The stability and therapeutic efficacy of antibody-drug conjugates (ADCs) are critically determined by the chemical linkers that connect the antibody to the cytotoxic payload, which is a key factor influencing drug release, plasma stability, and off-target toxicity. However, the current linker design space remains highly constrained, with most approved ADCs relying on a narrow set of established motifs. This limitation highlights an urgent need for computational tools capable of generating structurally diverse and synthetically accessible linkers. In this study, we introduce Linker-GPT, a Transformer-based deep learning framework leveraging self-attention mechanisms to generate novel ADC linkers with high structural diversity and synthetic feasibility. The model integrates transfer learning from large-scale molecular datasets and reinforcement learning (RL) to iteratively refine molecular properties such as drug-likeness and synthetic accessibility. During transfer learning, a pre-trained model was fine-tuned on a curated linker dataset, yielding molecules with high validity (0.894), novelty (0.997), and uniqueness (0.814 at 1k generation). RL further optimized the model to prioritize synthesizability and drug-like properties, resulting in 98.7% of generated molecules meeting target thresholds for QED (> 0.6), LogP (< 5), and synthetic accessibility score (SAS < 4). Linker-GPT demonstrates strong potential as a computational platform for accelerating the discovery and optimization of novel ADC linkers, offering a scalable solution for early-stage linker design. While these results are currently computational, they provide a foundation for future experimental validation and optimization.

Antibody-drug conjugates; Linker design; Machine learning; Molecule generation; Reinforcement learning; Transfer learning; Linker design; Machine learning; Molecule generation; Reinforcement learning; Transfer learning