Protenix-v2: Broadening the Reach of Structure Prediction and Biomolecular Design

Protenix-v2 is a robust biomolecular modeling system that significantly advances high-accuracy structure prediction and design, achieving record-breaking success rates in antibody-antigen modeling, GPCR target discovery, and diverse therapeutic applications like SARS-CoV-2 binder design.

The Gist

Imagine you are a master architect trying to design a key that fits a very specific, complex lock. In the world of medicine, the "locks" are proteins on the surface of viruses or cancer cells, and the "keys" are antibodies or drugs designed to stop them.

For a long time, figuring out exactly what these locks look like and designing the perfect key was like trying to solve a 3D puzzle in the dark. You had to guess the shape, build a prototype, test it, and if it didn't fit, start over. This took years and cost millions of dollars.

Protenix-v2 is like a super-powered, AI-driven architect that has just leveled up from "good guesser" to "master builder." Here is what this new system can do, explained through simple analogies:

1. The Crystal Ball (Structure Prediction)

Before you can build a key, you need to know exactly what the lock looks like.

• The Old Way: Previous AI models were like looking at a blurry photo of a lock. They could tell you it was a lock, but the details were fuzzy. If you tried to build a key based on that blurry photo, it often wouldn't fit.
• The Protenix-v2 Way: This new model is like a high-definition, 3D hologram. It can predict the exact shape of complex biological "locks" (specifically antibody-antigen interactions) with incredible precision.
• The Magic: It's not just accurate; it's fast. In the past, to get a good answer, you might have to run a simulation 1,000 times (like rolling dice 1,000 times to get a lucky number). Protenix-v2 gets a better result in just 5 tries. It's like having a master archer who hits the bullseye on the first five shots, whereas others needed a thousand arrows to get close.

2. The Master Key Designer (Zero-Shot Design)

Now that we know what the lock looks like, we need to design the key from scratch.

• The Challenge: Designing a key that fits a lock you've never seen before is incredibly hard. Most AI systems need to be taught with thousands of examples of similar locks first.
• The Protenix-v2 Way: This system is a "Zero-Shot" designer. Imagine being handed a blueprint for a brand-new, strange lock you've never seen, and the AI immediately draws a working key without needing to study a library of old keys first.
• The Results: In tests, when asked to design keys for new targets, it succeeded 100% of the time (at least one working key was found for every target). Even better, about half of the keys it designed actually worked in the lab immediately.

3. Tackling the "Impossible" Locks (GPCRs)

Some locks are notoriously difficult. They are small, wiggly, and hidden (like G-protein coupled receptors, or GPCRs). Traditional drug discovery often gives up on these because they are so hard to grab.

• The Analogy: Trying to design a key for a GPCR is like trying to grab a slippery, tiny pebble in a fast-moving river with a giant net.
• The Breakthrough: Protenix-v2 managed to design keys for these "slippery pebbles" with a success rate of up to 88%. It found working keys even when the team only had a very small budget to test them (like having only 20 chances to try a key in a lock).

4. The "Chemical Reality Check" (Ligand Plausibility)

Sometimes, AI designs a key that looks perfect on a computer screen but is physically impossible to build in real life (like a key made of twisted metal that would snap instantly).

• The Fix: Protenix-v2 has a new "reality check" feature. It doesn't just look at the shape; it checks the chemistry. It ensures the atoms are arranged in a way that obeys the laws of physics and chemistry. It's like an architect who not only draws a beautiful building but also ensures the beams can actually hold the weight and the materials won't melt in the sun.

5. The "Universal Key" (Cross-Variant Design)

Viruses mutate; they change their locks to avoid our keys.

• The Challenge: If a virus changes its shape (like the Omicron variant of SARS-CoV-2), your old key might not fit anymore.
• The Solution: Protenix-v2 successfully designed "universal keys" that fit both the original version of the virus and the mutated version. It's like designing a key that works on both a standard door and a slightly warped door, ensuring your security system stays effective even as the enemy changes.

Why This Matters

Think of drug discovery as a race against time and complexity.

• Before: Scientists were running a marathon in heavy boots, tripping over every step, and taking years to find a single working drug.
• Now: Protenix-v2 is like giving them a pair of jet-powered boots and a perfect map. It allows scientists to skip the years of trial and error, design drugs that actually work, and tackle diseases that were previously considered "impossible" to treat.

In short, Protenix-v2 is a massive leap forward in turning the science of "guessing" into the science of "knowing," speeding up the journey from a computer screen to a life-saving medicine.

Technical Summary

1. Problem Statement

Biomolecular modeling faces two critical challenges in drug discovery:

1. Structure Prediction Accuracy: While general protein structure prediction has advanced, predicting the precise geometry of antibody-antigen interfaces remains exceptionally difficult due to conformational flexibility and the complexity of binding pockets. Furthermore, existing models often struggle with ligand plausibility, generating small-molecule poses that satisfy geometric metrics but fail basic chemical validity (e.g., incorrect chirality or bond planarity).
2. Zero-Shot Design Efficiency: Moving from structure prediction to de novo design requires models that can generate novel binders for diverse targets (including difficult membrane proteins like GPCRs) without extensive experimental retraining. Previous efforts often lacked the ability to simultaneously achieve high hit rates, structural diversity, and "developability" (biophysical properties suitable for drug development) in a zero-shot setting.

2. Methodology

Protenix-v2 is a unified biomolecular modeling system that operates in two primary modes: structure prediction/ranking and generative design.

• Training & Data Cutoff: The model was trained without any wwPDB entries released on or after September 30, 2021. This ensures that all reported results on structure prediction and zero-shot design are evaluated against truly novel, held-out data, preventing data leakage.
• Architecture & Optimization: Protenix-v2 retains the input-output settings of its predecessor (Protenix-v1) but incorporates architectural refinements and training optimizations.
• Generative Design Capabilities:
  • Supports flexible, target-conditioned generation for diverse formats: miniproteins, VHH (single-domain antibodies), Fv (VH+VL), and full-length mAbs.
  • Allows granular control over binding regions (e.g., independent assignment of CDR loops with specified lengths) and integration of predefined scaffolds.
• Training-Free Guidance (TFG) for Ligands: To address ligand plausibility, the authors introduced Protenix-v2-TFG. Drawing from diffusion models and classical molecular dynamics constraints (like SHAKE), this variant imposes constraints during inference on:
  • Chirality
  • Planarity (around sp2 carbons and amide groups)
  • Non-planarity (at sp3 centers)
  • Torsional geometry and pairwise distances.
  • Note: This is applied without retraining the base model, using inference-time steering.

3. Key Contributions

The paper presents four major advancements:

1. Superior Antibody-Antigen Structure Prediction: Significant gains in predicting interface geometry compared to AlphaFold3, Boltz-1, and Protenix-v1.
2. High-Efficiency Sampling: Demonstrates that 5 inference seeds in Protenix-v2 outperform 1000 seeds in previous models, drastically reducing computational cost.
3. Robust Zero-Shot Antibody Design: Achieves high hit rates on novel soluble targets and, crucially, on difficult GPCR targets (small, flexible epitopes) in both VHH-Fc and full-length mAb formats.
4. Enhanced Chemical Realism & Breadth: Improved ligand plausibility via TFG and successful design of dual-specific binders against SARS-CoV-2 variants (Prototype and Omicron).

4. Key Results

A. Structure Prediction Performance

• Antibody-Antigen Benchmarks: Tested on three collections (PXMeter-AB, FoldBench-AB, AF3-AB).
  • Protenix-v2 achieved 9–13 percentage point gains over Protenix-v1 at the DockQ > 0.23 threshold.
  • It showed comparable large gains in the high-precision regime (DockQ > 0.8).
  • Sampling Efficiency: Protenix-v2 at 5 seeds surpassed the performance of Protenix-v1 at 1000 seeds, indicating a massive leap in inference efficiency.
• Ligand Plausibility: On the PXM-22to25-Ligand benchmark (recent structures), Protenix-v2-TFG achieved a 60.46% joint success rate (RMSD < 2Å + Validity) under a strict revised validity criterion. This outperformed Boltz-1x (53.96%) and approached Boltz-2x (62.86%), while correcting specific chemical failures like twisted amides and distorted rings.

B. Zero-Shot Antibody Design

• Soluble Targets:
  • Achieved a 100% target-level success rate (at least one confirmed binder per antigen).
  • BLI-confirmed hit rates ranged from 2% to 48% across novelty-filtered targets.
  • Developability: Hits showed exceptional pass rates: 100% (thermostability), 98% (self-interaction), and 93% (polyreactivity).
  • Diversity: Successful binders spanned multiple structural clusters, avoiding repetitive solutions.
• Challenging GPCR Targets:
  • Tested on 4 GPCRs with limited budgets (16–30 designs per target).
  • VHH-Fc Hit Rates: 16% – 88%.
  • mAb Hit Rates: Up to 50%.
  • One VHH-Fc binder against GPRC5D achieved a 112 pM KD (measured under avidity conditions).
• Ranking Capability: Model rankers (Ranker A/B) identified more high-response binders than human experts, though humans selected more structurally diverse candidates. The models captured features orthogonal to human intuition.

C. Multi-Variant Design

• SARS-CoV-2 Mini-Binders: Designed dual-specific mini-binders targeting both Prototype and Omicron (B.1.1.529) RBDs.
  • 2 out of 4 tested designs showed nanomolar-scale KD against both variants.
  • Structural analysis revealed compensatory mechanisms (e.g., reorientation of residues to accommodate the Q493R mutation in Omicron).

5. Significance

Protenix-v2 represents a significant step forward in accelerated drug discovery by bridging the gap between high-accuracy structure prediction and practical biomolecular design.

• Therapeutic Impact: It enables the discovery of binders for "undruggable" or difficult targets like GPCRs, which have historically been challenging for antibody discovery due to small, flexible epitopes.
• Efficiency: The dramatic improvement in sampling efficiency (5 seeds vs. 1000) makes high-throughput computational screening more feasible and cost-effective.
• Chemical Reliability: The introduction of Training-Free Guidance (TFG) addresses a critical gap in AI-driven drug design: ensuring that predicted structures are not just geometrically accurate but also chemically valid, reducing the risk of experimental failure due to impossible molecular geometries.
• Versatility: The system's ability to handle diverse formats (VHH, mAb, mini-binders) and targets (soluble, membrane, viral variants) establishes it as a robust, general-purpose platform for next-generation therapeutic development.