ProteinEBM Online: Score Protein Structures and Explore Folding Dynamics

Introduction

Protein structure predictors usually try to return one answer. ProteinEBM takes a different approach: it learns a sequence-conditioned energy landscape over protein conformations. That learned landscape can be queried to score an uploaded structure, rank competing models, compare alternative conformations, estimate fixed-backbone mutation effects, or generate coarse-grained folding trajectories [1].

Neurosnap exposes those capabilities through two services. ProteinEBM Scoring evaluates coordinates supplied by the user, while ProteinEBM Dynamics uses gradients of the learned energy to explore conformational space. This tutorial connects both services using NuG2, an engineered fast-folding variant of the Protein G B1 domain that is also analyzed in the ProteinEBM paper.

From Structure Prediction to an Energy Landscape

For a protein sequence s and conformation x, ProteinEBM learns a scalar energy E_θ(x, s, t), where t is the diffusion noise level. Conceptually, lower-energy conformations should be more compatible with the sequence under the learned model:

The normalization constant is not evaluated, so the useful quantities are relative energies and rankings rather than absolute probabilities. ProteinEBM is trained by denoising score matching, but unlike a standard diffusion model its structural score is explicitly the negative coordinate gradient of a scalar energy:

That scalar energy is the connection between the two Neurosnap services. It can rank fixed input structures, while its gradient can guide reverse diffusion or Langevin-style sampling.

What ProteinEBM Learned

The published model is an approximately 85-million-parameter, sequence-conditioned backbone model derived from diffusion modules used in AlphaFold3 and Boltz-1 [1]. It was pretrained on experimental CATH domains, AlphaFold Database domains defined with TED, and protein complexes. Selected variants were then fine-tuned on molecular-dynamics frames to improve conformational sampling.

The Neurosnap scoring workflows use the paper's expert ProteinEBM-x checkpoint. The recommended Scoring Time of 0.05 is a diffusion noise level, not a runtime setting. It is the low-noise level used for the paper's principal structure-ranking results.

ProteinEBM Scoring and Dynamics

Example: NuG2, an Engineered Protein G Variant

NuG2 is a 56-residue engineered variant of the Protein G B1 domain. Wild-type Protein G preferentially forms its C-terminal hairpin during the folding transition, whereas stabilization of the N-terminal hairpin in NuG2 shifts its pathway toward earlier N-terminal contact formation [2,3]. The ProteinEBM authors use PDB 1MI0 as the NuG2 input for their direct-folding example [1,4].

This makes NuG2 useful for more than a runtime demonstration. It lets us ask whether ProteinEBM assigns sensible relative energies to structural alternatives and whether its sampled trajectories recover a qualitative pathway reported experimentally and in the paper.

When to Use ProteinEBM Online

Use Structure Scoring when you need a baseline energy for one backbone, and Decoy Ranking when you already have several models of the same sequence. Conformational Comparison adds a relative-to-minimum column for alternative states. Mutation Stability is useful for rapid fixed-backbone triage when you want to ask whether a residue identity is compatible with a supplied conformation.

Use Fast-Folder Dynamics for broad sampling from high noise through progressively finer energy landscapes. Use Folding Simulation when the question is specifically how an initially unfolded chain moves under the low-noise expert energy. Neither mode replaces atomistic MD: ProteinEBM operates on a learned coarse-grained backbone representation and its trajectory steps are not calibrated physical time.

Configure the Scoring Jobs

The first baseline job scored the exact cleaned NuG2 structure distributed with ProteinEBM. We used the recommended low-noise settings:

The completed job returned an energy of 1524.279. That number is a baseline for subsequent NuG2 comparisons, not a standalone measure of quality: ProteinEBM energies have no useful interpretation without structurally and procedurally matched alternatives.

Open the shared Structure Scoring job, or inspect its live summary.csv and metadata.json.

Reading Structure and Decoy Energies

The standalone NuG2 score was 1524.279. ProteinEBM energy is not expressed in kcal/mol, and its zero point is not physically calibrated. A single value cannot tell us whether a structure is correct or stable. Structure scoring is most useful when the same job ranks matched candidate structures under identical settings.

The paper demonstrates that use case on Rosetta decoys. Across its test set, ProteinEBM-x energy had a mean Spearman correlation of 0.838 with decoy TM-score, compared with 0.757 for Rosetta energy [1]. The mean TM-score of the minimum-energy ProteinEBM decoy was 0.905, while Rosetta's top-ranked decoy averaged 0.899; the latter top-one difference was not statistically significant. The result supports ProteinEBM as a ranking function, not as a guarantee that every lowest-energy model is native.

We did not reproduce that benchmark here. The official Rosetta decoy archive used by the paper is 5.18 GB and does not provide a NuG2-matched subset. Creating arbitrary coordinate distortions would produce an attractive plot but a weak benchmark, so this tutorial keeps the NuG2 evidence to traceable service runs.

Fixed-Backbone Mutation Screening

The 1MI0 structure record identifies Ala47 in the cleaned ProteinEBM numbering as an engineered replacement for the parent Asp [4]. We therefore tested the reverse mutation A47D, using the same 0.05 scoring time and template self-conditioning as the baseline run.

Within this paired job, replacing Ala47 with Asp lowered the learned energy by 76.979 model units. The cautious interpretation is that ProteinEBM scores Asp as more compatible with the fixed NuG2 backbone than Ala at this position. It does not show that A47D experimentally stabilizes the protein: the backbone and conformational ensembles were not relaxed or resampled.

The wild-type value in this job (1573.236) also differs from the standalone structure score (1524.279). The worker randomly centers and rotates coordinates before evaluation, while ProteinEBM uses a non-equivariant architecture trained with spatial augmentation. For that reason, compare wild type and mutant from the same paired job rather than subtracting results from separate submissions.

This Neurosnap mode is deliberately simpler than the stability protocol reported in the paper. The paper approximates mutation ΔΔG using both folded and heuristic unfolded-state contributions. The service changes the amino-acid token while retaining the uploaded backbone coordinates, so its result is best described as a fixed-backbone mutation energy difference.

Open the shared Mutation Stability job, or inspect its live summary.csv and metadata.json.

Explore NuG2 Dynamics

The paper uses two related protocols. In fast-folder Langevin annealing, a base checkpoint explores high-noise conformational space before ProteinEBM-x takes over below a threshold of t=0.1. Direct folding instead begins from a Ramachandran-random unfolded chain and follows low-noise ProteinEBM-x dynamics at t=0.05 [1].

For a minimal end-to-end Neurosnap smoke test we used:

The job completed with one round, one batch, and one scoring level. It reported a native energy of 1531.584 and exported dynamics_trajectory.pt (53.9 KB), confirming that the service executed the upstream dynamics program and preserved coordinate output.

This run validates the workflow, not a folding claim. One step and one sample cannot characterize an energy landscape or folding pathway. The paper used hundreds to thousands of trajectories for its folding analyses: 800 trajectories for NuG2 were required to obtain 20 independent paths that reached within 3 Å RMSD of the native state [1]. A research-scale reproduction should therefore increase the sampling budget substantially and analyze native-contact formation across many successful trajectories.

Open the shared Fast-Folder Dynamics job, or inspect its live summary.csv, metadata.json, and dynamics_trajectory.pt.

Practical Interpretation

Use this order when reviewing a ProteinEBM study:

1. Confirm that every compared PDB represents the same sequence and residue numbering.
2. Compare energies produced within the same job and at the same scoring time.
3. Check whether low energy agrees with an independent structural measure or known biology.
4. Inspect the coordinates for non-native compact states, swapped strands, and broken topology.
5. Treat mutation deltas as hypotheses until backbone relaxation, ensemble effects, or experiments support them.
6. For dynamics, examine multiple trajectories rather than selecting one visually convincing path.

Practical Tips

• Compare energies only within runs that use the same model settings and structurally matched inputs.
• Keep Scoring Time at 0.05 for the primary tutorial unless the experiment explicitly studies noise level.
• Remember that scoring workflows use only the first protein chain from the uploaded PDB.
• Treat mutation results as fixed-backbone screening values.
• Treat ProteinEBM trajectories as learned coarse-grained dynamics, not atomistic molecular dynamics with calibrated physical time.
• Check low-energy structures visually and with an independent structural metric; learned landscapes can contain incorrect minima.

Limitations

ProteinEBM energy is learned, coarse-grained, and not reported in kcal/mol. A low value does not independently establish experimental stability, function, or a physical population. The paper reports both strong ranking results and failure modes, including non-native low-energy strand-swapped structures and cases where sampling fails to reach a native structure even when the native structure would score favorably [1].

Conclusion

ProteinEBM turns one learned object, an energy landscape, into several practical workflows. On Neurosnap, the scoring service queries that landscape at supplied coordinates, while the dynamics service follows its gradients to generate structural trajectories. The NuG2 runs show the right way to use those capabilities: preserve provenance, compare paired values, and keep the result tied to the exact model protocol.

The most important interpretive boundary is simple. ProteinEBM energy is valuable for ranking and hypothesis generation, but it is not an experimental free energy or an atomistic force field. Its strongest results come from combining learned energy with careful sampling and independent validation.

Sources

1. ProteinEBM: Energy Based Models for Protein Structure Generation and Evaluation, Roney et al.; bioRxiv, version 2, 2026.
2. How fast-folding proteins fold, Lindorff-Larsen et al.; Science, 2011.
3. Computer-based redesign of a protein folding pathway, Nauli, Kuhlman, and Baker; Nature Structural Biology, 2001.
4. PDB 1MI0: Crystal structure of the redesigned Protein G variant NuG2, RCSB Protein Data Bank.
5. ProteinEBM source code