In 2023, RFdiffusion was the only serious open tool for de novo protein binder design. In 2026, the landscape looks very different. BoltzGen has joined as a tightly integrated partner to the Boltz-2 co-folding model. DeepMind shipped AlphaProteo as a closed commercial offering. And a handful of smaller generators are chasing niche targets.
This post compares the three headline models head to head, so you can pick the right starting point for your project. If you only want the short answer, scroll to the bottom. Otherwise let us work through the details.
The contenders, in one paragraph each
RFdiffusion
The 2023 Baker Lab model that made diffusion-based protein design practical. Open source, well documented, and the foundation of an entire ecosystem of follow-up tools. RFdiffusion2, the 2025 upgrade, improved multi-chain conditioning and enzyme active-site scaffolding. Still the default for classical binder campaigns. Read our RFdiffusion primer for a deeper intro.
BoltzGen
A newer generative family from the Boltz team, designed to share representations with the Boltz-2 co-folding model. BoltzGen generates a binder backbone and the interface with the target in the same forward pass, which gives it a much tighter coupling between generator and validator than the classic RFdiffusion-plus-AlphaFold pipeline. Open source and hosted on SciRouter.
AlphaProteo
DeepMind's entry into the field, released in 2024. Not open source, not self-hostable. Available only through DeepMind's commercial offering. Reports competitive in-silico and wet-lab hit rates, but direct comparisons are hard because the code and the benchmark splits are not public.
Feature comparison
License and access
- RFdiffusion: open source with a permissive license. Weights, training data description, and code are public. Can be fine-tuned.
- BoltzGen: open source with a permissive license from the Boltz team. Weights are public. Can be fine-tuned.
- AlphaProteo: closed source. Available only through a hosted commercial API from DeepMind. Cannot be fine-tuned.
Hit rates (rough, order of magnitude)
We have to be careful here because each team benchmarks on different targets with different definitions of success. With that caveat, the landscape looks like this:
- Vanilla RFdiffusion plus ProteinMPNN plus AlphaFold: single-digit to low double-digit percent wet-lab success on most targets.
- BindCraft (BoltzGen plus Boltz-2 plus ProteinMPNN): higher double-digit percent in-silico and 10 to 30 percent wet-lab success across the reported benchmarks.
- AlphaProteo: reports in the same range as BindCraft on DeepMind's internal benchmarks, though without public head-to-head runs on identical targets.
Compute cost
- RFdiffusion: runs on a single A100. About 30 seconds per design. A 40-design batch finishes in 20 minutes.
- BoltzGen: also runs on a single A100, but joint generation is slightly heavier than pure RFdiffusion. About 45 seconds per design. Similar batch time.
- AlphaProteo: DeepMind does not publish compute details. You pay a per-call fee through the hosted offering.
Ease of use
RFdiffusion has the largest community and the most tutorials. It is the safest place to learn the concepts, and there are preset configs for most common workflows.
BoltzGen is newer, but the BindCraft wrapper is the easiest end-to-end pipeline in the field. You supply a target PDB and hotspot residues and it returns ranked candidates without any tuning.
AlphaProteo is only accessible through DeepMind's interface. If you want code-level control you cannot get it.
When to use each
Pick RFdiffusion when
- You want the most mature documentation and tutorials.
- You are teaching or learning and want to understand each step of the pipeline.
- You need enzyme active-site scaffolding, where RFdiffusion2 is still the best open option.
- You need to fine-tune on proprietary structures that cannot leave your cluster.
Pick BoltzGen (via BindCraft) when
- You want the highest hit rate on a standard binder target.
- You do not want to manage the generator and validator as separate pipeline stages.
- You are running many targets in parallel and need consistent ranking across campaigns.
- You are new to binder design and want a one-shot API. See the BindCraft tutorial for step-by-step instructions.
Pick AlphaProteo when
- You are already locked into DeepMind's platform for other reasons.
- You do not need self-hosting, fine-tuning, or open weights.
- You want to benchmark your open pipeline against a well-regarded closed system.
The meta lesson: use more than one
The single biggest thing you can do to improve your hit rate is not to optimize one generator. It is to run two or three generators in parallel and merge their outputs before filtering. Diversity in the candidate pool beats sophistication in any single model.
A practical modern campaign looks like this:
- Generate 40 designs with BoltzGen, ranked by Boltz-2 metrics.
- Generate 40 designs with RFdiffusion, ranked by AlphaFold pLDDT.
- Merge, deduplicate, and apply ProteinMPNN for sequences.
- Re-score with Boltz-2 and sort by predicted interface quality.
- Pick the top 10 unique designs for wet-lab testing.
On SciRouter you can do all of this from a single dashboard. See the Binder Design studio for the managed workflow, or wire the individual endpoints together in Python.
Bottom line
RFdiffusion is still the friendliest place to learn. BoltzGen inside BindCraft is the highest hit rate you can get on most targets today. AlphaProteo is competitive but closed. If you are starting a new project, begin with BindCraft and add RFdiffusion as a second generator when you need more diversity.