Membrane proteins play crucial roles in organismal transduction of information. Molecular recognition in biological membrane systems is the initial process triggering external factors to be passed into the inner part of a cell. G protein-coupled receptors (GPCRs), which consist of 7 transmembrane segments, play pivotal roles for signal transduction in higher eukaryotes and organize the largest families of proteins in the human genome [1]. The fact that drugs targeting GPCRs constitute over one-third of pharmaceuticals shows the importance of GPCRs as drug targets [2].

Adenosine receptors (AdoRs) belong to a family of rhodopsin-like class A GPCRs, which constitute the largest portion of GPCRs in humans, including olfactory receptors, and there are four known subtypes of AdoRs-namely, AdoR1, AdoRA2A, AdoRA2B, and AdoR3 [3, 4]. Although these AdoR subtypes have distinct amino acid sequences and tissue-specific distribution, their endogenous agonist and ligand-binding pockets are highly conserved.

The biogenic nucleoside adenosine (Fig. 1) exerts its effects by interacting with AdoRs that are involved in various diseases, such as cardiac ischemia, arrhythmia, neurodegeneration, diabetes, glaucoma, and inflammation [5]. The physiological effects of AdoRs can be used for diagnoses and planning of the surgical strategies, such as in coronary artery diseases [6].

Olfaction, the sense of smell, is also mediated by class A GPCRs-namely, olfactory receptors. The fundamental mechanism of olfaction is now under some controversy [7-9]. Various attempts were introduced to describe the molecular mechanism of receptor-ligand recognition in olfaction, such as the binding theory and vibration theory [10-12]. The former, where the specificity of ligand is explained by molecular shape, has been developed into a concept of a pharmacophore and is generally accepted by researchers. Nevertheless, this is not sufficient to give the account of ligand variety and agonism complexity of GPCRs.

In recent decades, there have been many efforts to find an efficient way to seek or make appropriate ligands working on the target receptors through a chemogenomic approach. Application of molecular descriptors can make an efficient way to discriminate ligand binding modes in receptor-ligand relationships in such cases as in silico chemogenomic screening [13]. Molecular descriptors encoding information about the molecular structure can be classified by the dimensionality of their molecular representation, and they are almost on the structural, topological, and geometrical bases [14]. They are also including information on dipole moment, electric polarizability, and electrostatic potential in their data types.

Lately, a computational approach was carried out to search for a molecular vibration-activity relationship in the agonism of histamine receptors, and the author suggested that the molecular vibrational frequency pattern may play a role as a possible molecular descriptor for histamine receptor ligands [15]. The EigenVAlue (EVA) descriptor is also based on infrared range molecular vibrations among the various molecular descriptors and is a unique one, based on eigenvalues corresponding to individual calculated normal modes among various molecular descriptors [16, 17].

In the present paper, we tested the potential of the corralled intensity of molecular vibrational frequency (CIMVF) [15] as a molecular descriptor, compared to the EVA descriptor and other descriptors for the classification of AdoR ligands.

As shown in Fig. 1, adenosine and caffeine, a typical agonist and antagonist of AdoRs, respectively, share the structure of purine. However, the Tanimoto coefficient between the two ligands is not high. The Tanimoto coefficient of some agonists to adenosine was lower than 0.3, though they did not take a large share. Moreover, the lowest Tanimoto coefficient of the agonists was 0.12766 (capadenoson) and is a lower value as an antagonist.

The results of hierarchical clustering of the similarity matrices from the MACCS keys, PubChem, and Klekota-Roth fingerprints; EVA; and CIMVF are shown in Fig. 2.

The dendrograms obtained from MACCS keys and PubChem fingerprint were shown to be similar to the Tanimoto coefficient pattern. The result of the Klekota-Roth fingerprint analysis showed a somewhat different pattern to the fingerprint-based one, as only two antagonists were located in the cluster of agonists. However, the three fingerprint-based results did not match the known facts of AdoR agonism.

In the case of EVA, Takane and Mitchell [27] reported a structure-odor relationship using the EVA descriptor and hierarchical clustering. The authors reported that the dendrograms that were produced by the EVA method outperformed those from UNITY 2D in a classification of odorant molecules. However, the EVA analysis in the current study was positive but not perfect for the molecular property-activity relationship in AdoR agonism. The larger σ got, the more the clusters in the dendrogram got entangled (data not shown). It might come from the Gaussian function that smeared out the vibrational frequencies, such that vibrations of similar frequency overlapped together. The dendrogram obtained from the EVA descriptor showed an eligible tree for the agonism of AdoRs on the whole (Fig. 2D). One exception in which an antagonist (ZM241385) was clustered in a subtree of agonists might have resulted from smoothing by the Gaussian convolution of molecular vibration data.

All antagonists were clustered into a subtree in the dendrogram obtained from CIMVF analysis, as shown in Fig. 2E. We can tell the regional difference between agonists and antagonists in the dendrogram and also find that the information from the CIMVF analysis can play a role in the discrimination of agonists from antagonists in AdoR agonism. The discreteness of CIMVF data seems to facilitate the performance of the binary classification of AdoR ligands.

To find the topological difference between dendrograms, similarity scores between every two dendrograms were shown in Table 2. The similarity between two dendrograms was represented as a percentage. For example, the score of topological similarity between the two dendrograms of CIMVF and EVA analyses is 59.6%. The value has the highest similarity among all comparisons. A distance meta-tree build from five dendrograms of 46 AdoR ligands was illustrated in Fig. 3. As shown in the tree, the resulting dendrogram produced from the CIMVF data analysis has the nearest position to that produced from the EVA data analysis and is located the farthest from the dendrogram produced from the analysis of MACCS keys.

It is generally known that structurally similar molecules have similar properties or functions; however, a small change in molecular structure can fairly affect its vibrational frequencies. The structural variety of AdoR ligands was shown in Fig. 4. The Tanimoto coefficients of LUF5835, BAY60-6583, and capadenoson to adenosine were smaller than 0.18, and the ligands share a hydroxyphenyl pyridine dicarbonitrile backbone (Fig. 4A). They seem to be dissimilar to other agonists that share the structure of purine. Several antagonists (Fig. 4B) do not have a similar structure to caffeine; however, they act as antagonists of AdoRs.

As mentioned above, the current experiment does provide circumstantial evidence for the molecular vibrational information to AdoR agonism, at least. With a more concentrated study on the relationship between the molecular vibrational frequency and pharmacological function of a ligand, the vibrational spectrum of a ligand molecule may explicitly propose a novel path to the field of receptor-ligand interaction mechanisms.