Abstract:DrugGPT presents a ligand design strategy based on the autoregressive model, GPT, focusing on chemical space exploration and the discovery of ligands for specific proteins. Deep learning language models have shown significant potential in various domains including protein design and biomedical text analysis, providing strong support for the proposition of DrugGPT. In this study, we employ the DrugGPT model to learn a substantial amount of protein-ligand binding data, aiming to discover novel molecules that can bind with specific proteins. This strategy not only significantly improves the efficiency of ligand design but also offers a swift and effective avenue for the drug development process, bringing new possibilities to the pharmaceutical domain.In our research, we particularly optimized and trained the GPT-2 model to better adapt to the requirements of drug design. Given the characteristics of proteins and ligands, we redesigned the tokenizer using the BPE algorithm, abandoned the original tokenizer, and trained the GPT-2 model from scratch. This improvement enables DrugGPT to more accurately capture and understand the structural information and chemical rules of drug molecules. It also enhances its comprehension of binding information between proteins and ligands, thereby generating potentially active drug candidate molecules.Theoretically, DrugGPT has significant advantages. During the model training process, DrugGPT aims to maximize the conditional probability and employs the back-propagation algorithm for training, making the training process more stable and avoiding the Mode Collapse problem that may occur in Generative Adversarial Networks in drug design. Furthermore, the design philosophy of DrugGPT endows it with strong generalization capabilities, giving it the potential to adapt to different tasks.In conclusion, DrugGPT provides a forward-thinking and practical new approach to ligand design. By optimizing the tokenizer and retraining the GPT-2 model, the ligand design process becomes more direct and efficient. This not only reflects the theoretical advantages of DrugGPT but also reveals its potential applications in the drug development process, thereby opening new perspectives and possibilities in the pharmaceutical field.### Competing Interest StatementS.H., Y.X., Y.L., C.G., X.S. and X.W. have filed a patent application (202310711607.X) relating to the autoregressive model-based drug design method in the name of the First Affiliated Hospital of Xi'an Jiaotong University School of Medicine.

AptaGPT: Advancing aptamer design with a generative pre-trained language model

In silico design of novel aptamers utilizing a hybrid method of machine learning and genetic algorithm

DrugGPT: A GPT-based Strategy for Designing Potential Ligands Targeting Specific Proteins

AptaDiff: de novo design and optimization of aptamers based on diffusion models

Machine learning guided aptamer refinement and discovery

A Novel Nucleic Acid Sequence Encoding Strategy for High-Performance Aptamer Identification and the Aid of Sequence Design and Optimization

In Silico Design and Validation of High-Affinity RNA Aptamers Targeting Epithelial Cellular Adhesion Molecule Dimers

Adapt-cMolGPT: A Conditional Generative Pre-Trained Transformer with Adapter-Based Fine-Tuning for Target-Specific Molecular Generation

Generative aptamer discovery using RaptGen

Identifying high-affinity aptamer ligands with defined cross-reactivity using high-throughput guided systematic evolution of ligands by exponential enrichment

Peptide-GPT: Generative Design of Peptides using Generative Pre-trained Transformers and Bio-informatic Supervision

AptaTrans: a deep neural network for predicting aptamer-protein interaction using pretrained encoders

cMolGPT: A Conditional Generative Pre-Trained Transformer for Target-Specific De Novo Molecular Generation

AptaTRACE: Elucidating Sequence-Structure Binding Motifs by Uncovering Selection Trends in HT-SELEX Experiments

TD-GPT: Target Protein-Specific Drug Molecule Generation GPT

Generative and interpretable machine learning for aptamer design and analysis of in vitro sequence selection

In Silico Discovery of Aptamers with an Enhanced Library Design Strategy.

Design and Prediction of Aptamers Assisted by In Silico Methods

RaptGen: A variational autoencoder with profile hidden Markov model for generative aptamer discovery

A high-dimensional microfluidic approach for selection of aptamers with programmable binding affinities