Transformer-Based Generative Model Accelerating the Development of Novel BRAF Inhibitors

Abstract

The de novo drug design based on SMILES format is a typical sequence-processing problem. Previous methods based on recurrent neural network (RNN) exhibit limitation in capturing long-range dependency, resulting in a high invalid percentage in generated molecules. Recent studies have shown the potential of Transformer architecture to increase the capacity of handling sequence data. In this work, the encoder module in the Transformer is used to build a generative model. First, we train a Transformer-encoder-based generative model to learn the grammatical rules of known drug molecules and a predictive model to predict the activity of the molecules. Subsequently, transfer learning and reinforcement learning were used to fine-tune and optimize the generative model, respectively, to design new molecules with desirable activity. Compared with previous RNN-based methods, our method has improved the percentage of generating chemically valid molecules (from 95.6 to 98.2%), the structural diversity of the generated molecules, and the feasibility of molecular synthesis. The pipeline is validated by designing inhibitors against the human BRAF protein. Molecular docking and binding mode analysis showed that our method can generate small molecules with higher activity than those carrying ligands in the crystal structure and have similar interaction sites with these ligands, which can provide new ideas and suggestions for pharmaceutical chemists.

Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation

“…9 a–c, and e the Tr model is much less stable under RL optimization compared to the RNN and more readily undergoes mode collapse i.e., it starts generating invalid or repeated (non-unique) molecules, as shown in Additional file 1 : Figures S18-19. In fact, very few implementations of transformers exist within a RL setting (e.g., [ 104 , 105 ]) likely due to this instability during training and computational expense, sometimes being distilled to an RNN for RL [ 10 ]. Therefore we additionally implemented a modified transformer architecture designed to stabilize model optimization during RL [ 69 ].…”

Transformer-based deep learning method for optimizing ADMET properties of lead compounds

“…Algorithms such as gated recurrent unit (GRU), long short-term memory (LSTM), and Transformer are widely used to extract sequence features and directly used to build predictive models, among which models with Transformer as the basic building block achieve state-of-the-art performance in most NLP-based classification and regression tasks. 10,19 Examples include the Bert model based on the Transformer encoder and the GPT model based on the Transformer decoder. 20,21 The Bert model has been well used in NLP tasks, such as a question answering system, text classification and entity extraction.…”

“…9 Among them, reinforcement learning (RL) is used as an optimization algorithm and docking software is used as a scoring function to find a small molecule with the highest affinity for binding to the target in the chemical space. 10,11 However, the coordinated optimization of PD and PK properties in molecule generation tasks remains a great challenge for most existing models, 12 especially when they focus on the pursuit of affinity for the lead compound. Therefore, the subsequent lead-optimization work is very necessary.…”

Recurrent neural network (RNN) model accelerates the development of antibacterial metronidazole derivatives

“…We found that the 1-layer GRU performs comparable to the 4-layer GRU and can be successfully used for small molecule library design. 22 Therefore, we use the above model to accelerate the development of antibacterial drugs, and the trained stack-augmented GRU is saved as a pretrained model to be weighted by a task-specific transfer learning model.…”