Today we announce NutraFold v3.0, the latest generation of our protein structure prediction engine. This release represents a significant leap in both accuracy and speed, achieving a mean GDT-TS of 0.92 across CASP15 benchmark targets while reducing inference time by 3x compared to v2.0.
The key architectural innovation in v3.0 is the integration of SE(3)-equivariant transformers with a novel iterative refinement module. Unlike previous approaches that predict structure in a single forward pass, NutraFold v3.0 uses a learned refinement loop that progressively improves atomic coordinates over 8-12 iterations, with each iteration conditioned on the previous prediction and the input MSA features.
Performance highlights include: sub-angstrom backbone RMSD on 67% of test targets (up from 48% in v2.0), improved accuracy on multi-domain proteins with flexible linkers, first-in-class performance on antibody CDR loop prediction (0.89 GDT-TS on SAbDab benchmark), and reliable confidence scores (pLDDT) that accurately distinguish high-quality predictions from uncertain regions.
Speed improvements come from optimized CUDA kernels for the attention mechanism, mixed-precision training on NVIDIA H100 GPUs, and a new MSA subsampling strategy that reduces memory requirements by 60% without sacrificing accuracy. A typical 300-residue protein can now be predicted in under 45 seconds on a single H100.
For drug discovery applications, we have added a structure-based virtual screening module that directly uses NutraFold predictions as input for molecular docking. Preliminary results show that using predicted structures achieves 91% of the docking accuracy obtained with experimental crystal structures, making NutraFold a viable alternative when experimental structures are unavailable.
NutraFold v3.0 is available immediately to all Pro and Enterprise customers. Researcher tier users receive 50 predictions per month. We have also released a comprehensive technical report with full benchmark details and ablation studies on our blog.
We thank our research partners at ETH Zurich, RIKEN, and Harvard Medical School whose collaborative feedback shaped many of the improvements in this release. The future of structural biology is computational, and NutraFold v3.0 brings us closer to the goal of predicting the structure of any protein with experimental-level accuracy.