Ablating Design Choices for Protein Ensemble Generators Across Structural Regimes

Eoin Quinn, Oliver BentarXiv 2026 · arXiv

TL;DR

This paper investigates which design choices actually matter when building generative models for protein ensembles. By systematically ablating a minimal flow-matching model, the authors demonstrate that training data quality is the primary bottleneck, while many standard architectural "tricks" borrowed from structure prediction—like pair representations or mixture-of-experts—offer surprisingly little benefit.

What Problem Does This Paper Address?

While tools like AlphaFold have revolutionized structure prediction for folded proteins, a huge portion of the proteome (20–40% of human proteins) is intrinsically disordered. These proteins don't have a single "correct" shape; instead, they exist as dynamic, heterogeneous ensembles.

Current generative models for these ensembles often inherit their architecture from successful folded-protein predictors. However, ordered and disordered proteins occupy very different structural regimes—one governed by tight packing and coevolution, the other by polymer physics and entropy. This paper asks: Do the design choices that work for folded proteins actually hold up when modeling disorder?

Methods

The researchers built a "minimal" base model: a sequence-conditioned $C\alpha$ flow-matching model (35.7M parameters). It uses frozen ESM2-650M embeddings and a transformer backbone.

From this base, they performed a series of ablations and variants:

  • Sequence Conditioning: Swapping the large ESM2-650M for a tiny ESM2-8M.
  • Pair Representation: Adding a static pair track (similar to AlphaFold's pair representation).
  • Disorder Adaptation: Testing a Mixture-of-Experts (MoE) approach and a dedicated "disorder probe" conditioning signal.
  • Kabsch Alignment: Testing whether aligning the prediction to the target during training (a standard practice) helps or hurts.
  • Training Source: Comparing coarse-grained simulations (CALVADOS) against fragment-assembled ensembles.

Datasets

The models were trained on a mix of three sources to cover all structural regimes:

  1. AFDB: High-confidence AlphaFold structures (ordered).
  2. IDRome-CALVADOS: Coarse-grained simulations of disordered chains.
  3. AF-CALVADOS: Simulations of multi-domain proteins (folded domains with disordered linkers).

For evaluation, they used PeptoneBench, which compares generated ensembles against experimental data:

  • SAXS (Small-angle X-ray scattering): Measures global dimensions and shape.
  • CS (NMR Chemical Shifts): Measures local backbone geometry.

Key Insights and Ideas

  • Data > Architecture: The most significant performance swings came from changing the training data source, not the model architecture. This suggests that for protein ensemble generation, we are currently "data-limited" rather than "architecture-limited."
  • The "Lever Arm" Effect: In structure prediction, Kabsch alignment is used to stabilize gradients. However, the authors found it degrades performance for disordered proteins. Because disordered chains have long, flexible "tails," these tails act as lever arms that dominate the alignment, causing the model to misalign the rest of the chain and injecting noise into the training process.
  • LLMs for Physics vs. Evolution: Large protein language models (like ESM2-650M) are essential for ordered proteins because they capture coevolutionary spatial contacts. However, for disordered proteins, a tiny 8M-parameter model performed just as well. Disordered regions rely more on bulk physicochemical properties (like charge and hydropathy) which are easier for smaller models to grasp.
  • Local vs. Global Trade-offs: There is a clear trade-off in training data. Coarse-grained simulations (CALVADOS) are great for global shape (SAXS) but poor at local geometry. Fragment assembly is the opposite. An equal mixture of both sources provided the best overall balance.
Figure 1: Change in geometric-mean RMSE of each variant relative to the base model, per cell and for the benchmark’s AUC, with 95% intervals. Ensembles are scored as generated, and lower is better. The shaded band marks ±5%. We regard a change as negligible when its interval lies within the band, and unresolved when its interval contains zero but extends beyond it. The AUC is a single score per run, so its interval reflects variation across training seeds only.

Key Results

  • Kabsch Alignment: Increased SAXS error by +26% on ordered proteins and a staggering +99% on disordered ones.
  • Model Scaling: Reducing the language model size (ESM2-8M) degraded performance on ordered proteins (SAXS +24%, CS +12%) but had no detectable effect on disordered ones.
  • Architectural "Bells and Whistles": Adding a pair representation, MoE, or a disorder probe resulted in negligible changes (within ±5% of the base model).
  • Training Source: Replacing simulation data with fragment-assembled data improved chemical shift agreement by 11–16% but worsened SAXS agreement on disordered proteins by 62%.

Limitations and Future Work

The study has a few notable constraints:

  • Resolution: The model operates at the $C\alpha$ level, requiring a separate "back-mapping" step to recover all-atom detail for evaluation.
  • Data Quality: All training data comes from approximations (simulations or fragment assembly), not direct experimental observations.
  • Evaluation: The benchmarks (SAXS and CS) are ensemble-averaged observables, which only constrain the ensemble indirectly.

The authors conclude that future progress in the field won't come from bigger or more complex transformers, but from developing training sources that are simultaneously accurate at both local and global scales.

Summary updated

Abstract

Generative models of protein conformational ensembles inherit much of their design from structure prediction for folded proteins. Ordered and disordered proteins, however, occupy distinct structural regimes, and we investigate how design choices hold up across them. Using a minimal C$α$ flow-matching model, we ablate choices of architecture, sequence conditioning, rigid-body alignment and training source, and evaluate each variant against experimental SAXS and NMR chemical-shift data partitioned into ordered, disordered, multi-domain and transitional proteins. Mixture-of-experts routing, a static pair representation and a disorder probe have no detectable effect, a smaller protein language model performs worse on ordered proteins but not on disordered ones, and alignment in the loss degrades agreement with SAXS. Replacing coarse-grained simulations of disordered proteins with fragment-assembled ensembles improves agreement with chemical shifts at a cost to SAXS, while an equal mixture of the two keeps most of the gain at a smaller cost. This points to the training data, rather than the architecture, as the main constraint.
BibTeX
@misc{quinn2026ablating,
  author = {Eoin Quinn and Oliver Bent},
  title = {{Ablating Design Choices for Protein Ensemble Generators Across Structural Regimes}},
  year = {2026},
  eprint = {2611.05055},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05055},
}
Figure