The Final Piece of the Puzzle

Last week, BindCraft was published in Nature, cementing its place as the breakthrough that democratized protein design. BindCraft has proven itself repeatedly, from winning multiple protein design competitions to achieving remarkable success rates in peptide design as validated by external researchers. We called its release a "Deepseek moment" in AI-powered drug discovery. Indeed, despite being an open-source platform, BindCraft set the state-of-the-art bar for miniprotein design when it was released in late 2024, more than half a year before comparable proprietary models like Chai-2 and Latent-X emerged.

But there was always one dependency that complicated things.

While BindCraft's core components—AlphaFold2, ColabDesign, ProteinMPNN—are genuinely open source, the final scoring step relies on PyRosetta. For academic researchers, RosettaCommons provides free licenses for non-commercial use. But for nascent startups trying to develop therapeutics, the situation becomes complex.

Commercial Rosetta licenses require negotiating with university tech transfer offices, navigating legal agreements, and paying thousands to tens of thousands of dollars annually, even for companies as small as six employees. For a nascent startup, this creates yet another barrier between a great idea and the ability to test it.

Today, we're excited to support a community solution to this problem. Meet FreeBindCraft: a completely open-source implementation of the BindCraft platform.

Understanding Rosetta's Role in BindCraft

When the BindCraft team originally developed their pipeline, they included Rosetta's sophisticated scoring functions for good reason. Rosetta analyzes physics-based properties like hydrogen bonding networks, shape complementarity at interfaces, and surface characteristics. These metrics are designed to predict which designs will successfully fold and bind.

But here's what the FreeBindCraft analysis revealed after examining over 20,000 BindCraft trajectories across dozens of protein design campaigns: For BindCraft's specific use case, Rosetta's contribution is surprisingly modest.

The numbers reveal the story:

• ~90% of designs that fail are rejected by AlphaFold metrics alone, before any PyRosetta scoring
• Surface hydrophobicity violations make up ~3.9% of rejections
• Hydrogen bond network assessments account for ~3.2% of rejections
• Structural alignment checks result in ~1.7% of rejections
• Shape complementarity filters contribute ~0.7% of rejections
• Interface energetic thresholds that Rosetta calculates? Only one rejection observed across ~400,000 analyses

Rosetta has been a historically important tool in molecular modeling and protein design. But for BindCraft's specific workflow of designing miniprotein binders, the data shows that AlphaFold metrics do the heavy lifting.

Building FreeBindCraft: Open Source All the Way Down

Rather than simply removing PyRosetta, FreeBindCraft rebuilds its key functions using open-source alternatives:

The Replacements:

• Structure relaxation: PyRosetta's FastRelax → OpenMM with GPU acceleration (2-4x faster than CPU-bound FastRelax)
• Surface analysis: PyRosetta → FreeSASA and Biopython
• Shape complementarity: PyRosetta → sc-rs (MIT licensed)
• Structural alignments: PyRosetta → Biopython

For hydrogen bond networks, one of the most significant Rosetta-specific filters, the FreeBindCraft team made a pragmatic choice. Evaluating hydrogen bond networks at protein-protein interfaces presents unique challenges, even for sophisticated tools like Rosetta. Interfaces are more solvent-exposed and dynamic than protein cores, making the predictive value of these metrics less certain. Given both the difficulty of accurately reimplementing these calculations and their uncertain contribution to design success, FreeBindCraft elected to focus on more immediately tangible metrics.

Head-to-Head Performance Comparison

FreeBindCraft was benchmarked in a direct comparison to BindCraft using the PDL1 example from the original BindCraft repository:

FreeBindCraft ran nearly 3x faster while producing designs with equivalent AlphaFold confidence scores. The speed boost comes from OpenMM's GPU acceleration compared to Rosetta's CPU-bound FastRelax, plus the fact that FreeBindCraft required fewer trajectories to reach the target number of accepted designs.

The FreeBindCraft designs were also rescored through the original PyRosetta pipeline to understand how they would perform under traditional filtering. 58% passed all traditional filters unchanged. Of the remaining 42%, two-thirds failed due to hydrogen bond assessments, while one-third fell marginally below thresholds for surface hydrophobicity or shape complementarity. Importantly, these "failures" were generally borderline cases—designs that scored just below BindCraft's cutoffs when calculated with Rosetta versus FreeBindCraft's open-source alternatives. FreeBindCraft's open-source calculations align closely, though not identically, with PyRosetta's methods. Such small differences in computational scoring seem unlikely to translate to meaningful differences in experimental outcomes.

These results suggest that FreeBindCraft designs should perform comparably to traditional BindCraft in experimental validation. The majority of designs pass even the original PyRosetta filters, and those that don't are typically borderline cases where minor scoring variations determine acceptance. While we await wet-lab validation, the computational evidence suggests FreeBindCraft provides a viable, fully open-source alternative for protein design.

Using FreeBindCraft on Ariax

Starting today, you can run FreeBindCraft on our platform without any licensing considerations. When setting up your BindCraft job, simply select your preferred scoring method:

• "Open-Source" for FreeBindCraft (no licenses needed)
• "PyRosetta" if you prefer traditional PyRosetta scoring and have an appropriate license

The new scoring method dropdown in the Ariax BindCraft project setup interface

The FreeBindCraft GitHub repository is also live for those who want to run it locally or contribute improvements.

The Path Forward

FreeBindCraft represents another step in our mission to democratize AI-powered drug discovery. As we've argued previously, the real revolution isn't for Big Pharma—they already have the infrastructure to develop practically any molecule they choose. The revolution is for the underdogs: small biotechs, independent researchers, and nascent startups who finally gain access to capabilities that were previously exclusive to pharmaceutical giants.

BindCraft was meant to be the great equalizer in protein design. Yet licensing requirements created an ironic barrier, keeping this democratizing tool out of reach for the very researchers who need it most—the ones willing to pursue novel mechanisms and challenge conventional wisdom while every dollar counts and time is precious.

With FreeBindCraft, we're continuing in our mission to remove every barrier to adoption of AI-powered drug discovery tools: compute complexity, infrastructure costs, and now licensing restrictions. The future of drug discovery will be written by those who can move fast and test bold hypotheses. By ensuring these tools remain accessible to everyone, we're enabling more of these ambitious therapeutic hypotheses to finally be tested.

Ready to design proteins without barriers? Start your first FreeBindCraft run →

FreeBindCraft is available immediately on Ariax Bio. Select "Open-Source" scoring when creating your BindCraft project. For local installation and technical documentation, visit the FreeBindCraft GitHub repository.