Something has quietly shifted in the last two years. Designing a new protein that binds a target used to be the kind of project that needed a full academic lab, a trained crystallographer, and a six-month head start. Now it is a weekend project for anyone who knows what a PDB file is. This guide is a roadmap for that weekend.
We will walk through why this matters, how to pick a target, which tools to use, what success looks like, and where to buy the physical DNA when you are ready to test your designs in real life.
Why 2026 is the right moment
Three things happened at roughly the same time that made this democratization possible:
- Generative protein models got good. RFdiffusion, BoltzGen, and AlphaProteo all deliver real wet-lab hit rates, not just pretty visuals.
- Co-folding models got accurate. Boltz-2 and AlphaFold 3 can now predict the bound interface between a designed binder and its target well enough to filter good candidates from bad ones before you ever order DNA.
- GPU compute got cheap and hosted. You no longer need a cluster. A one-off campaign on a serverless A100 costs a few dollars.
The result is that a hobbyist with a laptop and a credit card has access to tools that were science-fiction five years ago.
Step 1: Pick a target you actually care about
The single biggest determinant of whether you finish a weekend project is whether you care about the answer. Pick a target that is personally interesting. Possibilities range from checkpoint inhibitors to viral spike proteins to bacterial toxins to metabolic enzymes. The bar is just that the target should have:
- A crystal or cryo-EM structure in the PDB, preferably at 3 Angstroms or better.
- A defined interface where you know roughly what you want to block.
- No fatal biosecurity concerns. Stick to well-known research targets.
Good starter targets include PD-L1, HER2 extracellular domain, lysozyme, SARS-CoV-2 spike RBD, and CD20 loop regions. If you want a fully worked example, see our PD-L1 tutorial.
Step 2: Compare the available tools
You have three realistic options for the generator stage:
- RFdiffusion — the classic. Easy to understand, lots of tutorials, and mature.
- BoltzGen — the current state of the art, tightly integrated with Boltz-2 co-folding.
- BindCraft — a one-shot pipeline that chains BoltzGen, Boltz-2, and ProteinMPNN for you.
For a first project, use BindCraft. It abstracts away the pipeline plumbing and gives you a ranked list of candidates from a single API call. Once you know what the output looks like, you can run the individual models separately for finer control.
If you want a deeper comparison of the three generators, see our head-to-head post.
Step 3: Run your campaign
Open the Binder Design studio, enter your target PDB ID, click hotspot residues, and press Run. A typical campaign generates 40 to 100 candidates in 5 to 15 minutes on a hosted A100. You can walk away and come back for a coffee.
Step 4: Pick your panel
From the ranked candidate list, pick 8 to 12 designs for wet-lab testing. Prioritize:
- The top 5 by overall pipeline score.
- 2 or 3 designs with different backbone topologies from the top 5.
- 1 negative control (a scrambled sequence) and 1 positive control (a published binder from the literature).
Diversity in your panel matters a lot. If all 10 of your designs are three-helix bundles and the model is wrong about three-helix bundles on your target, you get zero hits and learn nothing. Mix it up.
Step 5: Order and test
Gene synthesis is the bottleneck most hobbyists worry about the most, but it is cheaper than you think. Twist's clonal genes service, IDT's gBlocks, and GenScript's custom genes all deliver in under two weeks for standard sequences. Your 10 designs will cost a few hundred dollars.
Express the resulting plasmids in E. coli with a 6xHis tag, purify by nickel-affinity chromatography, and test binding against your target in a simple assay like biolayer interferometry or a fluorescence polarization readout. A well-organized lab can do the whole cycle in three to four weeks.
What success looks like
A realistic expectation for a first-time panel of 10 designs against a medium-difficulty target is 1 to 3 hits. If you get zero, something went wrong, usually with hotspot choice or target chain selection. If you get 5 or more, you picked an easy target and should move to something harder next time.
A “hit” in this context means a designed protein that shows specific, sub-micromolar binding to the target with sigmoidal behavior on a concentration sweep. You can measure this with standard biophysical tools. You do not need a crystal structure to confirm.
The bigger picture
The barrier to entry for protein design in 2026 is not the models, the compute, or the DNA synthesis. It is knowing that this is possible and being willing to try. Pick a target, run a campaign, and see what the tools can actually do. Even one weekend spent here will change how you think about what is achievable in biology.