MolT5-Linker: A Transformer-Based Sequence- and Putative Site-Guided Generative Framework for Antibody-Drug Conjugate Linkers

The rational generation of antibody-drug conjugate (ADC) linkers remains challenging due to the need to balance linker stability, payload release, and compatibility with antibody-payload components. We propose MolT5-Linker, a Transformer-based generative framework conditioned on antibody sequences and payload molecular structures. In the encoder stage, the model integrates a GAT-based Attachment Site Attention Network that infers putative attachment sites from antibody and payload representations. This site-related information guides the decoder to generate chemically compatible linker candidates, improving compatibility between generated linkers and ADC components. Following fine-tuning on a self-constructed ADC data set, MolT5-Linker achieves a generation validity of 0.8986 while maintaining a balance between molecular recovery (0.5802) and uniqueness (0.5438). Furthermore, the generated linker candidates exhibit medicinal chemistry property distributions─including molecular weight, lipophilicity, and topological polar surface area (TPSA)─consistent with ground-truth ADC linkers. These results highlight MolT5-Linker as a computational framework for ADC linker candidate generation.