**MTMol-GPT: De novo multi-target molecular generation with transformer-based generative adversarial imitation learning**
**Authors:** Chengwei Ai, Hongpeng Yang, Xiaoyi Liu, Ruihan Dong, Yijie Ding, Fei Guo
**Abstract:**
De novo drug design is crucial for advancing drug discovery, aiming to generate new drugs with specific pharmacological properties. Recent deep generative models have achieved significant progress in generating drug-like compounds, but they primarily focus on single-target drug generation, neglecting the complex mechanisms of diseases and multiple influencing factors. To address this, we propose MTMol-GPT, an upgraded GPT model with generative adversarial imitation learning for multi-target molecular generation. The model employs a dual discriminator using the Inverse Reinforcement Learning (IRL) method to generate multi-target molecules concurrently. Extensive results show that MTMol-GPT generates various valid, novel, and effective multi-target molecules for complex diseases, demonstrating robustness and generalization. Molecular docking and pharmacophore mapping experiments validate the drug-likeness and effectiveness of the generated molecules, particularly in improving neuropsychiatric interventions. The model's generalizability is further demonstrated through a case study on multi-targeted drug design for breast cancer. MTMol-GPT provides a valuable tool for de novo drug design, enhancing the efficiency and quality of multi-target drug discovery.
**Introduction:**
De novo drug design aims to generate novel compounds with specific pharmacological and chemical properties. The scale of possible structures for drug-like compounds is vast, and traditional methods often rely on high-throughput technology and virtual screening. While single-target drug design has shown promising results, it often requires higher doses and can cause mutations. Multi-target molecules, which target multiple proteins simultaneously, are highly sought after for treating complex diseases. However, generating such molecules remains a challenging task. Deep generative models have facilitated the generation of novel molecules within a vast chemical space, but they often focus on single-target drug generation. GPT variants, such as MTMol-GPT, use generative adversarial imitation learning to optimize the generation of multi-target molecules. The model pre-trains transformer neural networks on the ChEMBL database to ensure the validity of generated structures and employs a dual-target molecular sequence generation system through the GAIL algorithm. Extensive experiments on DRD2 and HTR1A receptors demonstrate the model's ability to generate novel chemical matter targeting multiple proteins. The model's performance is superior to other state-of-the-art methods, and its generalizability is validated through a case study on breast cancer.
**Results:**
MTMol-GPT extends GAIL for multi-target molecular generation and is built on a GPT model. It involves five key steps: pre-training, generating compounds, collecting compounds in a replay buffer, updating the generator and discriminator, and evaluating the model. The model's performance is evaluated using MOSES metrics, which show high validity and structural similarity**MTMol-GPT: De novo multi-target molecular generation with transformer-based generative adversarial imitation learning**
**Authors:** Chengwei Ai, Hongpeng Yang, Xiaoyi Liu, Ruihan Dong, Yijie Ding, Fei Guo
**Abstract:**
De novo drug design is crucial for advancing drug discovery, aiming to generate new drugs with specific pharmacological properties. Recent deep generative models have achieved significant progress in generating drug-like compounds, but they primarily focus on single-target drug generation, neglecting the complex mechanisms of diseases and multiple influencing factors. To address this, we propose MTMol-GPT, an upgraded GPT model with generative adversarial imitation learning for multi-target molecular generation. The model employs a dual discriminator using the Inverse Reinforcement Learning (IRL) method to generate multi-target molecules concurrently. Extensive results show that MTMol-GPT generates various valid, novel, and effective multi-target molecules for complex diseases, demonstrating robustness and generalization. Molecular docking and pharmacophore mapping experiments validate the drug-likeness and effectiveness of the generated molecules, particularly in improving neuropsychiatric interventions. The model's generalizability is further demonstrated through a case study on multi-targeted drug design for breast cancer. MTMol-GPT provides a valuable tool for de novo drug design, enhancing the efficiency and quality of multi-target drug discovery.
**Introduction:**
De novo drug design aims to generate novel compounds with specific pharmacological and chemical properties. The scale of possible structures for drug-like compounds is vast, and traditional methods often rely on high-throughput technology and virtual screening. While single-target drug design has shown promising results, it often requires higher doses and can cause mutations. Multi-target molecules, which target multiple proteins simultaneously, are highly sought after for treating complex diseases. However, generating such molecules remains a challenging task. Deep generative models have facilitated the generation of novel molecules within a vast chemical space, but they often focus on single-target drug generation. GPT variants, such as MTMol-GPT, use generative adversarial imitation learning to optimize the generation of multi-target molecules. The model pre-trains transformer neural networks on the ChEMBL database to ensure the validity of generated structures and employs a dual-target molecular sequence generation system through the GAIL algorithm. Extensive experiments on DRD2 and HTR1A receptors demonstrate the model's ability to generate novel chemical matter targeting multiple proteins. The model's performance is superior to other state-of-the-art methods, and its generalizability is validated through a case study on breast cancer.
**Results:**
MTMol-GPT extends GAIL for multi-target molecular generation and is built on a GPT model. It involves five key steps: pre-training, generating compounds, collecting compounds in a replay buffer, updating the generator and discriminator, and evaluating the model. The model's performance is evaluated using MOSES metrics, which show high validity and structural similarity