A list of pan-PI3K and selectively targeting PI3K-alpha drugs was prepared and sourced from ChEMBL [47], Guide to Pharmacology [48], and DrugBank [49] databases, and then further removing duplicates to initialize a set of 25 distinct effective drugs. For the natural product dataset, comprehensive database repositories are acquired through the COCONUT (Collection of Open Natural Products) [50] database which had 407,270 natural compounds (Table 1). A molecular weight (MolWt) of 300–600 Da was taken as the size range to work further with the dataset, to eliminate non-druglike larger compounds and smaller compounds with a lack of useful scaffolds, for further processes.

The crystal structure of wild-type PI3K-alpha subunit inbound complexed with an inhibitor taselisib (2-methyl-2-(4-{2-[3-methyl-1-(propan-2-yl)-1H-1,2,4-triazol-5-yl]-5,6-dihydroimidazo[1,2-d][1,4]benzoxazepin-9-yl}-1H-pyrazol-1-yl)propanamide) having a resolution of 1.99 Å [PDB (Protein Data Bank) ID: 8EXL] [51] from the PDB was prepared using Protein Preparation Wizard in GLIDE in Schrodinger Maestro Suite 2023-1 [52], by filling missing side chains, removing water beyond 8 Å and restrained minimization in OPLS4 (Optimized Potentials for Liquid Simulations 4) force field, optimize overlapping hydrogens and replacing them, capping of termini, converting selenomethionines to methionines, filling in missing loops using Prime [53], generating het states at pH of 2.0 with Epik, create disulfide bonds and zero-order binding to metals. This is performed to rectify structural aberrations and optimize the protein structure for subsequent molecular docking studies.

Furthermore, due to the lack of good quality protein structures (parameters such as low resolution, similar co-crystal, similar method of structure characterization) available for mutational models, mutations of single residues were introduced in the copies of the prepared proteins in Maestro (Schrodinger Suite 2023-1). Three mutated proteins were created having mutations—H1047R, E545K, and E542K and then minimized again using Protein Preparation Wizard. Receptor grid generation was performed with a van der Waals scaling factor of 1.0 and partial charge cut-off of 0.25, using GLIDE to prepare the docking grid in the binding pocket of the protein, which facilitates structure-based ligand docking within the prepared protein binding site.

Drug and natural ligand structures were prepared using LigPrep in Schrodinger Suite 2023-1 [54] to generate optimized 3D conformations suitable for molecular docking analyses and virtual screening. A maximum ligand size of 500 atoms, OPLS4 force field, at the target pH of 7 ± 2, using Epik, and generated at most 1 stereoisomer at most per ligand with retention of specified chiralities, during LigPrep. All drugs underwent this preparation process leading to a total of 213,343 prepared natural ligands, while the selected subset of natural ligands from COCONUT underwent this preparation to produce 50 structures. The drugs were then docked against the receptor grid of the PI3K-alpha protein grid, initially using high throughput virtual screening (HTVS) docking, then standard precision (SP) docking in GLIDE on default settings. This was followed by refinement using extra precision (XP) docking in the same settings, for improved accuracy, to determine best poses of the ligand drugs. Simultaneously, natural ligands were made to undergo HTVS with GLIDE, and 153,409 compounds were docked. Those in the top 1% (1,534) with the lowest docking scores from HTVS underwent further refinement through SP docking. Finally, the top 10% (153) of this refined subset based on SP docking scores underwent XP docking for additional precision. This method, thus, ensured a thorough exploration of drug and natural ligand interactions with the PI3K-alpha protein, enhancing the identification of potential candidates with strong binding affinities and therapeutic promise.

In the fragmentation process, the docked drugs and ligands were broken down using the Schrodinger Python [55] utility of fragment_molecule.py with the carbon hetero option. This utility applied specific rules to ensure precise fragmentation and the bonds within the molecules were selectively cut, prioritizing two main types: bonds associated with rings, except for hydrogen bonds and internal ring bonds (with allowances made for bonds between ring-carbon and non-carbon atoms using the -c option), and carbon-carbon bonds that were preserved unless they are a part of a ring structure. The carbon hetero option enabled cutting bonds between ring-carbon atoms and hetero atoms, potentially increasing the variety of generated fragments. Additionally, Murcko fragments of the drugs were generated. This method divides compounds into components such as R-groups, linkers, and ring systems. From each compound, a scaffold or framework is derived by removing all R-groups while retaining core ring systems and linkers. This systematic approach categorizes the atoms of each drug molecule into ring, linker, framework, and side chain components, offering a thorough breakdown of structural data essential for detailed analysis and interpretation. This process resulted in four categories of fragments, drug fragments with carbon hetero setting and Murcko scaffolds of drugs, and natural compounds fragments and Murcko scaffolds similarly. Fragments of natural compounds were further filtered out by keeping fragments below 300 Da MolWt such that the hybridization process is streamlined and does not produce large fragments which are unlike small molecules.

In the hybridization process, fragments extracted from both drugs and natural ligands were merged using the BREED module in Maestro to create novel hybrid entities. The Murcko scaffolds of drugs were combined with fragments from the natural ligands (Category 1), and vice versa, Murcko scaffolds of natural ligands were combined with fragments from drugs (Category 2). This was performed to ensure that the combination of core structures of one category are hybridized with functional groups of the other, thereby ensuring the success of the hybrids by improving on existing features of drugs and natural ligands. The process involved aligning the structures of ligand-bound targets and identifying overlapping bonds across ligand pairs. Fragments on each side of these overlapping bonds were then exchanged to generate a new single generation of molecules, following established methods described in the literature [56] with default parameters which are maximum atom-atom distance of 1.0 Å and maximum 15 degrees angle of distance in bond overlap criteria. Specifically, fragments from one ligand could replace fragments from another if the bond connecting them in the first ligand matched a bond in the second ligand. Constraints were applied to maintain bond order consistency and to ensure that swapped bonds were not part of a ring structure. Each overlapping bond between pairs of input molecules resulted in the creation of a new pair of molecules, with duplicates removed. This iterative swapping process constituted one “generation” of hybridization, ultimately yielding a total of 3,744 unique natural-drug hybrids.

Following the creation of hybrids, they underwent HTVS to identify promising candidates, and 2,190 and 1,479 hybrids were docked from Category 1 and Category 2 respectively. The top 10% of hybrids from Category 1 (219) and Category 2 (147) with the most favorable docking scores were chosen for SP docking. Subsequently, the top 10% of hybrids from Category 1 (22) and Category 2 (15) from SP docking underwent XP docking. The top 5 hybrids from each category were further docked at SP followed by XP in the mutant protein structures (H1047R, E545K, and E542K). The total ten hybrids were analyzed using molecular mechanics/generalized Born and surface area (MM/GBSA) solvation to evaluate their binding affinities with the PI3K-alpha protein wild-type and mutants using the Prime module.

The validation of the docking protocol has been performed by re-docking the co-crystal ligand in the receptor at the same site, measuring the root mean square deviation (RMSD) in the poses between the original co-crystal and the docked molecule [57, 58]. The co-crystal inbound receptor taselisib is docked in the receptor grid of the protein and RMSD value is checked at SP and XP docking. A small RMSD value (< 1.5–2 Å) and superposition indicate that the docking strategy applied is working correctly [59–61].

Induced fit docking (IFD) in the wild-type receptor of top hybrids from MM/GBSA was performed to analyze the binding of the hybrids created. IFD ensures that the docking evaluated binding in a flexible protein-rigid ligand binding conformation, unlike a rigid protein-flexible ligand sampling during molecular docking analyses [62–65]. SP protocol for pose generation was used at 0.5 van der Waals scaling. The hybrids showing comparable scores to the docked ligands should be selected for further studies.

QikProp module in Schrodinger Suite 2023-1 [66], was utilised to predict the pharmacokinetic properties—absorption, distribution, metabolism, and excretion (ADME), of the shortlisted hybrids. It predicts pharmaceutical and physical descriptors of molecules, on the basis of known druglike parameters of compounds. The prediction and calculation of ADME properties are essential due to their importance in the physiological systems. The good druglike parameters, QPlogS (aqueous solubility), QPlogPo/w (octanol/water partition coefficient), hydrogen bond donors and acceptors (donorHB, accptHB), MolWt, polar surface area (PSA), and solvent accessible surface area (SASA) are important parameters, based on Lipinski’s rule of five. The percentage of human oral absorption (%HumOralAbs) was also predicted to determine the absorption capacity of the hybrids.

The danger posed by high expression of the p110α subunit of PI3K has been studied variably in breast cancer, amounting to about 30–40% of breast cancer cells presenting elevated and upregulated expression of PI3K [67]. On studying the binding mechanism of PI3K-alpha, it was revealed that the ligand binding and kinase-catalytic domain of the protein lies in the amino acid residue ranges 765–1,051, wherein key oncogenic driving mutation—H1047R also occurs. The inbound co-crystal ligand displayed binding and interactions with the residues in the ligand binding domain, on docking. Three main loops form the functional structure of the domain, along with the PI3K/PI4K conserved region from 801–815. These are the: G-loop which ranges from 771–777 residues (IMSSAKR), catalytic loop which ranges from 912–920 (GIGDRHNSN) which is conserved, and the activation loop which lies in 931–957 (HIDFGHFLDHKK-KKFGYKRERVPFVLT) [68, 69].

213,343 natural ligands were prepared and 50 structures of the drugs were generated by LigPrep. The best docking scores of –9.697 kcal/mol by alpelisib, –10.471 kcal/mol by taselisib, and –12.121 kcal/mol by GSK-615, were achieved in HTVS, SP, and XP docking of drugs, respectively. On the other hand, the best docking score of –10. 527 kcal/mol by CNP0097211 (3-(3,4-dihydroxyphenyl)-2-({3-[7-(3,4-dihydroxyphenyl)-3,4-dihydroxy-5,6-dihydronaphthalen-1-yl]prop-2-enoyl}oxy)propanoic acid), –11.353 kcal/mol by CNP0419682 (5-acetyl-7-(5-carboxy-7-hydroxy-4-oxo-4H-chromen-2-yl)-2,3-dihydroxy-6-methyl-9-oxo-9H-xanthene-1-carboxylic acid) and –17.190 kcal/mol by CNP0093692 (2-(3,4-dihydroxyphenyl)-5,8-dihydroxy-7-{[4-hydroxy-4-(hydroxymethyl)-3-{[3,4,5-trihydroxy-6-(hydroxymethyl)oxan-2-yl]oxy}oxolan-2-yl]oxy}-4H-chromen-4-one), were achieved in HTVS, SP, and XP docking of natural ligands, respectively.

On fragmentation of drugs, 25 Murcko scaffolds and 1,372 fragments were generated. For natural ligands, 151 Murcko scaffolds and 49,114 fragments were produced and 4,894 were retrained for further analysis. Hybridization of Murcko scaffolds of drugs with fragments of natural ligands, produced 2,225 hybrids (Category 1), whereas hybridization of Murcko scaffolds of natural ligands and fragments of drugs produced 1,519 hybrids (Category 2). The top 10 hybrids were selected based on further virtual screening (Figures S1–S10).

The best docking scores of –10.614 kcal/mol by CNP0302777 (frag 7) + 23582824 (frag 1) and –11.080 kcal/mol by CNP0171717 (frag 5) + 23582824 (frag 1) were achieved in HTVS and SP docking of Category 1, respectively. On the other hand, the best docking scores of –10.055 kcal/mol by CNP0382214 (frag 1) + 72709209 (frag 22) and –11.112 kcal/mol by 50990924 (frag 15) + CNP0475782 (frag 1) were displayed in HTVS and SP docking of Category 2, respectively. The best scores in XP docking by the two categories were –13.354 kcal/mol by 23582824 (frag 1) + CNP0243567 (frag 44) (NH-01) from Category 1 and –12.670 kcal/mol by 56649450 (frag 59) + CNP0392152 (frag 1) (NH-06) from Category 2. The top 5 hybrids based on docking score from each category were selected for further analysis (Table S1). The molecular docking and MM/GBSA analysis of the mutants of PI3K-alpha showed that the hybrids showed slightly better binding affinity for mutant proteins as compared to wild-type (Table S1). NH-01 performed the best in binding with all three mutant protein receptors, followed by NH-06 similar to the wild-type. However, the ranking of the structures was different when compared with the docking ranking against wild-type receptors.

On docking of the originally bound ligand, taselisib, a docking score of –9.571 kcal/mol on XP was achieved, with an RMSD value of 0.4307, thereby waiving a green flag for the screening strategy. The IFD corroborated the results achieved in XP docking. The IFD docking scores were in a different order from the XP docking but displayed comparable results. MM/GBSA revealed high negative free energy (dG) release on the binding of ligands (Table 2 and Table S1).

The best hybrid molecule from Category 1 (NH-01), displayed several interactions with the binding site residues. Three hydrogen bonds with VAL851, TYR836, and ASP810 are a part of the catalytic domain, while ASP810 additionally is a residue in the conserved domain. A charged interaction with residues like ASP810, LYS802, TYR836, and LEU807 was exhibited by the hybrid, while most of the interaction was hydrophobic with residues lying between ILE848 to GLN859 (Figure 1). In IFD, two more pi-pi stacking interactions were observed with TYR836 and TRP780, with a more polar interaction with residues lying from ASN853 to GLN859 (Figure 2). The hybrid shows similar pi-pi stacking with TYR836 as the drug GSK-615 from which the Murcko scaffold is derived and hydrogen bonding at VAL851. Whereas the hybrid exhibited the interactions of the parent flavonoid-7-O-glucuronide natural derivative (CNP0243567), as a hydrogen bond at ASP810, TYR836, and VAL851 (Figure 3).

The best hybrid molecule from Category 2 (NH-06) forms a salt bridge with GLU768 and GLU798 from the imidazole ring end of the hybrid, while two hydrogen bonds with VAL851 are observed. These interactions lie in the G-loop domain of the protein. Hydrophobic interactions from ILE848 to VAL851 were observed, while charged positive and polar interactions were seen with SER854 to ASN853, MET772 to SER774, and LYS802 to ILE800 (Figure 4). No changes were observed in interactions in the best IFD pose (Figure 2). The natural scaffold of CNP0392152 displayed a similar salt bridge with GLU768 and GLU798, while the drug fragment is from alpelisib and displays a similar hydrogen bond with VAL851 (Figure 3).

Thus, these interactions indicate a good binding potential of the hybrids with very low bad clashes and no ugly interactions according to Maestro, increasing the affinity and selectivity of natural hybrids and reducing the risk potential of the drugs.

The desired values of the ADME parameters predicted in the pharmaco-analysis should be such that the MolWt lies between 130.0–500.0 Da, donorHB should be between 0–5, accptHB should be between 2–10, the predicted QPlogPo/w should lie in the range of –2.0–6.5 mol dm^–3, the predicted QPlogS should be preferably within –6.5–0.5 mol dm^–3, the %HumOralAbs should be more than 70%, the desirable range of values of PSA is 7.0–200.0, the SASA should be 300.0–1,000.0 and the violations of rule of five should be a maximum of 4.

The 10 hybrids were predicted to have good druglike parameters. Only one hybrid NH-09 exhibited two violations of rule of five, while all the other hybrids showed 0 violations, therefore, displaying good characteristics (Table S1). This makes it evident that the safety and ADME profile of parent molecules are highly increased when they are hybridized (Table 3 and Table S2).