Designing a PD-L1-binding protein in 16 seconds: one DGX Spark is enough

There is a protein design model on Hugging Face called RFD4-Proteína. The most striking set of numbers on its model card is: for designing a PD-L1-binding protein, it produces a design in 16 seconds on a single DGX Spark.

Protein design can be understood as: given a target, generate protein sequences and structures that can bind to it. RFD4-Proteína generates the sequence and complete 3D structure together; inputs can be a target protein, a motif, or a small molecule. PD-L1 is a common immune checkpoint target, and making binding proteins for it has clear application scenarios.

The model card lists 9 examples, 7 of which ran successfully as provided. A single design takes 3 to 46 seconds. The PD-L1 case is 16 seconds.

VRAM requirement is 1.5 GB. A single Spark is more than enough. The H200 time given on the model card is about 7× faster. Being 7× faster matters, but the 1.5 GB threshold is more worth noting: it brings protein design from multi-GPU queues down to a desktop machine.

The video below shows results for several categories of design. Colored chains are new designs; gray is the target. It should be clear: these are computational designs without wet-lab validation—that is, they have not yet actually been made and tested in the lab.

The source is ata's post on X; the model card is on Hugging Face; original post link: https://x.com/atasrb/status/2117711776242442616.

One judgment: when generation speed reaches the second level and VRAM reaches 1.5 GB, the bottleneck will shift toward wet-lab validation and screening. Whether something can be designed is one thing; whether the design can bind, whether it is stable, and whether it has activity is up to the lab.