Generating Highly Designable Proteins with Geometric Algebra Flow Matching 1. This study introduces GAFL (Geometric Algebra Flow Matching), a cutting-edge generative model for protein backbone design that leverages projective geometric algebra (PGA) and Clifford Frame Attention (CFA) to enhance designability, diversity, and novelty of generated proteins. 2. GAFL’s innovation lies in Clifford Frame Attention (CFA), which extends the AlphaFold2 invariant point attention (IPA) mechanism, incorporating geometric features such as residue frames in projective geometric algebra to create expressive higher-order messages between residues. 3. Unlike traditional IPA, CFA allows GAFL to capture a more realistic secondary structure distribution, with an emphasis on balancing α-helices and β-sheets, matching the natural diversity found in proteins. 4. GAFL outperforms baselines like RFdiffusion, FrameDiff, and FoldFlow across metrics including designability, diversity, and novelty, particularly excelling in producing small, highly designable backbones with diverse secondary structures. 5. Benchmarking on protein datasets such as PDB and SCOPe shows GAFL’s capacity to generate diverse structures more efficiently, achieving high alignment with natural protein structural distributions. @JanStuehmer @graeter_mbm 💻Code: github.com/hits-mli/gafl 📜Paper: arxiv.org/abs/2411.05238 #ProteinDesign #GeometricAlgebra #MachineLearning #ComputationalBiology #GenerativeModel