Tmol: GPU-accelerated Rosetta energy function in PyTorch, up to 362x faster

ArashVahdat · x · 2026-11-20

Tmol is a GPU-native PyTorch library that reimplements the RosettaGenFF forcefield as batched tensor ops, making Rosetta-style all-atom energetics practical inside modern ML training and inference pipelines with autograd-ready derivatives. Reported speedups reach 233–362x for score+gradient and 38–102x for torsion-space relax vs single-threaded PyRosetta, while evaluating 60% more rotamers during relax. Its core is PoseStack, a GPU-friendly batch representation using dense coordinate tensors [nposes, maxatoms, 3] plus padded per-block tensors, with systems decomposed into blocks (residues/ligands/fragments). It supports batched differentiable scoring and whole-structure refinement (packing + minimization + FastRelax).

Related event: Baker Lab open-sources tmol: GPU-native Rosetta energy function with up to 362x throughput(6 posts)→

Original post →

More from Research

Research channel →