LocAlign: Local Protein Structural Alignment with Geometric Deep Learning

Identifying common function-determining structural motifs among proteins with different folds is a foundational task in computational biology with no go-to solution. Indeed, standard alignment tools like TM-align are ill-suited for matching small, sequence-order-independent motifs, while specialized tools have limited success. Here, we introduce LocAlign, a local structural alignment algorithm based on geometric deep learning. Given two protein structures, LocAlign iteratively predicts atom-level correspondences and a 3D superimposition. By formulating training as a weakly supervised task on pairs of proteins bound to identical ligands, we bypass the need for ground-truth alignments. We find that LocAlign recovers known functional motifs without explicit supervision, identifying high-quality alignments for 87% and 37% of protein pairs with similar and dissimilar folds, respectively. We show that, equipped with confidence scoring and motif-conditioning capabilities, LocAlign supports diverse applications, including functional annotation of the dark proteome and drug off-target screening. LocAlign is thus a potent, versatile framework for protein functional site comparison.