Ablating Design Choices for Protein Ensemble Generators Across Structural Regimes

Eoin Quinn; Oliver Bent

q-bio.BMNov 3, 2026

Ablating Design Choices for Protein Ensemble Generators Across Structural Regimes

Authors: Eoin Quinn, Oliver Bent

Organizations: InstaDeep London, UK · InstaDeep Paris, France

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.

Figures & tables

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

\pm 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.

	SAXS					CS
	ordered	disordered	multi-domain	transitional	AUC	ordered	disordered	multi-domain	transitional	AUC
BioEmu	2.06	3.50	3.37	2.54	2.57	1.04	1.48	1.25	1.22	1.24
IDPFold2	2.51	2.43	3.92	3.21	2.68	1.10	1.24	1.14	1.21	1.17
PepTron	2.23	4.60	3.66	2.60	3.01	1.03	1.80	1.41	1.46	1.42
Base	2.42	2.01	3.06	2.90	2.29	1.24	1.40	1.34	1.31	1.31
Kabsch	3.04	3.99	4.67	4.29	3.68	1.27	1.39	1.35	1.33	1.33

Table 1: Absolute scores for the three baselines, the base model, the Kabsch variant and the two training-source variants. Each entry is the geometric-mean normalised RMSE over the proteins in a cell, with ensembles scored as generated and without the default physicality filter (Appendix E ). The base model and variants are first averaged geometrically over training seeds. AUC is the benchmark’s LOWESS-AUC. Lower is better. Cells contain 313/34/42/42 (SAXS) and 237/220/40/150 (CS) proteins.

	ordered	disordered	transitional
BioEmu	0.41 $\pm$ 0.02	0.69 $\pm$ 0.02	0.48 $\pm$ 0.02
IDPFold2	0.34 $\pm$ 0.01	0.56 $\pm$ 0.02	0.42 $\pm$ 0.02
PepTron	0.38 $\pm$ 0.01	0.50 $\pm$ 0.02	0.45 $\pm$ 0.02
Base	0.34 $\pm$ 0.01	0.54 $\pm$ 0.01	0.45 $\pm$ 0.02
Kabsch	0.35 $\pm$ 0.02	0.58 $\pm$ 0.01	0.48 $\pm$ 0.02
Fragment	0.33 $\pm$ 0.02	0.61 $\pm$ 0.01	0.49 $\pm$ 0.02

Table 2: Flory scaling exponent

\nu

from a population fit of

R_{g}\sim N^{\nu}

, with bootstrap SE, for the models of Table 1 . The multi-domain cell is omitted, as its narrow length range makes

\nu

unreliable. The other variants are within error of the base model in these cells.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Architecture (35.7M trainable; frozen ESM2-650M excluded)
hidden size / layers / heads / head dim	384 / 12 / 6 / 64
feed-forward	SwiGLU, $\tfrac{8}{3}$ expansion ( $\approx$ 1024)
per-block norm / gating	RMSNorm on Q, K; sigmoid attention-output gate [ 35 ]
conditioning	AdaLN-zero (time)
positional encoding	RoPE, half of head dims rotated
coordinate scaler	divide by 1.5 nm, centred

Appendix

Table 3: Base model architecture and training hyperparameters.

	SAXS					CS
	ordered	disordered	multi-domain	transitional	AUC	ordered	disordered	multi-domain	transitional	AUC
BioEmu	1.39	1.36	1.48	1.27	1.21	0.93	1.10	1.05	0.90	0.99
IDPFold2	1.39	1.15	1.45	1.34	1.15	0.96	0.89	0.91	0.84	0.91
PepTron	1.57	1.76	1.84	1.48	1.47	0.93	1.33	1.08	1.08	1.11
Base	1.36	1.15	1.32	1.41	1.14	1.12	1.18	1.16	1.04	1.14
Kabsch	1.75	1.70	2.27	2.03	1.69	1.15	1.18	1.18	1.07	1.15

Appendix

Table 4: As Table 1 , after PeptoneBench’s reweighting. Lower is better.

Figure 2: As Figure 1 , after PeptoneBench’s reweighting. Points give the change in geometric-mean RMSE relative to the base model, per cell and for the benchmark’s AUC, with 95% intervals. Lower is better, and the shaded band marks

\pm 5\%

. The AUC interval reflects variation across training seeds only.

Protein Conformational Ensemble Generation

See all 8 papers

CoMoFlow: Controllable Protein Conformational Ensemble Generation via Explicit Collective-Mode Modeling
Nov 6, 2026Yusong Wang, Zheyuan Gu, Ailiang Lin +5Protein Conformational Ensemble Generation Protein Structure Generation
Beijing University of Chinese Medicine · Peking University · New York University +2
Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards
Mar 27, 2026Senura Hansaja Wanasekara, Minh-Duong Nguyen, Xiaochen Liu +2Generative Modeling Protein Design
Biomedical Engineering, The University of Sydney, Darlington, NSW 2006, Australia. · the College of Engineering and Computer Science, VinUniversity, Vietnam. · School of Computer Science and School, The University of Sydney, Darlington, NSW 2006, Australia.
Emyx: Fast and efficient all-atom protein generation
Jun 12, 2026Nicholas J. Williams, Ward Haddadin, Matteo P. Ferla +6Conditional Flow Matching 3D Molecular Generation
Xyme, Oxford, UK

Nov 6, 2026J/K move · Enter open · S save

CoMoFlow: Controllable Protein Conformational Ensemble Generation via Explicit Collective-Mode Modeling

Yusong Wang, Zheyuan Gu, Ailiang Lin, Minghao Shao+4

Beijing University of Chinese Medicine · Peking University · New York University+2

Protein Conformational Ensemble Generation Protein Structure Generation Flow-Based Generative Modeling

Open paper AskPDF arXiv

Protein conformational ensemble generation aims to efficiently characterize the structural variations accessible to a protein beyond a single static structure. Recent flow-based methods efficiently generate conformational ensembles, but they model coordinated structural variation implicitly, which can limit ensemble fidelity and hinder direct control. To address this limitation, we propose CoMoFlow, a controllable framework for generating protein backbone conformational ensembles by explicitly modeling collective variation. CoMoFlow constructs protein-specific collective modes from the equilibrium structure, learns the distribution of their coefficients across the molecular dynamics (MD) ensemble, and uses sampled collective variations to guide SE(3) flow matching through feature conditioning and prior recentering. Experiments on ATLAS and de novo proteins show that CoMoFlow improves reconstruction of MD conformational distributions and coordinated structural variation relative to strong flow-based baselines, while enabling direct control over targeted conformational changes.

Abstract

Protein Conformational Ensemble Generation

Explore similar work

CoMoFlow: Controllable Protein Conformational Ensemble Generation via Explicit Collective-Mode Modeling

Generative Modeling in Protein Design: Neural Representations, Conditional Generation, and Evaluation Standards

Emyx: Fast and efficient all-atom protein generation