In order to better formalize the notorious Inverse-QSAR problem (finding structures of given QSAR-predicted properties) is considered in this paper as a two-step process including (i) finding “seed” descriptor vectors corresponding to user-constrained QSAR model output values and (ii) identifying the chemical structures best matching the “seed” vectors. The main development effort here was focused on the latter stage, proposing a new Attention-based Conditional Variational AutoEncoder (ACoVAE) neural-network architecture based on recent developments in attention-based methods. The obtained results show that this workflow was capable of generating compounds predicted to display desired activity, while being completely novel compared to the training database (ChEMBL). Moreover, the generated compounds show acceptable druglikeness and synthetic accessibility. Both pharmacophore and docking studies were carried out as “orthogonal” in silico validation methods, proving that some of de novo structures are, beyond being predicted active by 2D-QSAR models, clearly able to match binding 3D pharmacophores and bind the protein pocket.

The data is the following : - model.yaml : the model parameters in yaml format - chembl23_umap1_std.smi : the standardized SMILES structures taken from ChEMBL23. Standardization rules are described in the publication. - chembl23_umap1.svm : the fragment descriptors used for the model training. The fragmentation is described in the publication. The first column contains the SMILES of the molecule. Other columns consist in a pair of values separated by a ":". The first value identifies the fragment’s index in the header file, the second one is the fragment count. - IA-FF-FC-AP-2-3.hdr : header file .hdr, containing the index and a string representing each fragment discovered into the SDF.

ISIDA/Fragmentor, 2022