LinkerMind: An Interpretable, Mechanism-Informed Deep Learning Framework for the De Novo Design of Antibody Drug Conjugate Linkers

Optimising linkers for Antibody-Drug Conjugates (ADCs) is difficult because a linker must stay stable in blood but still cleave efficiently inside target cells. Many computational studies also suffer from biased datasets, where models learn shortcuts such as molecular 1 2 3 4 size rather than true cleavable chemistry. Here, we present LinkerMind, a mechanism- 5 informed framework built on careful data curation and strict validation. We constructed 6 a property-matched dataset of 4,906 molecules and reduced the molecular-weight gap 7 between linkers and decoys from 627 Da to 113 Da, limiting size-based confounding. 8 Wethen trained a multi-modal model that combines structural fingerprints with explicit 9 mechanistic descriptors. The model achieves strong generalisation under scaffold-split 10 testing (ROC-AUC = 0.878), while a Random Forest baseline reaches ROC-AUC = 0.895. 11 Ablation experiments show the model relies on cleavable motifs rather than size, since 12 removing molecular weight causes only a small performance drop. Finally, we integrate 13 the classifier into a generative pipeline and propose novel linker candidates optimised for 14 stability, cleavability, and clinical feasibility. Overall, LinkerMind connects data-driven 15 prediction with chemical reasoning to support next-generation ADC linker design