Caspase is a family of cysteinyl aspartate-specific proteases, which plays a critical role in the cell regulatory networks controlling inflammation and programmed cell death.[1] Up to now, 11 functional caspase subtypes (i.e. caspase 1–10, 14) have been found in human encode proteins, of which caspase-1, -4 and − 5 are related to inflammatory response, caspase-14 to keratinocyte differentiation and others to apoptosis. The apoptotic caspases are further divided into two subcategories, namely apoptotic initiator and executioner caspases according to their functions in apoptosis processes. The initiator caspases (caspases-2, -8, -9, and − 10) can be recruited and activated by either death receptors or apoptosomes, while the downstream executioner caspases (caspases-3, -6, and − 7) are responsible for the actual cell destruction.[2–4]
Accumulated evidences have suggested that the activation of caspase-6 is responsible for neuronal apoptosis and amyloid β peptide (Aβ) deposition, which is highly involved in age-dependent axon degeneration and neurodegenerative diseases, such as Huntington's disease and Alzheimer's disease.[5–7] Due to the potencies in the treatments of neurodegenerative diseases, caspase-6 inhibitors have attracted intensive attentions. Recently, a series of aza-peptides,[8] acyl dipeptides,[9, 10] and non-peptide benzenesulfonyl chloride, isatin sulfonamide,[11–15] tetrafluorophenoxy methyl ketone,[16] phenothiazin-5-ium derivatives,[17] heteroaryl propanamido hexanoic acid,[18] vinyl sulfone,[19] furoyl-phenylalanine derivatives[20] have been identified as caspase-6 inhibitors with nanomolar to micromolar potencies (Fig. 1). In spite of promising efficacies, the available caspase-6 inhibitors also showed inevitable pharmacokinetic deficiencies (e.g. cell toxicity, low permeability, metabolic instability) that restrict their clinical applications.[21]
Over the last decade, deep learning (DL) technologies, such as convolutional networks (CNN), restricted Boltzmann machine (RBM), recurrent neural networks (RNN), and generative adversarial network (GAN) have been gradually applied in drug design and proven to be promising approaches for artificial intelligence-based drug design.[22], [23, 24] Recently, RNN-based molecular generative network has attracted particular attentions duo to its unique features in de novo molecular design.
In this paper, gated recurrent unit (GRU)-based RNN network combined with transfer learning and traditional machine learning were employed for de novo molecular design of caspase-6 inhibitors. The results showed that the established generative RNN model can generate efficiently potent caspase-6 inhibitors with the similar chemical space distribution to the known caspase-6 inhibitors, which can be easily incorporated with the traditional molecular design methods. Collectively, this paper provides an efficient combinational strategy for de novo molecular design of caspase-6 inhibitors.