Transformer models have become central to biological data analysis, yet most genomic workflows still incorporate linkage disequilibrium (LD) through explicit preprocessing, typically relying on pairwise correlation estimates that are computationally demanding and task-specific. Here, we introduce LD-Attention, a generalized LD-aware attention layer as a modular and reusable transformer component. By embedding LD directly within attention, the model learns locus-to-locus dependency structure implicitly during end-to-end training, removing the need for precomputed LD matrices. In this formulation, LD is not treated as an external feature but as an inductive bias encoded within the architecture, enabling attention to align with the correlation structure of genetic variation. The resulting framework is task-agnostic and readily extensible across LD-dependent analyses, including genotype imputation, genome-wide association studies, fine-mapping, haplotype phasing, and polygenic risk prediction. By abstracting LD awareness into a portable attention layer, LD-Attention provides a scalable foundation for integrating LD structure into transformer-based models for genetic data.

Co-authors: Davoud Torkamaneh