#38 · Structural bio & protein design

Context parallelism and TensorRT acceleration scale biomolecular structure prediction

The dominant thread is GPU-level engineering that removes throughput limits in biomolecular modeling, applied to structure prediction and cryo-EM reconstruction. Fold-CP distributes co-folding model inference and training across GPUs using context parallelism, enabling structure prediction of assemblies over 30,000 in size. OpenFold-TRT pairs TensorRT-accelerated OpenFold inference with MMseqs2-GPU homology search, reporting up to 131x speedup over AlphaFold2. torch-projectors is a differentiable PyTorch Fourier-projection library for cryo-EM single-particle and tomography workflows, reported as 1-2 orders faster than torch-fourier-slice. What is new is the shift from algorithmic novelty toward parallelization, runtime optimization, and differentiable GPU-native implementations within standard deep learning frameworks. All three are preprints, so the reported speedups and scaling figures await peer review. torch-projectors addresses a different modality from the two structure-prediction tools.