Protenix-v1: Toward High-Accuracy Open-Source Biomolecular Structure Prediction

Protenix-v1 is the first fully open-source biomolecular structure prediction model that surpasses AlphaFold3 in accuracy while matching its constraints, featuring inference-time scaling, expanded template and RNA capabilities, and a larger 2025-trained variant alongside improved benchmarking tools.

The Gist

Imagine you are trying to build a complex 3D puzzle, but instead of cardboard pieces, you are assembling proteins, RNA, and other tiny biological machines. For a long time, only one company (DeepMind) had the "master blueprint" (AlphaFold3) that could solve these puzzles with near-perfect accuracy. Everyone else was trying to build their own blueprints using open-source tools, but they kept falling short of that master level.

Enter Protenix-v1.

Think of Protenix-v1 as the first fully open-source team that finally built a blueprint that is just as good as, and in some cases better than, the master blueprint. Here is how they did it, explained through simple analogies:

1. The "Fair Race" (Strict Rules)

Usually, when people compare these models, it's like a race where one runner gets a head start or uses a different track. DeepMind's model was trained on data up to a certain date, while others might have peeked at newer data.

Protenix-v1 decided to run a strictly fair race. They said, "We will use the exact same training data cutoff, the same computer power, and the same time limit as the master model." Even with these strict rules, Protenix-v1 didn't just catch up; it started winning in specific categories, proving that the gap between "open-source" and "top-tier" isn't a fundamental law of physics—it was just a gap in engineering.

2. The "Practice Makes Perfect" Superpower (Inference Scaling)

Most open-source models are like a student who studies hard, takes a test once, and submits their answer. If they get it wrong, they can't really fix it without re-studying.

Protenix-v1 is different. It has a superpower called Inference-Time Scaling. Imagine you are trying to guess the shape of a hidden object.

• Old way: You guess once and hope you're right.
• Protenix way: You guess 100 times. You look at all 100 guesses, pick the best one, and realize, "Hey, the more I try, the better my final answer gets!"

The paper shows that if you give Protenix-v1 more computer time to generate more "guesses" (samples), its accuracy goes up in a straight, predictable line. This is a feature previously only seen in the closed-source "master" models. It gives users a "volume knob": want higher accuracy? Just turn up the budget for more guesses.

3. The "Swiss Army Knife" (New Features)

Previous open-source models were like a basic screwdriver. Protenix-v1 is a Swiss Army Knife.

• RNA Support: It can now handle RNA (a cousin of DNA) alongside proteins.
• Templates: It can look at old, similar puzzles to help solve new ones (like using a reference photo).
• Drug Discovery: They released a special "Pro" version (Protenix-v1-20250630) trained on even more recent data. This is like giving a detective a file cabinet that was updated yesterday, making them better at solving brand-new crimes (like designing new drugs for diseases discovered recently).

4. Cleaning Up the Scoreboard (Better Benchmarks)

The authors noticed that the "scoreboards" used to judge these models were messy. Sometimes, a model would fail to run on a specific puzzle, but the scoreboard would still count it as a "pass" because it didn't check the details. This was like grading a math test but ignoring the questions the student didn't answer.

The Protenix team built a new, cleaner scoreboard. They made sure every model was tested on the exact same set of puzzles, and they used statistics to smooth out the "luck" factor. They even created new tests specifically for tricky things like antibodies (the body's defense soldiers) and drug molecules, which were previously too hard to measure accurately.

Why Does This Matter?

• For Scientists: It means they can now use a free, open tool to design drugs and understand diseases with the same confidence as expensive, closed tools.
• For the Future: It proves that open-source collaboration can reach the highest levels of science.
• For You: Better tools mean faster discoveries for new medicines and a deeper understanding of how life works.

In short: Protenix-v1 is the open-source community's "Golden Ticket." It's a free, powerful, and adaptable tool that finally plays on the same field as the giants, but with the added benefit of being transparent, customizable, and constantly improving.

Technical Summary

1. Problem Statement

Despite the rapid expansion of the open-source ecosystem for biomolecular structure prediction, a significant performance gap persists between current open-source models and AlphaFold3 (AF3), the state-of-the-art closed-source model.

• The Gap: Existing open-source models generally underperform AF3, particularly on complex tasks like antibody-antigen interactions and protein-ligand docking.
• Evaluation Issues: Current benchmarking practices suffer from methodological flaws, including inconsistent data coverage (different models failing on different subsets), lack of variance reporting, and dataset bias due to small sample sizes in specific domains (e.g., antibody complexes).
• Transparency: The closed nature of AF3 prevents systematic, fair comparisons and hinders the community's ability to understand the fundamental limits of open-source approaches.

2. Methodology

The Protenix Team developed Protenix-v1, a fully open-source model designed to match or exceed AF3's performance under strictly controlled conditions (identical training data cutoff, model size, and inference budget).

Core Architecture & Training

• Architecture: Protenix-v1 utilizes a diffusion-based inference mechanism consistent with AF3's hyperparameters.
• Data Processing: The model incorporates enhanced data pipelines, including:
  • RNA MSA Support: Explicit support for RNA Multiple Sequence Alignments.
  • Protein Template Integration: Leveraging structural templates.
  • Distillation: Expanded disorder-focused distillation and large-scale monomer distillation based on MGnify data.
• Training Variants:
  • Protenix-v1: Trained with a data cutoff matching AlphaFold3 for fair benchmarking.
  • Protenix-v1-20250630: A variant trained on an expanded dataset (cutoff: June 30, 2025) to optimize for real-world applications like drug discovery where recent structural data is available.

Inference Strategy

• Sampling: The model employs a stochastic diffusion process.
• Scaling Behavior: A key methodological feature is inference-time scaling. By increasing the sampling budget (e.g., generating hundreds of candidates via random seeds), the model demonstrates consistent, log-linear improvements in prediction accuracy, particularly for challenging targets like antibody-antigen complexes.
• Confidence Filtering: The model utilizes confidence scores (ipTM, pTM, pLDDT) to select high-quality structures from the sampled ensemble, significantly boosting performance over median predictions.

Evaluation Framework

To address benchmarking flaws, the authors introduced:

• Strict Intersection Protocols: Evaluations are performed only on the common intersection of successfully evaluated samples across all models to ensure fair comparison.
• Variance-Aware Protocols: Extensive bootstrapping (e.g., 20 random seeds) is used to calculate confidence intervals, mitigating stochastic variance in data-scarce domains.
• New Benchmarks:
  • PXM Suites: Year-stratified benchmarks (PXM-2024, PXM-2025, PXM-2025H2) covering diverse modalities (Monomer, Protein-Protein, Antibody-Antigen, Ligand, RNA, DNA).
  • Corrected FoldBench: A re-evaluation of the FoldBench dataset using common intersections and bootstrapping.

3. Key Contributions

1. Performance Parity with AlphaFold3: Protenix-v1 is the first fully open-source model to achieve AF3-level performance under matched constraints, proving the performance gap is not fundamental but solvable.
2. Inference-Time Scaling: The paper demonstrates that open-source models can exhibit the "scaling law" behavior previously observed only in AF3, where increased computational budget (sampling) directly translates to higher accuracy.
3. Methodological Rigor in Benchmarking: The authors identified and corrected critical flaws in existing benchmarks (data coverage bias, lack of variance reporting) and released updated tools and year-stratified datasets (PXM) for the community.
4. Dual-Release Strategy: Providing both a "benchmark-aligned" model (strict cutoff) and a "deployment-optimized" model (expanded data cutoff) to serve both research comparison and practical application needs.
5. Enhanced Capabilities: Integration of RNA MSA and protein templates, expanding the model's utility beyond simple protein folding.

4. Results

Protenix-v1 was evaluated across multiple domains and benchmarks, consistently outperforming other open-source baselines (Chai-1, Boltz-1, Boltz-2) and often surpassing AlphaFold3 in specific categories.

• FoldBench (Corrected):
  • Protenix-v1 achieved a DockQ Success Rate (SR) of 72.70% on protein-protein interfaces, outperforming AF3 (71.73%) and other open-source models.
  • On Antibody-Antigen (Ab-Ag) tasks, Protenix-v1 reached 52.31% SR, significantly higher than AF3 (48.75%) and Chai-1 (23.12%).
• PXM Benchmarks (2024/2025):
  • Protenix-v1 showed a 10–30% relative improvement over leading open-source baselines on protein-protein interface prediction.
  • Confidence Selection: Selecting structures based on confidence scores improved median performance by 10–30% across domains, with the largest gains in Ab-Ag and Protein-DNA tasks.
• Inference-Time Scaling:
  • On the PXM-22to25-Antibody benchmark, increasing the seed count from 1 to 80 improved the DockQ SR from 36.01% to 47.68%, demonstrating a clear pathway for optimization via computational budget.
• Protein-Ligand & RNA:
  • Achieved the lowest Ligand RMSD (2.772 Å) and highest RMSD < 2Å success rate (72.79%) among evaluated models.
  • Demonstrated strong performance in Protein-RNA and Protein-DNA docking, outperforming all open-source baselines.
• Design Filter Performance:
  • In a binary classification task (distinguishing binding vs. non-binding designs), Protenix-v1 achieved the highest AUC and Average Precision across nearly all targets, validating its utility in drug discovery workflows.

5. Significance

• Democratization of SOTA: Protenix-v1 breaks the monopoly of closed-source models, providing the scientific community with a tool that rivals AlphaFold3 in accuracy while remaining fully open and reproducible.
• Paradigm Shift in Evaluation: The paper sets a new standard for how biomolecular models should be evaluated, emphasizing statistical rigor, common data intersections, and variance reporting. This prevents misleading conclusions caused by dataset bias.
• Practical Impact: The release of the Protenix-v1-20250630 variant and the demonstrated ability to scale inference time make the model immediately applicable to real-world challenges in drug discovery and antibody design.
• Future Direction: The identification of inference-time scaling as a key performance driver suggests that future open-source research should focus not just on model architecture, but on efficient sampling strategies and distilling high-quality sampling signals into base models.

In conclusion, Protenix-v1 represents a major milestone in open-source AI for biology, closing the gap with proprietary models and establishing a robust, transparent framework for future advancements in biomolecular structure prediction.