SiD-Protein

SiD-Protein addresses a practical bottleneck in generative protein design: the cost of sampling. Modern backbone generators built on diffusion and flow matching, such as Proteina, produce high-quality de novo structures but require many sequential denoising steps per sample, making large-scale generation slow. SiD-Protein applies score distillation—compressing a slow, many-step teacher into a fast, few-step student—to protein backbone generation, delivering more than a 20-fold speedup while preserving the quality of the original model.

The method adapts Score Identity Distillation (SiD), a state-of-the-art distillation strategy from image generation, to protein structure. A key finding is that naive single-step distillation is insufficient for proteins: the authors show that multistep generation combined with inference-time noise modulation is essential to retaining the teacher's designability, diversity, and novelty. The released student is a 16-step generator distilled from Proteina.

SiD-Protein was developed by researchers at the University of Texas at Austin and Memorial Sloan Kettering Cancer Center, and released as a preprint in October 2025 with open code and distilled checkpoints.

#Key Features

• Score Identity Distillation for proteins: Adapts SiD, proven in image diffusion, to distill a protein backbone generator into far fewer sampling steps.
• Over 20x faster sampling: The distilled generator produces backbones with more than a 20-fold reduction in sampling time relative to the Proteina teacher.
• Quality-preserving distillation: Designability, diversity, and novelty of generated backbones remain comparable to the teacher rather than degrading with speed.
• Multistep generation with noise modulation: The authors identify that combining a few generation steps with inference-time noise modulation—rather than a single step—is the key to matching teacher quality.
• Open checkpoints: A distilled 16-step generator and PyTorch implementation are publicly released.

#Technical Details

SiD-Protein distills the Proteina backbone generator—specifically its no-triangle, flow-matching variant—using pretrained Proteina checkpoints from NVIDIA's Clara catalog as the teacher. The distilled student is trained to reproduce the teacher's outputs in far fewer steps, with a default 16-step generator whose step count is configurable. Generation quality is measured with the standard de novo backbone metrics: designability (via self-consistency after inverse folding and refolding), diversity across generated samples, and novelty relative to the PDB and AlphaFold databases. Across these metrics the distilled model remains close to Proteina while achieving more than a 20x sampling speedup. Code is released under the Apache-2.0 license and checkpoints are hosted on Hugging Face.

#Applications

SiD-Protein is aimed at protein designers who need to generate large numbers of candidate backbones cheaply—for instance, screening thousands of scaffolds for a binder or enzyme campaign, or embedding structure generation inside iterative design loops where sampling latency is limiting. By cutting per-sample cost by more than an order of magnitude without sacrificing quality, it makes exhaustive backbone exploration and higher-throughput pipelines more feasible on modest compute budgets.

#Impact

SiD-Protein demonstrates that score distillation, which dramatically accelerated image diffusion models, transfers to protein structure generation, and it clarifies what makes that transfer work—multistep sampling with noise modulation rather than single-step collapse. By open-sourcing a fast, quality-matched alternative to a leading backbone generator, it lowers the compute barrier to de novo design. Its benchmarks are in-silico structural metrics; experimental validation of designs remains the ultimate test, as it is for the teacher model it accelerates.