🤖 machine learning

Emyx: Fast and efficient all-atom protein generation

Emyx is a highly efficient 140M-parameter conditional flow matching model that outperforms state-of-the-art all-atom protein generators like RFdiffusion and ProteinComplexa in enzyme design benchmarks while requiring four times less training compute by leveraging lightweight conditional representations and a novel reparametrisation for optimal sampling.

Original authors: Nicholas J. Williams, Ward Haddadin, Matteo P. Ferla, Constantin Schneider, Nicholas B. Woodall, Ruby Sedgwick, Christian D. Madsen, Andrew L. Hopkins, Edward O. Pyzer-Knapp

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicholas J. Williams, Ward Haddadin, Matteo P. Ferla, Constantin Schneider, Nicholas B. Woodall, Ruby Sedgwick, Christian D. Madsen, Andrew L. Hopkins, Edward O. Pyzer-Knapp

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect trying to build a custom house. But there's a catch: you don't get to design the whole thing from scratch. You are given a specific, non-negotiable room (let's call it the "catalytic room") where a very delicate piece of machinery must sit. Your job is to build the rest of the house (the "scaffold") around this room so that the whole structure is stable, looks unique, and doesn't collapse.

This is exactly what scientists do when they design new enzymes (nature's tiny machines). They need to generate a protein structure that holds a specific chemical "room" in place.

The paper introduces Emyx, a new computer program designed to solve this architectural puzzle faster, cheaper, and better than previous tools. Here is how it works, broken down into simple concepts:

1. The Problem: The "Heavy" Architects

Previous computer models for designing proteins were like trying to build a house using a giant, heavy crane that was originally built to predict what an existing house looks like.

The Issue: These old models were massive, expensive to run (like paying for a luxury construction crew), and often produced houses that looked okay from the outside but had weird, broken walls inside. They were also very similar to each other, lacking creativity.
The Paper's Insight: The authors realized that to build a new house, you don't need the same heavy machinery used to analyze an old one. You just need a lightweight, agile tool that focuses on the specific constraints (the "catalytic room") rather than trying to memorize every possible house ever built.

2. The Solution: Emyx (The "Lightweight" Architect)

Emyx is a new model that is much smaller and smarter.

The "Rep14" Trick: Instead of trying to track every single brick (atom) individually in a messy way, Emyx groups them into "tokens" (like residues). It treats each group as a small, fixed-size box containing 14 slots. If a box doesn't have 14 real bricks, it fills the empty spots with "ghost bricks" that act like placeholders. This keeps the math clean and fast.
Sparse Connections: Imagine a city where every building is connected to every other building by a road. That's a traffic jam (too much data). Emyx only builds roads between buildings that actually need to talk to each other (neighbors or specific chemical bonds). This saves a massive amount of energy and memory.
The "Transformer" Engine: At its heart, Emyx uses a standard, efficient engine (a Transformer) that is very good at understanding patterns, but it strips away all the extra, heavy accessories that other models carry.

3. The Secret Sauce: The "EDM" Upgrade

The paper mentions a technical trick called EDM reparametrisation.

The Analogy: Imagine you are trying to find your way out of a foggy maze.
- Old Way (SDE): You take small, random steps, hoping you don't hit a wall. It's safe but slow and sometimes you get stuck.
- Emyx's Way (EDM): The authors figured out a way to translate their map so they could use a "smart compass" (a method originally designed for a different type of AI). This compass allows them to take bigger, more confident steps and correct their path instantly if they drift off course.
The Result: They didn't have to rebuild the whole model to get this compass; they just changed how they read the map. This made the final designs much more likely to succeed.

4. The Results: Faster, Cheaper, and Better

The authors tested Emyx against two other top-tier models (RFdiffusion3 and Proteína-Complexa) using a strict test called the AME benchmark.

The Test: Can the model build a stable protein around a specific chemical room? And does the whole protein fold correctly?
The Score:
- RFdiffusion3: Solved about 6.7% of the puzzles.
- Proteína-Complexa: Solved about 8.8%.
- Emyx: Solved 13.4%.
The Efficiency: Emyx did this while using 4 times less computing power (training cost) than the next best model. It's like building a better house in half the time with a quarter of the budget.
Creativity: The houses Emyx built were also more unique. While the other models tended to build variations of the same old designs, Emyx created structures that looked more like new, never-before-seen inventions.

5. Why It Matters (According to the Paper)

The paper claims that by simplifying the architecture and using this new "compass" (EDM), they have proven that you don't need a massive, expensive AI to design complex proteins.

Strict Standards: They also introduced a stricter way of grading these designs. They realized that previous tests were too lenient—they would pass a design just because the "catalytic room" looked right, even if the rest of the house was falling apart. Emyx passed the strict test, proving the whole structure is sound.
Accessibility: Because it's cheaper and faster to run, this technology could potentially be used by more researchers, not just those with massive supercomputers.

In summary: Emyx is a lean, mean protein-designing machine. It ditched the heavy, expensive tools of the past, adopted a smarter way of navigating the design process, and proved that a smaller, simpler model can actually build better, more unique, and more stable protein structures than its giant competitors.

Technical Summary: Emyx – Fast and Efficient All-Atom Protein Generation

Problem Statement

Computational enzyme design requires generating protein scaffolds that precisely position catalytic residues and ligands. This task demands both geometric accuracy and structural diversity. Current all-atom generative models, such as RFdiffusion3 and Proteína-Complexa, inherit complex architectures from protein structure prediction (e.g., dense pair representations, heavy embedding stacks). These architectures result in high training costs, limited sample diversity, and often require expensive inverse folding steps to generate high-quality sequences. Furthermore, existing evaluation metrics (e.g., heavy-atom self-consistency RMSD) may overstate success by failing to verify global fold recovery, leading to designs that satisfy local motif constraints but fail to form stable, foldable proteins.

Methodology

Model Architecture

Emyx is a 140M-parameter conditional flow matching model designed for all-atom protein generation. Unlike predecessors that rely on dense pairwise updates and domain-specific embedding layers, Emyx concentrates its capacity within standard transformer blocks.

Representation: The model uses a unified two-level representation. At the atom level, residues and ligands are mapped to a fixed-size "Rep14" matrix (14 atom slots), where unused positions are filled with "ghost atoms" inheriting backbone coordinates. At the token level, residues and HETATMs constitute single tokens.
Connectivity: Instead of dense pair representations, Emyx utilizes a sparse connectivity graph with a fixed edge budget. Edges are constructed based on sequence neighbors, chemical bonds, ligand k-NN, and motif connectivity.
Transformer Design: The architecture employs a "patchify–process–unpatchify" pipeline:
1. Atom Encoder: Lightweight transformer blocks encode local atomic detail within the Rep14 representation.
2. Downcast: Gated cross-attention pools atom representations into token-level embeddings.
3. Token Trunk: A deep transformer trunk (18 layers) processes long-range structural context using sparse self-attention with learned pair biases.
4. Upcast: A second cross-attention broadcasts token context back to atoms.
5. Atom Decoder: Final transformer blocks refine local atomic detail before the output velocity head.
Modifications: The standard DiT (Diffusion Transformer) blocks are modified with bottleneck projections for conditioning, sigmoid gating to bound residual updates, and stochastic depth (drop-path) for regularization.

Training and Objective

Emyx is trained using conditional flow matching over Cartesian coordinates.

Objective: The model learns a time-dependent velocity field $v_\theta(x_t, t, c, G)$ to transport samples from a noise distribution to the data distribution. The training loss combines an $\ell_2$ regression on velocities for non-motif atoms with an auxiliary differentiable local distance difference test (lDDT) loss to enforce structural quality.
Conditioning: The model conditions on sparse geometric constraints (motif geometry, ligand positions) and global features (secondary structure composition, radius of gyration). These features are stochastically masked during training to enable both conditional and unconditional generation.
Efficiency: Training utilizes a linear interpolant between noise and data, which yields lower-variance gradients and faster convergence compared to variance-preserving diffusion schedules. The model was trained on experimental PDB data in approximately 3.55 days on 8 NVIDIA H200 GPUs (681.6 GPU-hours).

Sampling and Reparametrisation

A key methodological contribution is the derivation of an exact reparametrisation of the flow matching interpolant into the EDM (Elucidating the Design Space of Diffusion-based Generative Models) noise-level framework.

This mapping allows the flow matching model to utilize the Karras noise schedule and stochastic churn mechanisms without retraining or architectural changes.
The reparametrisation bridges the training efficiency of flow matching with state-of-the-art sampling methods designed for diffusion models, significantly improving success rates over standard SDE Euler–Maruyama integration.

Key Contributions

Efficient All-Atom Generator: Emyx achieves state-of-the-art performance with a significantly smaller parameter count (140M) and training cost (4× less than RFdiffusion3) by dispensing with expensive prediction-inherited architectures.
Strict Evaluation Protocol: The authors propose a "strict sc-RMSD" protocol that requires both global fold recovery (backbone RMSD < 2.0 Å) and local catalytic geometry accuracy (motif tip-atom RMSD < 1.5 Å). They demonstrate that previous metrics (heavy-atom sc-RMSD) overstate success rates by roughly 2× because they fail to penalize designs that recover the active site but not the global fold.
EDM Reparametrisation: The paper introduces a model-agnostic method to convert flow matching velocity models into the EDM framework, enabling superior sampling strategies (Karras schedule, stochastic churn) without retraining.
Spectral Analysis: The authors provide a spectral analysis revealing a fundamental "generator–predictor gap." Predictors initialized with rich MSA/template signals exhibit higher effective rank and lower overtraining, whereas generators initialized from noised coordinates operate in a lower-rank regime. Emyx demonstrates the lowest overtraining fraction among generators, attributed to its architectural bottlenecks and stochastic depth.

Results

Evaluated against the AME (Atom-level Motif Enzyme) benchmark (41 catalytic active sites):

Success Rate: Under the strict sc-RMSD protocol, Emyx achieves a 13.4% success rate (39/41 targets solved), outperforming Proteína-Complexa (8.8%) and RFdiffusion3 (6.7%).
Structural Novelty: Emyx produces the most structurally novel scaffolds, with a median TM-score of 0.475 to the closest PDB hit (lower is more novel), compared to 0.499 for Proteína-Complexa and 0.512 for RFdiffusion3.
Diversity: Emyx yields the highest number of unique structural clusters (441) among successful designs.
Geometric Validity: Emyx passes geometric validity checks (bond lengths/angles) in 82.2% of designs, significantly outperforming RFdiffusion3 (0.8%) and Proteína-Complexa (49.9%). The authors note that RFdiffusion3's poor geometry scores are likely due to post-processing artifacts where input motif coordinates overwrite generated ones, creating unphysical bonds.
Efficiency: Emyx trains in ~682 GPU-hours (4× faster than RFdiffusion3) and is >2× faster at inference on a single A10G GPU across all chain lengths.

Significance and Claims

The paper claims that architectural simplicity is beneficial for protein generators. Unlike predictors, which require rich co-evolutionary signals (MSA, templates) to initialize complex pair representations, generators condition on sparse geometric constraints. Therefore, the heavy architectures inherited from structure prediction are unnecessary and detrimental to efficiency and diversity.

Emyx demonstrates that concentrating capacity in standard transformer blocks, combined with efficient flow matching training and EDM-style sampling, can achieve superior performance in enzyme design tasks with significantly lower computational resources. The work also highlights the importance of rigorous evaluation protocols (strict sc-RMSD) to distinguish between locally accurate but globally unstable designs and truly viable enzyme scaffolds. Finally, the proposed EDM reparametrisation offers a pathway for other flow matching models to adopt advanced sampling techniques without architectural overhaul.