Understanding how molecular structures are involved in recognition by a biological receptor is a decisive step in drug design, and could constitute an intricate problem because of existence of several binding sites, as is the case of GPCRs (1) that constitute the largest class of membrane receptors. In this context, identification of pharmacophores that should differentiate multiple binding modes is of particular interest. We have recently applied to ligands of a human olfactory receptor an original sorting-out procedure carried out using Catalyst/HypoGen software (Accelrys Ltd) (2). We aimed to validate this sorting out procedure using literature data, and in this way, we focused on CB1 agonists and antagonist. Indeed, CB1 ligands present several qualities: they have been extensively studied since several years, and it is now admitted that sites of cannabinoid agonists are distinguishable from binding sites of cannabinoid antagonists (3). Furthermore, CB1 receptor possess two distinct subsites for ligand binding (4,5). We used two training sets: the first one (P-23) constituted by 23 classical cannabinoids CB1 agonists (6), the second one (C-29) by 29 CB1 antagonists (7). Hypothesis generation was carried out separately on the two sets P-23 and C-29, and on the whole set (PC-53), testing several features associations of five features: Hydrogen Bond Acceptor (HBA), Hydrogen Bond Donor (HBD), Hydrophobic (HY), Hydrophobic Aliphatic (HYAL), Ring Aromatic (RA). The best significant hypothesis obtained for group P-23 was constituted by 1 HY, 3 HYAL and 1 RA (cost=160, correl=0.89, Config=15.5, fixed cost=66.7, null cost=500), whereas 1 HY, 2 HYAL and 1 HBD constituted the best significant hypothesis obtained for group C-29 (cost=186, correl=0.89, Config=15.9, fixed cost=80, null cost=554). Ten hypotheses were obtained starting from the merged group PC-53; two HBA and two HY constituted the best significant hypothesis (cost=569, correl=0.84, Config=14, fixed cost=128, null cost=1585). On the basis of compound's alignment, we identified two subsets that allowed to obtain significant models. The first one is constituted by 6 agonists (group P-6), and the second one by 5 antagonists (C-5). Activities estimation and addition of well estimated compounds respectively in the subsets P-6 and C-5 led to obtain two new groups: P-13 (best significant hypothesis: 2 HY, 2 HYAL, 1 RA, cost=47, correl=0.99, Config=15, fixed cost=44.5, null cost=237) and C-23 (best significant hypothesis: 2 HY, 2 HYAL, 1 HBD, cost=75, correl=0.98, Config=15, fixed cost=67, null cost=243). We have thus successfully separate agonists from antagonist; moreover our procedure allowed to improve the quality of both agonists and antagonists models, and to identify several outliers in agonist group that should be related to the existence of two distinct subsites for CB1 agonist binding (4,5).