SELFIES-based adversarial autoencoder framework for de novo drug-like molecule generation

Deep learning techniques have become increasingly prominent in drug discovery, particularly for generating new molecular structures. This study presents a novel approach that combines an autoencoder with an adversarial autoencoder (AAE) to facilitate de novo molecular design. The proposed framework is tested in two contexts: producing random drug like molecules and generating molecules tailored to specific targets. The AAE-based model performs well in both scenarios, according to experimental results. The produced molecules contain a significant number of distinct structures and cover a chemical space similar to that of the training data. Furthermore, the drug likeness scores of these molecules are comparable to those in the training set. The AAE-generated molecules show dramatic differences compared to molecules generated with a recurrent neural network (RNN)-based model, showcasing the complementary strengths of both approaches.