Protein degradation targeted chimera connector generation method based on deep reinforcement learning

Abstract

The invention discloses a protein degradation targeted chimera linker generation method based on deep reinforcement learning, which comprises the following steps: expanding a constructed ZINC data set and a PROTAC data set by using a data enhancement method, and taking the obtained first ZINC data set and first PROTAC data set as a supervision training set; constructing a Transform model, and setting a training step number, a network layer number, an attention layer number and optimizer parameters; training a Transform model by using the first ZINC data set, and training and updating the Transform model by using an objective function and a back propagation algorithm; the first PROTAC data set is used for training the updated Transform model, the Transform model is migrated to a PROTAC target domain, and a Prior prior model is obtained; inputting the segment pair SMILES into a Prior prior model for batch generation, scoring the generated PROTAC by using a scoring function, introducing a strategy gradient algorithm of reinforcement learning, and updating an Agent model; repeating until the PROTAC score is no longer increased or the training step number is reached; and the updated Agent model is utilized to realize large-scale generation of the protein degradation targeted chimera linker conforming to expected attributes under the condition of given fragment pairs.

Description

technical field [0001] The present invention relates to the technical field of artificial intelligence and protein degradation targeting chimera design, and more specifically, to a method for generating protein degradation targeting chimera linkers based on deep reinforcement learning. Background technique [0002] Protein degradation targeting chimeras (PROTACs), bifunctional small molecules capable of inducing degradation of proteins of interest (POIs), have shown intense interest in drug discovery in recent years. PROTAC mainly consists of three parts: a ligand (warhead) targeting POI, a ligand that recruits E3 ubiquitin ligase, and a linker connecting the two. Due to this bifunctional structural property, PROTAC can simultaneously bind POI and E3 ligase to form a ternary compound, and make POI be labeled with ubiquitin, and then sent to the ubiquitin protein system to be degraded by the 26S proteosome. This process is different from the occupancy-driven approach of trad...

AI Technical Summary

Problems solved by technology

[0006] Although these two deep generative models have achieved good results in generating connectors after training on ChEMBL and ZINC datasets, both methods are difficult to generalize to the generation of PROTAC connectors

The SMILES generated by SyntaLinker on the PROTAC dataset has very low legitimacy and cannot get the correct linker at all

However, DeLinker is trained in a graph-based manner, and its legality is guaranteed, but the structure of the generated linker is quite different from the real linker, and it cannot be applied to the real PROTAC design.

At the same time, the two acquire the molecular distribution of the training set, so the properties of the generated linkers can only be close to the training set, and molecules with any desired properties cannot be generated, so as to solve some problems such as the oral absorption and availability of PROTAC.

Images