Triangle Multiplication Is All You Need For Biomolecular Structure Representations

A major bottleneck lies in the Pairformer backbone of AlphaFold3-style models, which relies on computationally expensive triangular primitives--especially triangle attention--for pairwise reasoning. AlphaFold (Senior et al., 2020; Jumper et al., 2021) has transformed protein structure prediction and At this scale, runtime and memory efficiency are critical bottlenecks: for example, Boltz-1 (Wohlwend et al., 2024) requires over 15 minutes to process a single 2048-token We introduce Pairmixer, a streamlined alternative to the Pairformer backbone of AlphaFold3 (Abram-son et al., 2024). Figure 1: Pairmixer is an efficient architecture for biomolecular structure prediction. By reducing both runtime and memory requirements, Pairmixer expands the scope of feasible downstream applications of structure prediction. Early approaches such as trRosetta (Y ang et al., 2020) and AlphaFold1 (Senior et al., 2020) leveraged convolutional neural networks to extract pairwise residue The Pairformer has since become the de-facto backbone architecture for biomolecular structure prediction (IntFold et al., 2025; Boitreaud et al., 2024; Wohlwend et al., 2024; ByteDance et al., 2025).