FrameBench: A Principled, Symmetry-Aware Evaluation of SE(3) Protein Generators

Wenran LI, Xavier Cadet, Cédric Damour, Yu LI, Alexandre G. de Brevern, Alain Miranville, Frederic CADET

Keywords: SE(3) group; protein generation; diffusion model; flow matching

Abstract: We propose an evaluation framework for the $SE(3)$ group symmetry diffusion model and flow matching models: FrameDiff, FrameFlow, and FoldFlow. We show how they expose trade-offs among diversity, domain stability, and compute-normalized efficiency. We instantiate this principle in FrameBench, which treats diversity, stability, and efficiency as quantities of the generated protein. Concretely, FrameBench reports (i) diversity via CATH entropy and coverage; (ii) stability via SWORD2-based domain partitions; and (iii) fairness via compute-normalized efficiency. Evaluating FrameDiff, FrameFlow, and FoldFlow under matched length/compute budgets reveals clear trade-offs, e.g., coverage versus stability, that simple wall-clock timings miss.

Submission Number: 37