compound 3 complex, the Backbone RMSD (Figure 20A) starts at 1.0 nm, rises quickly to 1.4–1.6 nm within the first 10 ns, then plateaus with minor fluctuations around 1.5–1.7 nm until 40 ns. A sharp spike occurs at 45 ns (up to 5 nm), after which it returns to 1.5 nm and remains relatively stable with small oscillations (1.4–1.8 nm) through the end. The Ligand RMSD remains very low and stable ( 0.6–1.0 nm) throughout most of the trajectory, with only brief spikes coinciding with the backbone spike at 45 ns, then returns to baseline. The results indicate a protein structure that is reasonably equilibrated after the first few tens of ns, and the ligand remains tightly bound (low RMSD 0.8 nm average), even during the protein spike, indicating strong anchoring in the binding pocket. The sharp spike at 45 ns is likely a transient event—possibly a loop movement, partial unfolding in a flexible region, or simulation artifact. Importantly, the system recovers quickly and returns to stable RMSD, suggesting no irreversible unfolding or major disruption. The Residue RMSF (Figure 20B) is generally moderate (0.2–0.4 nm) across most of the protein. In Figure 20C, the radius of gyration starts higher ( 3.2 nm), decreases sharply to 2.8 nm by 20 ns, then fluctuates around the average of 2.859 nm with several large spikes (up to 3.4 nm). This trend indicates that the protein–ligand complex experiences noticeable global compaction early on, followed by breathing/expansion events (spikes in Rg), highlighting partial unfolding or significant flexibility in some domains, rather than a tightly stable, compact structure throughout. The number of hydrogen bonds (Figure 20D) between ligand and protein averages 1.81, with frequent oscillations between 1 and 3 bonds, indicating that the ligand forms moderately persistent hydrogen bonds (average 1.8).

2.5.8. Molecular Dynamic Simulation for 5YA5–Compound 3 Complex

In Figure 21A, the Backbone RMSD starts at 0.1 nm, rises quickly to 0.25–0.30 nm within the first 10 ns, then stabilizes with low-amplitude fluctuations around 0.25–0.30 nm for the remainder of the trajectory. The Ligand RMSD follows a very similar profile, equilibrating at 0.20–0.30 nm with minimal deviation. These trends indicate excellent structural stability, major conformational changes, or ligand dissociation. The Residue RMSF (Figure 21B) is generally low (<0.2 nm) across most of the protein, with moderate peaks (0.3–0.5 nm) at a few localized regions such as around residues 1150–1200 and 1250–1300. The radius of gyration plot in Figure 21C, shows that it fluctuates tightly around the average of 1.990 nm, with deviations typically < 0.03 nm (maximum excursion 0.05 nm), indicating that the protein–ligand complex maintains excellent global compactness throughout the simulation. In Figure 21D, the number of hydrogen bonds averages 1.27, with most frames showing 1 bond and frequent oscillations between 0, 1, and 2 bonds (rare spikes to 3), indicating that the ligand forms moderately persistent hydrogen bonds.

2.5.9. Systemic Toxicity Considerations of Polypharmacological Agents

We have to emphasize that this study does not assume direct therapeutic translatability of $\beta$-adrenergic-designed scaffolds to oncological targets. Rather, the cardiovascular origin of the scaffolds serves as a structural starting point, driven by shared intracellular signaling. Also, we note that the predicted interactions of the AI-generated scaffolds with diverse cancer targets may lead to unintended effects on cardiovascular, neuronal, or immune function. Indeed, broad polypharmacological activity, as suggested by our multi-target docking and MM-GBSA results, carries an inherent risk of systemic toxicity due to off-target engagement of non-oncological pathways. This underscores the necessity of subsequent experimental in vitro selectivity profiling and in vivo toxicity studies to determine therapeutic windows, beyond the computational feasibility and mechanistic plausibility presented here.

Hence, the computational study presented in this work does not claim therapeutic safety or efficacy of cardiovascular-derived scaffolds in oncology, but rather identifies hypothesis-generating interaction profiles to guide downstream prioritization and medicinal chemistry refinement under strict physicochemical and safety constraints.

3. Materials and Methods

Figure 22 shows the schematic block diagram of the virtual screening framework used in this work to computationally screen and profile the novel AI-generated virtual lead antihypertensive molecules. The inputs to the framework are the AI-generated molecules, the known cancer drugs, and the known cancer target proteins. The AI-generated lead drug molecules are analyzed for their docking scores, MM-GBSA, ADMET properties, and drug-likeness against the target proteins using a virtual screening platform.

3.1. Materials

3.1.1. Dataset of AI-Generated Virtual Lead Antihypertensive Molecules

This dataset contains the Simplified Molecular Input Line Entry System (SMILES) of 60 lead antihypertension molecules randomly selected from a set of molecules generated using the iFragGenMoS framework presented in [9]. In the iFragGenMoS, FDA-approved beta-blockers were used as seed molecules to generate the novel lead antihypertensive molecules. This random selection was performed to ensure the chosen subset for experimental validation remained structurally unbiased and preserved the broad chemical diversity and scaffold representation of the parent AI-generated library, rather than being guided by any prior property assumptions.

The iFragGenMoS framework [9] is illustrated in Figure 23. The fragment-and-build method was used to generate a hundred thousands molecules from a set of FDA-approved beta-blocker-like molecules, The built molecules are compared with the seed molecules, and the ones with Tanimoto scores < 1.0 and >0.65 are used to train a generative model and generate molecules. In [20], molecular pairs with Tanimoto scores $\ge 0.65$ are classified as similar in 2D. The choice of a Tanimoto similarity threshold between 0.65 and 1.0 is a strategic design decision. The upper bound (<1.0), excludes identical molecules (Tanimoto = 1.0), ensuring novelty and avoiding redundancy in the training set. On the other hand, the lower bound (>0.65) retains molecules that share meaningful structural similarity with the seed set, while still allowing sufficient chemical diversity for the generative model to explore new regions of chemical space. For the Tanimoto scoring we first converted the molecules to Molecular ACCess System key (MACCS Key) fingerprints; then, the RDKit Fingerprint Similarity (the Tanimoto coefficient equivalent) function is used to determine the similarity between molecules.

Further, we note that applying a generic $0.65$ cut-off without domain-specific justification can be overly arbitrary, particularly in polypharmacology or drug repurposing studies where structural similarity does not necessarily correlate perfectly with functional similarity. This is because Tanimoto similarity thresholds are inherently fingerprint-dependent and context-specific (e.g., agonist versus antagonist behavior, receptor versus enzyme targets). To statistically justify the use of a $0.65$ Tanimoto similarity threshold, we extracted a curated set of 29 known active ligands and 29 known inactives for the ${\beta }_1$-adrenergic receptor from the ChEMBL database. The inactive compounds were defined as experimentally confirmed inactive molecules using an IC₅₀≥$10\mathsf{\mu }\mathrm{M}$ cutoff. Using the MACCS keys fingerprinting method, we computed the intra-class Tanimoto similarity within the active set and the inter-class Tanimoto similarity between the active and inactive sets. The mean intra-class (actives) Tanimoto similarity score was $0.516$, whereas the mean inter-class (active–inactive) similarity score was $0.258$. The substantially higher similarity observed among known actives compared to the active–inactive comparisons demonstrates meaningful structural clustering of bioactive compounds. By conservatively setting the similarity threshold to $0.65$, which lies above the observed mean intra-class similarity and well above the inter-class similarity distribution, we effectively exclude the majority of inactive-like chemical space while preserving a structurally coherent relationship with known active ligands. This threshold therefore represents a statistically informed and chemically rational compromise between selectivity and structural fidelity.

For the generative model, a set of 13,313 built molecules was used to train a variational autoencoder (VAE)-based generative model. The VAE structure consists of an encoder, a decoder, and a latent space defined by the mean, ${\mathbf{z}}_{\mu }$, and standard deviation, ${\mathbf{z}}_{\sigma }$, of a Gaussian distribution.

Let $\mathbf{x}$ denote the one-hot encoded molecular representation and $\mathbf{z}$ the latent variable vector. We assume the joint distribution factorizes as

$$p(\mathbf{x},\mathbf{z})=p(\mathbf{x}\mid \mathbf{z})p\left(\mathbf{z}\right).$$

(1)

The objective of the VAE is to approximate the intractable posterior $p(\mathbf{z}\mid \mathbf{x})$ by introducing an encoder-defined variational distribution $q_{\mathit{\varphi }}(\mathbf{z}\mid \mathbf{x})$, parameterized by $\mathit{\varphi }$, which is modeled as a multivariate Gaussian distribution,

$$q_{\mathit{\varphi }}(\mathbf{z}\mid \mathbf{x})=\mathcal{N}\left(\mathbf{z};{\mathit{\mu }}_z\left(\mathbf{x}\right),diag\left({\mathit{\sigma }}_z^2\left(\mathbf{x}\right)\right)\right),$$

(2)

where ${\mathit{\mu }}_z\left(\mathbf{x}\right)$ and ${\mathit{\sigma }}_z\left(\mathbf{x}\right)$ denote the learned mean and standard deviation of the latent variables, respectively.

During training, latent variables $\mathbf{z}$ are sampled from $q_{\mathit{\varphi }}(\mathbf{z}\mid \mathbf{x})$ and decoded through the likelihood model $p_{\mathit{\theta }}(\mathbf{x}\mid \mathbf{z})$, where $\mathit{\theta }$ denotes the decoder parameters. The loss function used to estimate $\mathit{\varphi }$ and $\mathit{\theta }$ is the evidence lower bound (ELBO),

$$\mathcal{L}(\mathit{\varphi },\mathit{\theta };\mathbf{x})={\mathbb{E}}_{\mathbf{z}\sim q_{\mathit{\varphi }}(\mathbf{z}\mid \mathbf{x})}\left[\log p_{\mathit{\theta }}(\mathbf{x}\mid \mathbf{z})\right]-D_{\mathrm{KL}}\left(q_{\mathit{\varphi }}(\mathbf{z}\mid \mathbf{x})\parallel p\left(\mathbf{z}\right)\right),$$

(3)

where the first term represents the reconstruction loss and the second term corresponds to the Kullback–Leibler divergence, which regularizes the latent space by enforcing proximity to the prior distribution $p\left(\mathbf{z}\right)=\mathcal{N}(\mathbf{0},\mathbf{I})$.

The VAE is implemented with 80 × 24 input dimension and latent dimensions of 8. The VAE encoder contains two hidden layers with 1120 and 560 neurons, and the decoder is symmetric. We trained the model using the loss function expressed in (3). The output (generated) molecules were evaluated and those that are valid, unique, and beta-blocker-like were filtered in. The output of this process is a set of 123 unique molecules generated by the VAE-based generative model, from which the sixty molecules used in this paper were randomly selected. Samples of the dataset are shown in Figure 24.

3.1.2. Dataset of Known Cancer Drugs

This dataset contains the SMILES of some cancer drug small-molecules that are either approved or in clinical trials. We categorized the drug molecules into groups based on their common target proteins, shown in

Table 6. Target categories and corresponding drugs.

Target Category	Drugs
P2X7	AZ10606120; Brilliant Blue G; GSK1482160; JNJ-54175446; JNJ-47965567; Probenecid; AZD9056; JNJ-42253432; AZ11645373; CE-224535
EGFR	Gefitinib; Vandetanib; Mobocertinib; Lazertinib; Lapatinib; Dacomitinib
VEGFR2	Axitinib; Pazopanib; Sunitinib; Sorafenib; Lenvatinib; Cabozantinib; Regorafenib; Tivozanib; Fruquintinib
c-MET	Crizotinib; Cabozantinib; Capmatinib; Tepotinib; Foretinib; Savolitinib
PDGFRA	Imatinib; Avapritinib; Sunitinib; Sorafenib; Nilotinib; Ripretinib

. Target proteins that are typically associated with cancer, such as Epidermal Growth Factor Receptor (EGFR), Vascular Endothelial Growth Factor Receptor 2 (VEGFR2), cellular Mesenchymal–Epithelial Transition Factor (c-Met), Platelet-Derived Growth Factor Receptor Alpha (PDGFRA), and Purinergic Receptor P2X, Ligand-Gated Ion Channel 7 (P2X7), are considered in this work.

3.1.3. Dataset of Known Cancer Target Proteins

This dataset (shown in

Table 7. Protein structures with PDB identifiers and target categories.

PDB Identifier	Target Category
5U1X	P2X7
7U9A	EGFR
5EW3	VEGFR
5YA5	c-MET
8XRR	PDGFRA

) contains the identifiers that represent the categories of the target proteins associated with cancer. The 5YA5 is the PDB entry for cyclin-dependent kinase 2 (CDK2) bound to an inhibitor [21]. CDK2 is a protein kinase that plays a crucial role in regulating the cell cycle, which is often overexpressed or dysregulated in many cancers, leading to uncontrolled cell proliferation [22]. The inhibition of its activities can halt uncontrolled cell proliferation.

The 5U1X is a crystal structure of the ATP-gated P2X7 ion channel bound to allosteric antagonist JNJ47965567 [23]. Its activation by extracellular ATP can lead to changes in cancer cell signaling, proliferation, metastasis, and even immune evasion. Hence, its inhibitors are considered promising anticancer therapeutics. The 7U9A is an EGFR in complex with a macrocyclic inhibitor [24] and like many EGFRs could be considered a representation of a major target in cancer treatment [25]. In many cancers, the EGFR becomes abnormally activated due to gene amplification, point mutations, or other alterations that lead to increased cell proliferation, angiogenesis, and metastasis. Hence, its selective inhibition can help in cancer treatment.

The 5EW3 is a human VEGFR2 kinase domain in complex with AAL993 [26]. Its overactive activation by Vascular Endothelial Growth Factor (VEGF-A) in tumor cells triggers a signaling cascade within the cancerous endothelial cell, leading to the formation of a dense network of new blood vessels that feed the tumor. Hence, it has become a major target for the development of antiangiogenic drugs. The 8XRR is a complex structure of PDGFRA with an inhibitor RH140 [27]. The 8XRR is a receptor tyrosine kinase (RTK) whose overreaction or mutation in cancer cells leads to tumor growth and angiogenesis. Hence, its inhibition can offer treatment to various cancers.

3.1.4. Protein and Ligand Preparation

Each of the target proteins was prepared using the protein preparation Wizard of the Schrödinger Suite Release 2025-3 [28]. For preparing a protein, the Protein Preparation Wizard module in Maestro version 13.6 automates the complex process to ensure the protein structure is chemically and sterically correct for subsequent calculations. The PDB file of the protein is first loaded into Maestro, and the wizard automatically identifies issues such as missing atoms, bond orders, or unexpected bond angles, and it assigns correct bond orders to the protein and adds all missing hydrogen atoms. A physiological pH of 7.0 is used to determine the protonation states of ionizable residues. The water molecules that are not in the active site are removed. The wizard corrects any errors in the orientations of side chains that may have been introduced during the crystallization process. Finally, the OPLS4 force field was used to execute restricted energy reduction, which relieved steric conflicts and optimized the protein’s local shape. Then, the LigPrep tool in Schrödinger was used to prepare the ligands. At physiological pH (7.0 ± 0.2), Epik was used to construct various ionization states and tautomeric forms by adding all hydrogen atoms. The desalting option was used to eliminate counterions while retaining just neutral ligand forms. Hydrogen-bonding networks were optimized by using PROPKA to estimate the pKa values of ionizable groups, resulting in precise protonation states for subsequent docking simulations.

3.1.5. Ligand Profiling

The profiling of the ligands on their ADMET and drug-likeness proterites is done using Schrödinger’s QikProp tool. All the ligands were prepared in neutralized form for the calculation of pharmacokinetic properties by QikProp. It computes the pharmacokinetic properties and descriptors such as octanol/water partitioning coefficient, aqueous solubility, brain/blood partition coefficient, intestinal wall permeability, plasma protein binding and others. The essence of performing the ligand profiling at this stage is to enrich the compound library with candidates possessing favorable pharmacokinetic and safety characteristics, thereby reducing attrition in downstream analyses.

3.1.6. Binding Free Energy Estimation

To be able to dock the ligands on the proteins, the binding site for each ligand on the protein of interest was created using Schrödinger’s Glide Receptor Grid Generation tool. The prepared ligands were docked on the grid using the Virtual Screening Wizard. Van der Waals radii of nonpolar atoms were scaled by 1.0, with a partial charge cut-off of 0.25. All other settings were left alone, and the docking process was run in extra precision (XP) mode and the MM-GBSA. The essence of performing molecular docking followed by MM-GBSA binding free energy calculations as secondary filter is to estimate the thermodynamic favorability and achieve a significantly higher confidence in the quantitative binding ranking of our lead compounds.

3.1.7. Polypharmacological Potential Analysis

To quantitatively assess the polypharmacological potential of the molecules, we employ a method that integrates MM-GBSA binding free energy scores across the cancer-associated protein targets. The rationale here is the MM-GBSA provides a more accurate estimate of ligand binding affinity than simple docking scores. By comparing these energetically refined scores across multiple targets, we can identify compounds with a high likelihood of engaging a desired subset of proteins, a hallmark of designed polypharmacology.

It is important to state that MM-GBSA estimates do not represent absolute experimental binding affinities, as they are influenced by force-field parametrization, implicit solvent approximations, entropic truncations, and sampling limitations. Nevertheless, MM-GBSA has been widely validated as a robust post-docking approach for comparative ranking of ligands across multiple targets, particularly when a consistent computational protocol is applied. In the context of polypharmacological prediction, where the objective is to identify compounds exhibiting balanced and concurrent binding tendencies across a panel of cancer-relevant proteins, MM-GBSA scores are sufficient to capture relative interaction strengths and cross-target consistency, thereby enabling reliable prioritization of multi-target lead candidates.

We express the term polypharmacological potential scores (PPSs) to polypharmacologically profile the multi-target potential of a molecule. The PPS is a composite score of the Span score (SS) and the Average Potency Score (APS). The SS is the number of targets for which a compound’s MM-GBSA score is lower than a stringent affinity threshold. Let $N_T$ denote the total number of evaluated targets, $\Delta G_{bind}^{(c,i)}$ the MM-GBSA binding free energy of compound c against target i, and $\Delta G_{\mathrm{thr}}$ a stringent affinity threshold. The ${\mathrm{SS}}^{\left(c\right)}$ is defined as

$${\mathrm{SS}}^{\left(c\right)}=\sum _{i=1}^{N_T}\mathbb{I}\left(\Delta G_{\mathrm{bind}}^{(c,i)}\le \Delta G_{\mathrm{thr}}\right),$$

(4)

where $\mathbb{I}(·)$ is the indicator function

$$\mathbb{I}\left(x\right)=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}x\mathrm{is}\mathrm{true},\hfill \\ 0,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(5)

The ${\mathrm{SS}}^{\left(c\right)}$ ranges from 0 to $N_T$, with higher values indicating stronger multi-target binding behavior and increased polypharmacological potential.

The APS is the mean potency (in positive kcal/mol) for the targets a compound effectively hits. It can be expressed as

$${\mathrm{APS}}^{\left(c\right)}={\displaystyle \frac{1}{{\mathrm{SS}}^{\left(c\right)}}}\sum _{i=1}^{N_T}\mathbb{I}\left(\Delta G_{\mathrm{bind}}^{(c,i)}<\Delta G_{\mathrm{thr}}\right)\left(-\Delta G_{\mathrm{bind}}^{(c,i)}\right),$$

(6)

A high APS indicates compounds with strong, balanced affinity against a set of specific targets.

3.1.8. Molecular Interaction Analysis

Molecular interaction analysis performed include ligand interaction mapping and pharmacophore modeling. The ligand interaction diagram is used to identify the specific, persistent, and crucial molecular/atomic determinants responsible for a ligand’s high affinity. This information is vital for future lead optimization. Further, pharmacophore modeling generalizes these findings into abstract chemical features such as H-bond donors/acceptors, hydrophobes, and aromatics that define the essential requirements for receptor recognition. This allows one to design or search for structurally diverse molecules that maintain critical binding interactions. Combining the two processes at this stage provides a detailed chemical blueprint of the binding interaction that can be used to strategically design more potent derivatives.

3.1.9. Molecular Dynamics Simulation

While the molecular docking/MM-GBSA studies identify the best binding pose and estimate binding affinity, MD simulations are used to observe and analyze the stability, fluctuations, and conformational changes of the molecules over time. In this study, a duration of 100 ns and a step size of 2 fs are used to run the MD simulation. Advanced metrics, such as root mean square deviation (RMSD), root mean square fluctuation (RMSF), radius of gyration, and hydrogen bonds, are calculated. In all MD simulations, the temperature was maintained at T = 300 K and the pressure at 1.01325 bar. MD simulation serves as a stability and dynamic behavior validation tool for the top-ranked ligand–protein complexes, thereby, placing high-confidence on the lead candidates.

We note that the choice of 100 ns for the MD simulations is to evaluate short-term stability and interaction persistence, not residence time or kinetic trapping, which would require extended simulations or enhanced sampling approaches.

4. Conclusions

In this study, we computationally investigated the polypharmacological anticancer potential of AI-generated antihypertensive virtual lead compounds by re-evaluating their interaction profiles across multiple cancer-associated protein targets. By integrating drug-likeness and pharmacokinetic screening with molecular docking, binding free energy estimation, pharmacophore modeling, and molecular dynamics simulations, we established a systematic in silico framework for uncovering latent multi-target activity within a chemically focused virtual lead library originally optimized for $\beta$-adrenergic receptor modulation. Our findings suggest that several of the hypothetically evaluated compounds exhibit favorable physicochemical properties, acceptable predicted pharmacokinetic and safety profiles, and consistent predicted binding affinities across multiple cancer-relevant targets. Importantly, binding free energy analyses and dynamic simulations provide support for the stability and persistence of key ligand–target interactions, increasing the plausibility of polypharmacological engagement rather than isolated target-specific effects. Collectively, these in silico results generate a robust hypothesis that select AI-designed compounds possess a polypharmacological profile relevant to oncology. These results underscore the utility of computational re-profiling strategies in expanding the hypothesized therapeutic relevance of AI-designed chemical libraries and advancing polypharmacological drug discovery paradigms. To test this hypothesis, some of the prioritized molecules identified in this work are currently being synthesized for crystallization purposes and eventual biological testing against cancer cells.