Template-based modeling of insect odorant receptors outperforms AlphaFold3 for ligand binding predictions Authors: Amara Jabeen, John Graham Oakeshott, Siu Fai Lee, Shoba Ranganathan, Phillip W. Taylor Published: 2024 ABSTRACT Insects rely on odorant receptors (ORs) to detect and respond to volatile environmental cues, so the ORs are attracting increasing interest as potential targets for pest control. However, experimental analysis of their structures and functions faces significant challenges. Computational methods such as template-based modeling (TBM) and AlphaFold3 (AF3) could facilitate the structural characterisation of ORs. This study first showed that both models accurately predicted the structural fold of MhOR5, a jumping bristletail OR with known experimental 3D structures, although accuracy was higher in the extracellular region of the protein and binding mode of their cognate ligands with TBM. The two approaches were then compared for their ability to predict the empirical binding evidence available for OR-odorant complexes in two economically important fruit fly species, Bactrocera dorsalis and B. minax. Post-simulation analyses including binding affinities, complex and ligand stability and receptor-ligand interactions (RLIs) revealed that TBM performed better than AF3 in discriminating between binder and non-binder complexes. TBM’s superior performance is attributed to hydrophobicity-based helix-wise multiple sequence alignment (MSA) between available insect OR templates and the ORs for which the binding data were generated. This MSA identified conserved residues and motifs which could be used as anchor points for refining the alignments.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-024-80094-x. Introduction Structural modeling is becoming a powerful tool for protein characterization. One such approach, homology modeling, also known as template-based modeling (TBM), could be applicable to any type and size of protein that has a structural homolog1 in Protein Databank (PDB)2. Other methods such as Raptor-X, AlphaFold (AF), trRosetta, MULTICOM, and C-I-TASSER3 have also become more powerful in recent years through their incorporation of deep learning. In particular, AlphaFold3 (AF3)4 is an updated version of AF25 based on end-to-end deep learning and utilizes structural data in PDB as well as multiple sequence alignment (MSA) to develop atomic resolution models. However, some families of membrane-spanning proteins such as the G-protein-coupled receptors (GPCRs) have remained problematic for both TBM and AF because experimental templates are scarce and high levels of accuracy are needed to predict binding outcomes for a range of potential ligands6. One group of membrane-spanning proteins for which reliable structural models would be particularly useful is the insect odorant receptors (ORs). These proteins are key components of insects chemosensory systems7 and potential targets for developing sustainable pest management strategies8. Like human ORs, which are GPCRs, insect ORs have seven canonical transmembrane domains but their topology is inverted compared to human ORs, with an intracellular N-terminus and an extracellular C-terminus. Also, mature odorant receptors in insects are heteromers composed of a variable ligand-specific OR and a highly conserved coreceptor called the OR co-receptor (Orco, also known as OR83b)9,10, and the residue composition and orientation of the seven transmembrane helices in the OR ligand binding site determine their ligand specificity. The number of ligand-specific ORs varies from about 40 to 80 in different insect species11 and many show substantial sequence variation, both between orthologs in different species and between paralogs within species12. Recently, multiple insect OR structures (Table S1) have been determined by cryogenic electron microscopy (cryo-EM). The functional homotetrameric structure of MhOR5 from the jumping bristletail Machilis hrabei13 has been elucidated, while very recently heteromeric structures for several other ligand-specific ORs have been reported, where the OR in question is complexed with three Orco subunits. These latter complexes, OR10 from Aedes aegypti, OR28 from Anopheles gambiae14 and OR5 from Acyrthosiphon pisum15, like MhOR5, have been solved in both their ligand-free (apo) and ligand-bound (holo) states. The structure of a homomeric in vitro functional Orco from a fig wasp, Apocrypta bakeri16 has also been resolved. Empirical functional analyses of insect ORs have relied heavily on their heterologous expression in Xenopus laevis and, recently, also CRISPR-Cas9 knock-out technology10,17 One pattern emerging from such studies is that insect odor recognition is a combinatorial process, akin to mammalian olfactory receptors; a single odor can often be perceived by multiple ORs and a single OR can perceive multiple odors, but with quantitative differences in binding affinities in both cases18,19 Site-directed mutagenesis (SDM) has also been used to identify particular residues impacting on ligand binding and ligand selectivity/specificity in specific ORs20–26. Although these studies also identify more distant residues that act allosterically27, several of the residues implicated are consistent with the experimental structural data, suggesting that helices 1 to 6 (predominantly 2, 3, 4, and 6, with 1 and 5 in some cases) are key transmembrane domains involved in ligand binding and can contain residues interacting directly with the ligands. An earlier study28 reported a TBM model for MhOR5 using iBio-GATS with Orco (PDBID: 67C0) as a template. An initial comparative analysis indicated that TBM outperformed AF2 in modeling the MhOR5 structure. Here we present a more detailed quantitative comparison of an optimized MhOR5 TBM with AF3 models, including docking followed by simulations with diagnostic ligands and quantitative analysis of receptor-ligand interactions (RLIs). The TBM has been optimized using a hydrophobicity-based helix-wise multiple sequence alignment (MSA) based on extensive sequence datasets available for the OR family of two important fruit fly pests, the oriental fruit fly Bactrocera dorsalis and Chinese citrus fly B. minax29,30. Of the numerous volatile compounds tested for their potential as lures or repellants for various fruit fly species, methyl eugenol (ME), raspberry ketone (RK) and a more volatile analog of the latter (cuelure) have been used to attract B. dorsalis, B. minax and a third species, B. tryoni, respectively, to insecticide-laced traps31–33. The interactions of certain B. dorsalis and B. minax ORs and these specific ligands have also been characterized experimentally17,20,34–37. We first computed the ability of TBM and AF3 to predict the experimentally established fold and ligand-binding site of MhOR5. We then evaluated their capability to predict whether five B. dorsalis and three B. minax ORs would bind to a panel of five odorants for which quantitative experimental data on the responsiveness of the eight ORs in question were available. These tests involved initial docking studies to develop the starting coordinates for subsequent molecular dynamics (MD) analyses. The MD generated quantitative energy and stability statistics, identifying specific residues that could be involved in RLIs. The results could then be evaluated against the empirical data on the RLIs. Results and discussion Docking studies for TBM and AF3 models of MhOR5 At present, nine different experimentally resolved OR structures are available in the PDB from different insect species (Table S1). Phylogenetically, AaOR10, AgOR28, and MhOR5 are more closely related to one another and many other ORs in the model insect Drosophila melanogaster than ApOR5 (Figure S1). Moreover, AgOR28, AaOR10, and ApOR5 are missing two or more ligand binding residues that have been confirmed through site-directed mutagenesis of D. melanogaster and B. dorsalis ORs (Figure S2)20–22,38. We further note that experimental structures of MhOR5 showed strong binding affinities (as predicted by binding free energies) for their native ligands compared to AgOR28, AaOR10, ApOR5 structures with their respective native ligands (Table S2 and see below). We therefore selected MhOR5 as the preferred structural template for insect OR TBM models. Our cross-validation of TBM and AF3 modeling of MhOR5 as a template against the experimental data first involved testing the ability of the two tools to reconstruct the experimental MhOR5 structure. The three experimentally resolved structures available for MhOR5 in PDB are the apo-structure (PDBID: 7LIC), along with complexes with eugenol (EOL) (PDBID: 7LID) and N, N-diethyl-m-toluamide (DEET) (PDBID: 7LIG)13. EOL is a natural product insecticide and DEET a synthetic repellant39,40. We compared our pre-docking models against 7LID (minus its ligand) because it was solved with higher resolution than 7LIC (2.9 Å vs. 3.3 Å). TBM predicted the MhOR5 fold better than AF3 across a range of measures. Template modeling scores (see Methods) were 0.55 for MhOR5_TBM and 0.47 for MhOR5_AF3 (where a TM-score > 0.5 indicates correct fold prediction). The root mean square deviation (RMSD; measuring structural similarity) between 7LID and MhOR5_TBM (0.31Å) was also lower than between 7LID and MhOR5_AF3 (1.41Å) (Fig. 1a–b). In MhOR5_AF3 model, beginning of helix 1 was predicted differently than that of MhOR5_TBM model and the 7LID structure (Fig. 1e). Kinks in helices 2 and 7 of 7LID were accurately represented in both models (Fig. 1c, f, k) but there were prominent differences between 7LID and MhOR5_AF3, but not MhOR5_TBM, in extracellular regions of 7LID, specifically helices 3, 5, 6 and 7 (Fig. 1c, g–k). These differences were also reflected in structure-based sequence alignments at the end of helices 3, 5 and 7 and the beginning of helix 6 (Fig. 1l). Moreover, helices 4 and 5 were longer in MhOR5_AF3 than in 7LID (Fig. 1c, d, h–i), which might be because intracellular loop 2 (ICL2) connecting helices 4 and 5 was not resolved in the experimental MhOR5 structure. Fig. 1Superimposition of the experimental 7LID structure on the TBM and AF3 models. (a) Side and extracellular views of 7LID (salmon color) superimposed on the TBM model (blue). (b) Side and extracellular views of 7LID (salmon) superimposed on the AF3 model (green). Images were generated with Chimera41. (c) Helix-wise superimposition of 7LID (salmon), MhOR5_AF3 (green), and MhOR5_TBM (blue). (d) Structure-based sequence alignment of 7LID, MhOR5_AF3 and MhOR5_HM generated with Geneious version 2023.242 and highlighted according to hydrophobicity, from hydrophobic (red) to hydrophilic (blue). Superimposition of the experimental 7LID structure on the TBM and AF3 models. (a) Side and extracellular views of 7LID (salmon color) superimposed on the TBM model (blue). (b) Side and extracellular views of 7LID (salmon) superimposed on the AF3 model (green). Images were generated with Chimera41. (c) Helix-wise superimposition of 7LID (salmon), MhOR5_AF3 (green), and MhOR5_TBM (blue). (d) Structure-based sequence alignment of 7LID, MhOR5_AF3 and MhOR5_HM generated with Geneious version 2023.242 and highlighted according to hydrophobicity, from hydrophobic (red) to hydrophilic (blue). The next step was to compare the binding conformations and interactions of the two modeled structures with docked ligands, EOL and DEET with those in the corresponding 7LID and 7LIG experimental structures. For this, we first downloaded the structures of EOL and DEET (Fig. 2a–b) from the PubChem database and re-docked them to 7LID and 7LIG, respectively. This was done to validate the recovery of RLIs after docking and, as such, to then provide controls for the comparison of the ICM (see Methods) docking scores43 and contact residues obtained with the models. The re-docking recovered 100% of the experimental RLIs in both the 7LID_EOL and 7LIG_DEET complexes, with no additional contacts, and yielded control ICM docking scores of − 24.33 for 7LID with EOL and − 20.41 for 7LIG with DEET (Table S3). The docking scores for MhOR5_TBM with the two ligands, EOL and DEET, were higher than the control values, at − 19.20 and − 12.96, respectively, while the scores for MhOR5_AF3 were closer to the control, at − 21.7 and − 20.6 (Table S3). On the other hand, the binding conformations of both ligands within the MhOR5_TBM binding site closely resembled the experimental structures (Fig. 2d–e), whereas MhOR5_AF3 yielded slightly different binding conformations of both ligands compared to the experimental conformations (Figure S3), possibly due to the AF3 training database containing both Orco and MhOR5 PDB structures. The ligand binding pocket of MhOR5 involves helices 2, 3, 4, and 6 and, using our generic residue numbering system for each helix (see Methods), the contact sites for 7LID with EOL are 2.3, 2.6, 2.7, 3.26, 3.29, 3.30, 3.33, 4.13, 4.17, 6.7 and 6.10, and those for 7LIG with DEET are 2.3, 2.6, 2.7, 3.25, 3.26, 3.29, 3.30, 3.33, 4.17, 6.6, 6.7 and 6.10. These residues are mostly aromatic and hydrophobic (Figure S4), and all are non-polar. Functional assays of site-directed mutants have also implicated homologs of six of these contact sites, 2.3, 2.7, 3.29, 3.33, 4.13 and 4.17, in ligand binding in ORs from other insect species10,20–25,38,44. Some other sites close to these, namely 3.34, ECL3, and 6.3, have also been implicated in these studies (Table S4). All the contact residues in the experimental structures were also predicted when EOL and DEET were docked into MhOR5_TBM, while MhOR5_AF3 only partially captured the contact residues of the contact residues for EOL (72%), and of those for DEET (66%) (shown in Figure S4). Notably, the MhOR5_TBM model also predicted three other contact residues not in the experimental structural data (2.2, 4.16, and 6.3) while the corresponding number for MhOR5_AF3 was seven (2.2, 3.32, 3.36, ECL2, 5.32, 6.3, 6.4). Molecular dynamics and simulations for TBM and AF3 models of MhOR5 MD simulations were executed for all four combinations of the two models and two ligands. We also executed MD simulations for the experimental 7LID_EOL and 7LIG_DEET complexes to validate our MD simulation protocol (through recovery of the experimental RLIs) and to provide positive control values for post-dynamics energy calculations. As detailed below, we found the simulations generally improved the performance of both models, with TBM continuing to perform better than AF3. This assessment was based on comparisons of binding free energies, complex and ligand stabilities and RLIs. Post-dynamics energy calculations were performed using relative binding free energy methods, specifically molecular mechanics generalized Born surface area (MM/GBSA for GB estimates) and molecular mechanics Poisson–Boltzmann surface area (MMPBSA for PB estimates) methods, which are computationally efficient, although they assume that the conformations of the protein and ligand in their bound states are representative of their unbound forms. Absolute free energy methods take into account both the bound and unbound states separately but are computationally expensive, particularly when scaled to large datasets, and are currently difficult to apply to membrane proteins45,46. MM/GBSA and MM/PBSA were appropriate to use here because our primary goal was a comparative assessment of alternative models in predicting experimental binding affinities. Our post-dynamics binding energy values for the control 7LIG_DEET (GB: − 32.00 kcal/mol, PB: − 9.02 kcal/mol) and 7LID_EOL (GB: − 35.75 kcal/mol, PB: − 9.62 kcal/mol) complexes were notably low, as expected, and those for the simulations using the two TBM complexes (GB: − 26.60 kcal/mol, PB: − 5.49 kcal/mol for TBM_MhOR5_EOL and GB: − 34.13 kcal/mol, PB: − 12.59 kcal/mol for TBM_MhOR5_DEET ) were generally closer to the corresponding controls than those based on the AF3 models (GB: − 25.16 kcal/mol, PB: − 2.33 kcal/mol for AF3_MhOR5_EOL and GB: − 33.89 kcal/mol, PB: − 9.03 kcal/mol for AF3_MhOR5_DEET) (Fig. 2c). Post-dynamics stability for both the complex and ligand was assessed from the standard deviations (CSD and LSD, respectively) for the last 10 ns of the standard 50 ns simulations. Values for the control experimental structures (0.09 Å for 7LID_EOL and 0.10 Å for 7LIG_DEET) were again low, as expected. Both the TBM and AF3 complexes likewise generated low CSD values, with the TBM values lower than the AF3 values (0.07 Å and 0.18 Å respectively) for MhOR5_EOL but no difference between them (0.15 Å for both) for MhOR5_DEET. However, the LSD values for the modelled complexes were generally somewhat higher than their experimental counterparts, with the AF3 generating lower values than TBM in this case (0.33 Å and 0.22 Å for TBM and AF3, respectively for MhOR5_EOL and 0.28 Å and 0.14 Å for TBM and AF3, respectively for MhOR5_DEET). Importantly, however, the post-simulation positions and orientations of EOL and DEET in the TBM complexes largely overlapped with those in the experimental structures, whereas the corresponding comparisons for the AF3 complexes showed less agreement, particularly for DEET (Fig. 2c, d). These differences were also reflected in the post-simulation RLIs of the various complexes (Fig. 2f and S5); TBM and AF3 recovered 100% and 81%, respectively, of the experimental RLIs for EOL, and 91% and 83%, respectively, of them for DEET (Fig. 2d and S5). The only differences in the post-simulation RLI sets predicted by TBM from the experimental set were the absence of 3.25 and the presence of 2.10, 3.32, 4.16, and 6.3 in the respective predicted sets. Fig. 2Performance of TBM and AF3 in predicting the binding of EOL and DEET by MhOR5. (a) 2D structures of EOL (above) and (b) DEET (below). (c) Comparison of ICM docking scores and post-dynamics binding free energy calculations (kcal/mol) using EOL and DEET. Differences between experimental and modelled EOL conformations (above) in 7LID (pink), TBM_EOL (blue), and AF3_EOL (green) and (e) differences in DEET conformations (below) in 7LIG (pink), TBM_DEET (blue), and AF3_DEET (green). (f) RLIs of MhOR5_TBM with EOL and DEET were more similar than those of MhOR5_AF3 to 7LID and 7LIG. Elaborated post-docking and post-simulation (PS) RLIs are illustrated in Figures S4 and S5 respectively and docking RLIs are shown in Figure S6. Performance of TBM and AF3 in predicting the binding of EOL and DEET by MhOR5. (a) 2D structures of EOL (above) and (b) DEET (below). (c) Comparison of ICM docking scores and post-dynamics binding free energy calculations (kcal/mol) using EOL and DEET. Differences between experimental and modelled EOL conformations (above) in 7LID (pink), TBM_EOL (blue), and AF3_EOL (green) and (e) differences in DEET conformations (below) in 7LIG (pink), TBM_DEET (blue), and AF3_DEET (green). (f) RLIs of MhOR5_TBM with EOL and DEET were more similar than those of MhOR5_AF3 to 7LID and 7LIG. Elaborated post-docking and post-simulation (PS) RLIs are illustrated in Figures S4 and S5 respectively and docking RLIs are shown in Figure S6. Hydrophobicity-based MSAs of key helices for the TBM The ability of TBM to predict the known functionalities of our panel of five B. dorsalis and three B. minax ORs for which no experimental structural data were available was dependent on optimizing the alignment of the sequences of their key helices to those in the known structures of the template1. This is challenging for insect ORs for two reasons. Firstly, sequence similarities between any two ORs within a species can be as low as 17%12. Secondly, unlike mammalian ORs, insect ORs have very few highly conserved residues to use as anchor-points for the alignments; the only well-known motif in insect ORs is a single tryptophan at the end of predicted helix 612. Exploiting the crucial importance of hydrophobicity to the structure and function of the key helices, and drawing on multiple transcriptome datasets for B. minax17,46–48 and B. dorsalis20,34,49–52 (Table S5, Supporting information Sheet 1), we developed an extensive database of hydrophobicity-guided alignments of all seven helices in MhOR5, ApOR5, AaOR10, AgOR28 and the 130 ORs from the two Bactrocera species for which high confidence sequence data were available. As expected, an MSA of the 84 BdORs and 46 BmORs generated through ClustalW showed very little conservation (data not shown) but a hydrophobicity-based, helix-wise MSA including MhOR5 produced by iterative application of AlignME revealed several conserved motifs in most of the seven shared helices (Supporting information: Sheet 2). Conservation was greatest in helix 5 and lowest in helices 1 and 3, although some conservation of hydrophobic residues was evident even in the latter (e.g. sites 1.6, 1.8, 1.11, 1.12, and 1.15). Highly conserved residues or short motifs identified in individual helices of this MSA were then used as anchor points to guide further manual curation of the alignments of the helices containing them (Fig. 3; Table 1). Table 1Conserved sites/motifs identified in four helices.HelixAnchoring site(s)Conservation (%)2 K/R (2.13) 974Y/F (4.2), D (4.24), L(4.41)755LxxxI/VxxH (5.4), F/LxQF (5.31), H (5.11)80, 84, 856Y/F(6.7), W(6.41)80, 90Anchoring sites used to facilitate pairwise alignments between the Bactrocera ORs and MhOR5 are bold. Minimum levels of conservation for the respective sites/motifs in our MSA are also indicated. Table 1Conserved sites/motifs identified in four helices.HelixAnchoring site(s)Conservation (%)2 K/R (2.13) 974Y/F (4.2), D (4.24), L(4.41)755LxxxI/VxxH (5.4), F/LxQF (5.31), H (5.11)80, 84, 856Y/F(6.7), W(6.41)80, 90Anchoring sites used to facilitate pairwise alignments between the Bactrocera ORs and MhOR5 are bold. Minimum levels of conservation for the respective sites/motifs in our MSA are also indicated. Conserved sites/motifs identified in four helices. Anchoring sites used to facilitate pairwise alignments between the Bactrocera ORs and MhOR5 are bold. Minimum levels of conservation for the respective sites/motifs in our MSA are also indicated. Fig. 3Sequence logo plots53 for the seven helices of B. dorsalis and B. minax ORs based on their helix-wise MSAs. Colors represent the chemical nature of amino acid side chains. Y-axis scaling has been adjusted based on conservation. Highly conserved residues are highlighted in yellow. Sequence logo plots53 for the seven helices of B. dorsalis and B. minax ORs based on their helix-wise MSAs. Colors represent the chemical nature of amino acid side chains. Y-axis scaling has been adjusted based on conservation. Highly conserved residues are highlighted in yellow. Only one anchor point was identified for helix 2, site K/R (2.13), which occurred in 97% of the sequences in the dataset. However, three such sites were identified for helix 4, namely Y/F (4.2), D (4.24) and L (4.41), which all occurred in over 75% of the sequences. Helix 5 contained three conserved multi-site motifs: LxxxI/VxxH at sites 5.4–5.11 which was conserved in more than 80% of the sequences; F/LxQF at sites 31–34, with F/L (5.31) more highly conserved than QF (84 cf. 45%); and L/IC/G at sites 40–41, which was conserved in 65% of the sequences. Since H (5.11) was the most conserved site of all these residues (85%), it was used as the primary anchor point for adjusting the alignment of helix 5. Three conserved sites or motifs were also identified in helix 6; site Y/F (6.7), which was conserved in about 80% of the sequences; CYxG/A, with 78% C (6.19), 62% Y (6.20) and 95% G/A (6.22) conservation across the dataset; and Y/FxxxW at sites 6.37–6.41, which showed over 90% conservation across the dataset. Note that Y/F (6.7) is a contact site for EOL and DEET in the binding pockets of 7LID and 7LIG respectively. W (6.41) alone has previously been reported as a binding motif in D. melanogaster and Anopheles gambiae12 and was used here as the primary anchor point for this helix. All the above conserved sites except Y/F (4.2) and D (4.24) were also detected as motifs by MEME suite (Figure S7). We could not identify any anchoring sites for helices 1, 3, and 7 and therefore used the alignment generated through AlignME for these helices 1 and 7. However, available structural and mutational data indicates that helix 3 is important for ligand binding (Fig. 2f, Table S4). So, following the approach of Jabeen et al.54, we manually adjusted its alignment in order to optimise the hydrophobicity correspondence (HC) per site, using the sum of squared differences (SSD) calculator embedded in Bio-GATS55. The approach is detailed and illustrated for helix 3 for BdOR88a and BmOR7a1 in Supporting information: note 1 and Figures S8 and S9, respectively. The adjustments were then validated through simulations; for instance, complexes based on adjusted helix 3 models yielded lower SSDs per residue than AlignME for both BdOR88a_ME and BmOR7a1_linalool (BmOR7a1_LIN) for both GB (− 27.62 kcal/mol vs. − 14.9 kcal/mol and − 28.34 kcal/mol vs. − 22.65 kcal/mol, respectively) and PB (− 7.5 kcal/mol vs. − 4.8 kcal/mol and − 8.78 kcal/mol vs. − 3.93 kcal/mol, respectively). The optimized helix 3 alignments were integrated into the full length pairwise alignments, serving as the basis for use in subsequent model building. Testing the abilities of TBM and AF3 to capture empirical functional evidence We identified a panel of 63 odorants which have been tested for binding against one or more B. dorsalis or B. minax OR (Fig. 4, Table S6). Five of these complexes have EC50 values in the µM range which is widely regarded as potentially relevant biologically; biokinetic studies of ORs typically define high Affinity binding as within the nM to µM range11,56,57. These five complexes plus the structurally validated 7LID-EOL and 7LIG-DEET were selected as a binder cohort for further analysis. A complementary cohort of non-binders was then selected from the many non-binding OR-ligand combinations in the panel. Selection of these was based on maximizing the overlap with the binder cohort in respect of the ligands tested but was otherwise random. In total, the 14 combinations selected involved eight ORs and four odorants (1-octen-3-ol (Oct-3OL), LIN, ME, and EOL) (Figure S10). Two of the ORs, BmOR13a1 and BdOR13a1, were orthologs in the two species58. Fig. 4Experimental EC50 values taken from the literature for functionally characterized B. dorsalis and B. minax ORs. Experimental EC50 values taken from the literature for functionally characterized B. dorsalis and B. minax ORs. TBM and AF3 modeling were then carried out on the eight ORs in these combinations, using 7LID as a template for the TBM because it was bound to the agonist EOL and most of the other ligands in our binder cohort were also agonists of the receptors. The TBM also used final pairwise alignments of the bactroceran ORs in the cohorts tested with 7LID (Figures S11, S12). Sequence identity between those ORs and 7LID varied from 6.9 to 20% (Table S7). The docking results for the ORs in the binder cohort modelled using either TBM or AF3 with their respective ligands did not reflect the corresponding empirical data. The ICM docking scores (Table S8) were higher than those of the MhOR5_EOL control (− 24.33) and not significantly different from those in the non-binder cohort (p-values of 0.76 for the TBM and 0.23 for the AF3 models). Therefore, we proceeded with MD simulations, which, for each model, generated GB, PB, CSD and LSD values as defined above, and a total of 62 RLIs. Thirty and 27 RLIs were retained for TBM and AF3 respectively, after filtering out several with low variances and high correlations (r2 ≥ 0.7) with other RLIs (Figure S13). Linear discriminant analysis (LDA) of the energy and stability variables showed that TBM successfully distinguished between the binders and non-binders (Fig. 5a), whereas AF3 left some overlap between the two cohorts (Fig. 5b). To probe these results more closely we then used the distributions of normalized values predicted by each model for each energy and stability variable across all the complexes in the two cohorts (Fig. 5c). Considering first the two key energy variables, TBM generated clearly negative normalized values (<= -0.25) expected of true binders for both GB and PB for five of the seven binder complexes (MhOR5_DEET, BdOR13a1_Oct-3OL, BdOR7a6_Oct_3OL, BmOR7a1_LIN and BdOR88a_ME), whereas AF3 did so for four complexes (MhOR5_DEET, BmOR13a1_Oct-3OL, BdOR13a1_Oct-3OL, and BmOR7a1_LIN). Conversely, TBM generated the clearly positive normalized values ( > = 0.25) expected of true non-binders for both GB and PB for four of the seven non-binder complexes (BmOR7a1_Oct-3OL, BdOR82a_EOL, BdOR82a_Oct-3OL, and BdOR94b2_ME), whereas AF3 did not predict BmOR13a1_LIN, BmOR63a1_LIN, and BdOR82a_Oct-3OL as clear non-binders. Moreover, TBM only produced the counter-expectation, clearly positive values for both GB and PB for one binder (BmOR13a1_Oct-3OL) and clearly negative values for one non-binder complex (BmOR13a1_LIN), whereas AF3 produced clearly positive values for three binders (MhOR5_EOL, BdOR7a6_Oct-3OL, and BdOR88a_ME) and three clearly negative values for three non-binders (BmOR13a1_LIN, BmOR63a1_LIN, and BdOR82a_Oct-3OL). Thus, while not in agreement with the empirical data in all individual cases, the energy estimates from TBM outperformed those from AF3 in predicting both binders and non-binders, with AF3 having little predictive value in either case. The corresponding analysis for the two stability variables showed neither TBM nor AF3 consistently predicted either the empirical binder or non-binder cohorts. This was largely because there was much less correlation between the two stability variables than there was between the two energy variables generated by either TBM or AF3 (Figure S14). Consequently, there were relatively few cases where CSD and LSD were both <= − 0.25 or both > = 0.25. However, there was a general trend for TBM to produce the more negative CSD and LSD values indicating higher stability in the empirical binder than non-binder complexes. This was particularly the case for MhOR5_EOL, MhOR5_DEET, BdOR7a6_Oct-3OL, and BdOR88a_ME, respectively, for CSD and BmOR13a1_Oct-3OL, BdOR13a1_Oct-3OL, BmOR7a1_LIN, and BdOR88a_ME, respectively, for LSD. By contrast, the opposite trend was evident in the AF3 results (MhOR5_EOL, MhOR5_DEET, BdOR13a1_Oct-3OL, and BmOR7a1_LIN, respectively for CSD; BmOR13a1_Oct-3OL, BdOR13a1_Oct-3OL, and BdOR88a_ME, respectively for LSD). Most of the experimentally validated RLIs in the MhOR5, AaOR10, AgOR28, and ApOR5 complexes (2.3, 2.6, 2.7, 2.10, 3.26, 3.29, 3.30, 3.33, 4.13, 4.17, 4.20, 4.21, 6.7, and 6.10)13–15 were frequently found in either or both of the TBM and AF3 complexes (Fig. 5(d), S15 and S16) and most of the others predicted by the two models also lay in the four helices containing RLIs in the experimental structures. However, a few helix 1 and helix 5 RLIs were also predicted, the former mainly by AF3 and the latter mainly by TBM. Helix 5 has been implicated in the binding of geranyl acetate to ApOR5 in experimental structural data (PDBID: 8Z9A)15 and in in silico docking of the ligand VUAA1(N-(4-ethylphenyl)-2-{[4-ethyl-5-(pyridine-3-yl)-4 H-1,2,4-triazol-3-yl]sulfanyl}acetamide) with the Orco co-receptor in D. melanogaster59, while experimental structural data has shown helix 1 is involved in the binding of 2,4,5-trimethylthiazole with AgOR28 (PDBID: 8V3D)14. Empirical evidence from site-directed mutagenesis of DmOR85b has also implicated its site equivalent to 3.34 in its binding to 2-heptanone22. TBM and AF3 each generated several differences in the frequencies of specific RLIs between the binder and non-binder cohorts but there was little overlap between TBM and AF3 in these RLIs (Fig. 5d). The TBM and AF3 models also differed in the orientation of the ligands in the binding pockets (Figures S15, S16). At the helix-level, the RLIs in helix 5 implicated by TBM and those in helix 1 implicated by AF3 were much more frequent in the binder complexes. To confirm the involvement of helix 5 and helix 1 in the TBM and AF3 binder models respectively, we extended the respective simulation times to 200 ns for one of the strongest binding complexes, namely BdOR13a1 with Oct-3OL. Helix 5 in TBM_BdOR13a1_Oct-3OL and helix 1 in AF3_BdOR13a1_Oct-3OL remained consistently present in the simulation trajectories throughout the 200 ns. Nor did values for GB, PB, CSD, and LSD change significantly throughout the simulation (Table S9). Eleven individual RLIs were also discriminating, as each was markedly more prevalent in just one of the four TBM/AF3/binder/non-binder cohorts than it was in the other three. These were 4.20, 5.35, 5.36 and 5.39 in the TBM binder cohort, 2.2 in the TBM non-binder cohort, ECL2 in the AF3 binder cohort, and 3.33, 4.20, 4.21, and 6.14 in the AF3 non-binder cohort. Thus, inspection of the individual discriminating RLIs again showed prominent representation of helix 5 in the TBM binding dataset, although it did not do so for helix 1 in the AF3 non-binder dataset. Direct side-by-side comparisons of predicted parameter values for binder and non-binder complexes were possible for two of the ORs in our panels, BmOR7a1 and BmOR13a1, the latter being of particular interest because it had been the most refractory to TBM modelling (Fig. 5c, S15 and S16). The binding of BmOR7a1 with LIN and its non-binding with Oct-3OL were both successfully predicted by TBM (GB -0.46 vs. 1.07; PB − 0.69 vs. 0.73; CSD − 0.11 vs. 2.48; LSD − 0.74 vs. 2.51). Its RLIs with LIN lay in three of the four helices expected above (2–4) plus helix 5 but only three of the four expected (3, 4, 6) with Oct-3OL. In contrast, AF3 was able to capture differences in energy values and ligand stability (GB − 0.46 vs. 1.61; PB − 0.47 vs. 0.46; LSD − 1.4 vs. − 0.59) but did not fully reflect the difference in complex stability (CSD 0.89 vs. − 0.58), and its RLIs with LIN were in all four of the helices expected plus helix 1 and ECL2, and all four expected helices with Oct-3OL. On the other hand, the interactions of BmOR13a1 with its binder and non-binder ligands proved problematic to predict for both AF3 and TBM. AF3 yielded GB and PB estimates (− 1.1 and − 1.4, respectively) correctly predicting its binding to Oct-3OL, although its CSD and LSD estimate (2.62 and 0.12, respectively) indicated that both the complex and Oct-3OL were unstable in the binding pocket (Figures S17a, S18a). However, the estimates of GB, PB and LSD (− 0.83, − 1.22, − 0.59, respectively), albeit not CSD (1.41), which AF3 produced also incorrectly suggested it bound to LIN. In contrast, TBM yielded GB, PB and CSD estimates (0.69, 1.89, and 1.01, respectively) incorrectly predicting non-binding to Oct-3OL, although the LSD (− 0.64) suggested ligand stability in the binding pocket. TBM also failed to predict LIN as a non-binder of BmOR13a1, although its GB and PB values (− 0.24 and − 0.4, respectively) were less extreme than those produced by AF3 (− 1.43 and − 1.59, respectively) and its CSD estimate (0.45) suggested complex instability. Both models also failed to produce RLIs which distinguished the binder and non-binder interactions; AF3 predicted RLIs in helix 1 and ECL2 as well as the four helices expected with both ligands, while TBM predicted RLIs in three of the four expected helices (2, 4 and 6) plus 5 with both. Importantly, none of the results for any of the parameters changed materially when the simulations of BminOR13a1 with LIN and Oct-3OL were extended to 200 ns (Table S9). The problems with this OR emphasise the scope for further refinement of both TBM and AF3. Fig. 5Comparison of post-dynamics energy and stability values and RLIs produced by TBM and AF3 in the empirical binder and non-binder cohorts. (a) Linear discriminant analyses of the five energy and stability, and RLI variables showed TBM models differentiated binders from non-binders while (b) AF3 models showed some overlaps. (c) Heat maps of normalized values for each energy and stability variable for the TBM and AF3 post-dynamics complexes show that each variable contributed to the resolution between the binders and non-binders. Values <= − 0.25 (pink), > = 0.25 (orange), and intermediate (gray). (d) Heat maps of RLI frequencies show many differences between TBM and AF3 binders and non-binders. Discriminatory sites are boxed red. Detailed RMSD plots for the various complexes and ligands are shown in Figures S17–S20 and energy values are detailed in Table S10. Comparison of post-dynamics energy and stability values and RLIs produced by TBM and AF3 in the empirical binder and non-binder cohorts. (a) Linear discriminant analyses of the five energy and stability, and RLI variables showed TBM models differentiated binders from non-binders while (b) AF3 models showed some overlaps. (c) Heat maps of normalized values for each energy and stability variable for the TBM and AF3 post-dynamics complexes show that each variable contributed to the resolution between the binders and non-binders. Values <= − 0.25 (pink), > = 0.25 (orange), and intermediate (gray). (d) Heat maps of RLI frequencies show many differences between TBM and AF3 binders and non-binders. Discriminatory sites are boxed red. Detailed RMSD plots for the various complexes and ligands are shown in Figures S17–S20 and energy values are detailed in Table S10. Conclusions While both AF3 and TBM could predict the experimental structural folds correctly, TBM clearly outperformed AF3 in predicting key ligand binding features and only TBM could separate binding and non-binding interactions in the LDA. The success of the TBM was achieved, despite the long evolutionary separation of B. dorsalis and B. minax from the jumping bristletail (Machilis hrabei) template (~ 300My60), because it exploited a helix-wise hydrophobicity-based MSA involving 130 high quality OR sequences from the two bactrocerans. This MSA revealed previously undetected anchor points and other sequence features that could be used to fine-tune the alignment of the individual ORs to the template. Although not used in the LDA, the RLI analysis also implicated some individual residues and motifs that provided at least quantitative discrimination between binding and non-binding OR-ligand interactions. The RLIs are therefore also potentially valuable indicators of productive binding interactions and as such warrant empirical validation. The TBM approach would still benefit from more comprehensive MSAs, because it relies heavily on accurately annotated OR sequences. This could be key to its extension to other insect groups such as mosquitoes, caterpillars and beetles where there is also great interest in pest detection and control using pheromones and kairomones61–63. Fortunately, the rapid uptake of the latest sequencing technologies is now generating high quality genome assemblies for diverse insects64, so this extension is well within reach. Moreover, the success of TBM in resolving binder and non-binder complexes herein suggests it could already provide a useful in silico pre-screen for productive OR-ligand interactions and thence potential lures for many major fruit fly pests around the world. While relative free energy estimates were sufficient for the comparative analyses of TBM and AF3 models in predicting experimentally established interactions, further developments of the tools for estimating absolute free energy values will also be important for future de novo predictions of potential OR-ligand interactions. Methods The overall methodology is summarized in Fig. 6. Fig. 6In silico workflow to compare the ability of TBM and AF3 to reproduce the experimental structural and functional data using molecular docking and simulation. Post-dynamics binding affinities were assessed in terms of estimates of GB (generalized Born) and PB (Poisson–Boltzmann) energy values, CSD and LSD stability values (expressed as root means square deviations) for the OR-ligand complex and ligand, respectively, in the last 10ns of MD and RLIs (post-simulation receptor ligand interactions). In silico workflow to compare the ability of TBM and AF3 to reproduce the experimental structural and functional data using molecular docking and simulation. Post-dynamics binding affinities were assessed in terms of estimates of GB (generalized Born) and PB (Poisson–Boltzmann) energy values, CSD and LSD stability values (expressed as root means square deviations) for the OR-ligand complex and ligand, respectively, in the last 10ns of MD and RLIs (post-simulation receptor ligand interactions). Dataset collection The datasets used in the current study comprise available insect OR structural data (Table S1)13–15, functionally characterized B. dorsalis20,34,36 and B. minax ORs17,37, and transcript sequences of all available B. dorsalis20,34,49–52 and B. minax ORs17,46–48. The eight functionally characterized ORs for which we collected publicly available sequences and odorant data were BdOR7a620, BdOR13a120,34 BdOR82a34, BdOR88a35, and BdOR94b236 in B. dorsalis, and BmOR7a117, BmOR13a1 and BmOR63a137 in B. minax. All these ORs have been heterologously expressed using the Xenopus oocytes expression system and deorphanized using two-electrode voltage-clamp electrophysiological recordings with known EC50 values. The five odorants docked and simulated in the current study were DEET, Oct-3OL, EOL, LIN and ME. Their structures were downloaded from the PubChem database65 and are displayed in Figure S10. They were classified as binders or non-binders with each of the ORs in question on the basis of the responses they elicited in electroantennography (EAG) recordings (Table S6). Only those odorants with well-defined EC50 values were considered as binders. The OR sequence dataset for the MSA comprises 84 and 46 OR transcript sequences reported for B. dorsalis and B. minax, respectively, in published transcriptomic studies17,20,34,46–52. Sequence alignment The MSA was primarily developed from progressive pairwise alignments of the B. dorsalis and B. minax OR sequences (Supporting information sheet 1) with MhOR5, ApOR5, AaOR10, and AgOR28 using AlignME66. AlignME bases its alignment on a hydrophobicity matrix that is suitable for membrane proteins such as ORs. We used it to align each of the seven helical regions forming the transmembrane structures of the ORs. The conserved residues in each helix were identified using logo plots generated in the R programming language with the ggseqlogo package53 and using MEME Suite 5.5.467. The maximum number of motifs were set at eight and their width was limited to 2–20 residues for MEME Suite. The pairwise alignments were then manually edited based on predicted conserved residues using MEGA 1168. Structural model predictions TBM models were built for the five B. dorsalis and four B. minax ORs using Modeller 10.369 with the electron microscopy structure of MhOR5 in complex with the agonist EOL (PDBID: 7LID)13 as a template and the pairwise sequence alignments derived in the MSA. Fifty models were generated through Modeller 10.3 and the model with lowest objective function was selected for the next step, i.e., molecular docking. AF3 models for the same ORs were generated using AlphaFold3 webserver4 with eight different seeds. Five models were generated for each seed and the relaxed model with the highest confidence score was selected for molecular docking. The TM-score program70 was used to calculate template modeling (TM) scores. Molecular docking The binding pockets of the AF3 and TBM models for each OR were predicted through ICMPocketFinder embedded in ICM-pro software43 and refined through available mutagenesis data (Table S4). The initial conformation and orientation of the ligands inside the binding pockets of both AF3 and TBM models were generated through ICM-pro. The ligand charges were automatically assigned by ICM-pro. The docking effort in the ICM-pro software was set to three and ten conformations were generated for each complex. The complexes with the lowest ICM score and the ligand inside the binding pocket were retained for further analysis. To simplify the interpretation of docking results a residue numbering scheme was introduced following the Ballesteros and Weinstein numbering scheme for GPCRs71. We assigned a two-number (x.y) identifier to each residue in the transmembrane helices of our models, where the first number indicates the order of the helix and the second the order of the residue in the helix in MSA. Membrane building The docked complexes were embedded inside a membrane in the full simulation system. CharmmGUI72 was used to build a lipid bilayer for each model containing 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphoethanolamine (POPE) and 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) molecules based on published lipodomics data for Drosophila melanogaster73. Each system was built in a 0.15 M NaCl solution, using the Monte Carlo method to place the Na + and Cl- ions. A CHARMM36 forcefield was assigned and the whole system was equilibrated at constant pressure and temperature (NPT ensemble; 298.15 K). The final system size for each complex is listed in Table S11. Molecular dynamics simulation The MD simulations were performed using the Amber22 package74. Each simulation was initiated with three steps of energy minimization across 7,000 iterations, with the algorithm switched from steepest descent to conjugate gradient After 1000 iterations in each step and the restraints of a force constant of 10 kcal mol− 1Å2 applied to the complex in the first two steps. After the third minimization step, the temperature of the system was gradually increased to 298.15 K in three heating steps over 205 ps, with the positional restraints applied to both the complex and the lipid part of the membrane in all three steps. The whole system was then equilibrated in an NVT ensemble for 6ns. The production simulation was performed at 298.15 K for 50 ns, with trajectories being stored every ns and a lipid21 forcefield applied to the lipid bilayer, ff14SB to the receptor, gaff to the ligand and the TIP3P model to the water molecules. The Austin Model 1-bond charge correction (AM1-BCC) method was employed for calculating the partial atomic charges of each ligand. The ANTECHAMBER module was used to create ligand topologies. Minimization and heating steps were carried out using PMEMD.mpi module followed by the GPU-accelerated Particle-Mesh Ewald Molecular Dynamics (PMEMD) cuda module to execute the equilibration and production steps. Binding-free energy calculation The binding free energies for selected receptor-ligand complexes were estimated through both MMPBSA and MMPBSA methods. As noted, tools for estimating absolute free energies are resource intensive75 and also, these workflows have yet not been optimized for membrane proteins. Water molecules, counter ions and lipid molecules were stripped out from all receptor-ligand complexes prior to calculations. The last 2 ns were selected for binding affinity predictions in interest of time and computational resources. The MMPB(GB)SA was calculated as in Eq. 1 using MMPBSA.py script embedded in Amber2274.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Delta {G_{MMPB\left( {GB} \right)SA}}={\left( {{G_c}} \right)_i} - ~{\left( {{G_r}} \right)_i} - ~{\left( {{G_l}} \right)_i}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left({G}_{c}\right)}_{i}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left({G}_{r}\right)}_{i}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left({G}_{l}\right)}_{i}$$\end{document} are average values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varDelta\:G}_{MMPB\left(GB\right)SA}$$\end{document} for the complex, receptor, and ligand for i snap shots of trajectories. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3