Beyond Static Structures: Linking Protein Dynamics to Function via DynaReason-Pro

Zhaokang Liang; Mingze Yuan; Quanzheng Li; Wengong Jin

cs.AIOct 24, 2026

Beyond Static Structures: Linking Protein Dynamics to Function via DynaReason-Pro

Authors: Zhaokang Liang, Mingze Yuan, Quanzheng Li, Wengong Jin

Organizations: Northeastern University · Harvard University · Massachusetts General Hospital, Harvard Medical School

Abstract

Intrinsically disordered proteins (IDPs) carry out their functions through heterogeneous conformational ensembles that reflect their dynamics, yet most protein function prediction methods rely on sequence or a single static structure. To this end, we introduce DynaReason-Pro, the first framework to jointly predict GO terms from conformational ensembles and generate natural-language functional rationales. We evaluate DynaReason-Pro on a curated IDP benchmark from the Protein Ensemble Database (PED), with sequence- and structure-cluster-disjoint splits to prevent data leakage. Variability-stratified evaluation shows the clearest gains in the high-conformational-variability regime.

Figures & tables

Panel of Figure 1 — Figure 1: a , Architecture of DynaReason-Pro. The ontology predictor integrates the protein sequence and residue-aligned conformational ensemble to generate GO terms; the LLM-based reasoning module then combines these ontology anchors with the protein sequence, organism information, and InterPro annotations to generate functional descriptions and natural-language rationales. b , Comparison with DeepGO, DeepFRI-GCN, and BioReason-Pro across $F_{\max}$ , $F_{\mathrm{wmax}}$ , AUPR, IA-MAP, NDCG@5, and IA-R@5. All spokes report raw scores on a common scale, with higher values indicating better performance.

Metric	Aspect	Statistics			Sequence
Metric	Aspect	GO prior	length 5NN	AA comp. 5NN	Sequence-only predictor	DeepGO	DeepFRI-CNN	Structure DeepFRI-GCN	LLM-based BioReason-Pro	Ensemble-aware DynaReason-Pro
$F_{\max}\uparrow$	Overall	0.4066	0.3392	0.4247	0.4226	0.4261	0.3983	0.4390	0.4251	0.6045
	MF	0.2944	0.2301	0.3050	0.3306	0.4470	0.4104	0.4081	0.3223	0.4814
	BP	0.2104	0.0852	0.1885	0.3020	0.3071	0.2948	0.2214	0.3064	0.3163
	CC	0.4827	0.4300	0.5120	0.4634	0.4537	0.4663	0.5424	0.4579	0.7004
	High-IA	0.3617	0.1991	0.3194	0.2337	0.3481	0.3318	0.3430	0.2356	0.4897

Table 1: CAFA-style metrics on the PED test set. Arrows indicate the preferred metric direction. Bold indicates the best method for each metric.

Metric	GO prior	length 5NN	AA comp. 5NN	Sequence-only predictor	DeepGO	DeepFRI-CNN	Structure DeepFRI-GCN	LLM-based BioReason-Pro	Ensemble-aware DynaReason-Pro
Metric	Statistics			Sequence
IA-MAP $\uparrow$	0.3494	0.2063	0.3412	0.2068	0.3294	0.3780	0.3951	0.2110	0.4309
N@1 $\uparrow$	0.0845	0.0739	0.0796	0.0729	0.1081	0.1395	0.1553	0.1002	0.2940
N@3 $\uparrow$	0.1029	0.1349	0.1777	0.1447	0.1260	0.1972	0.1988	0.1510	0.3557
N@5 $\uparrow$	0.1096	0.1618	0.2026	0.1515	0.1624	0.2633	0.2815	0.1646	0.3704
N@10 $\uparrow$	0.2072	0.1863	0.2388	0.1626	0.2055	0.3068	0.3208	0.1751	0.3779

Table 2: Information-aware and top-

k

retrieval metrics on the PED test set for the same methods in Table 1 . All metrics are reported with

\uparrow

, indicating that higher is better. Bold indicates the best method for each metric. N@

k

= NDCG@

k

, J@

k

= Jaccard@

k

, and R@

k

= IA-R@

k

Figure 2: Overall ablation across architectural and retrieval variants.

F_{\max}

and

F_{\mathrm{wmax}}

compare sequence-only prediction, AFDB single-structure conditioning, naive averaging over 32 PED conformers, ESM2 sequence retrieval, ESM3 AFDB-structure retrieval, and DynaReason-Pro.

Figure 3: DynaReason-Pro has its largest relative advantage in the highest-variability quartile. a ,

F_{\max}

; b ,

F_{\mathrm{wmax}}

. Both metrics are shown for DynaReason-Pro and four fixed-split retrieval variants across TM-score quartiles (Q1–Q4;

n=30,30,29,29

). Dashed error bars show 95% bootstrap confidence intervals from 10,000 within-quartile resamples (2.5th–97.5th percentiles); markers are horizontally offset for visibility. DynaReason-Pro leads all four variants in Q4.

Figure 4: Ensemble conditioning removes incorrect predictions and recovers specific GO functions. Across three PED proteins, DynaReason-Pro replaces broad sequence-only outputs with exact reference terms: (a) cytosol and keratin filament, (b) zinc ion binding and ubiquitin protein ligase activity, and (c) two RNA polymerase II-specific DNA-binding functions.

Figure 5: DynaReason-Pro resolves domain architecture into candidate protein-specific mechanisms. Panel (a) shows the complete trace for PED02763_e001, panel (b) for PED01773_e001, and panel (c) for PED03119_e001. Within each panel, the full trace is reflowed into three semantic blocks: domain evidence; function and cellular context; and mechanisms and candidate interaction partners.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Quantity	Count
PED entry/ensemble records	5,902
Distinct UniProt accessions	5,760
Distinct model-input sequences	5,744
Training / validation / test records	5,654 / 130 / 118
Accession groups spanning splits	0
Exact model-input-sequence groups spanning splits	0

Appendix

Table 3: Retained benchmark records and distinct identities. Counts of accessions and sequences refer to distinct values.

Condition or aggregation	$F_{\max}$	$F_{\mathrm{wmax}}$
Full set-attention ensemble	0.62267	0.52527
Mismatched ensemble	0.42431	0.25065
Repeated single conformer (range)	0.50902–0.55238	0.38598–0.44531
DeepSets	0.60275	0.50732
Attention pooling	0.57603	0.46753

Appendix

Table 4: Controlled evidence for protein-specific ensemble conditioning. The full, mismatched, and repeat-single inputs are evaluated with the same ensemble reference model. DeepSets and attention pooling are trained ensemble alternatives. Repeat-single ranges summarize five choices of one conformer duplicated 32 times.

Conformer count	$F_{\max}$	$F_{\mathrm{wmax}}$
$K=8$	0.571077	0.454456
$K=16$	0.585052	0.458119
$K=32$ reference	0.622673	0.525268

Appendix

Table 5: Conformer-count sensitivity for trained

K=8

K=16

, and

K=32

models.

Subset size	$F_{\max}$	$F_{\mathrm{wmax}}$
$K=8$	$0.611430\pm 0.011644$	$0.509754\pm 0.019627$
$K=16$	$0.623143\pm 0.004861$	$0.527054\pm 0.011763$

Appendix

Table 6: Random-subset robustness. Values are means

\pm

standard deviations over five subsets per conformer count.

Selection	$F_{\max}$	$F_{\mathrm{wmax}}$
Diversity-based $K=8$	0.62200	0.52444
Diversity-based $K=16$	0.61336	0.51258
Full $K=32$ reference	0.62267	0.52527

Appendix

Table 7: Diversity-based subset selection compared with the full ensemble.

Input	$F_{\max}$	$F_{\mathrm{wmax}}$
Repeated single conformer 1	0.50902	0.39536
Repeated single conformer 2	0.55108	0.44531
Repeated single conformer 3	0.51110	0.38598
Repeated single conformer 4	0.53560	0.42086
Repeated single conformer 5	0.55238	0.44521
Full ensemble reference	0.62267	0.52527

Appendix

Table 8: Repeat-single negative control. Each row supplies repeated copies of one structural state; the full reference uses 32 distinct conformers.

Contrast	Estimate	95% CI
Q4 minus rest, protein bootstrap	0.291746	[0.095152, 0.487881]
Q4 minus rest, related-group bootstrap	0.291746	[0.094825, 0.492712]
90th minus 50th embedding-variability percentile	0.134365	[0.006773, 0.252423]
Matched high-variability gain difference (13 pairs)	0.501261	—

Appendix

Table 9: Selected covariate-adjusted variability contrasts relative to ESM2 sequence 5NN.

Measure	Paired change	95% CI
Reference-supported fraction	+1.8406 pp	[1.0029, 2.6816]
Reference-unverifiable fraction	$-2.0520$ pp	[ $-2.9159$ , $-1.1842$ ]
Supported units per response	+0.423729	[0.220339, 0.627119]

Appendix

Table 10: Ensemble-conditioned GO anchors increase reference-supported content. Differences are ensemble minus sequence. Fractions are response-level means; pp denotes percentage points. Intervals are paired protein-bootstrap 95% CIs.

Control	Support-rate ratio	95% CI
Checkable-unit count	1.970460	[1.433944, 2.840275]
Response word count	1.947284	[1.405645, 2.819655]

Appendix

Table 11: Reference-supported content after controlling for response size. Ratios compare ensemble-GO with sequence-GO responses.

Aspect	$n$	Truth labels	Returned	Recall	IA recall	AP
BP	31	17.48	4.74	0.2190	0.1910	0.1970
CC	107	6.84	4.00	0.6790	0.5730	0.5820
MF	28	8.61	5.68	0.4160	0.3190	0.3060

Appendix

Table 12: Aspect-specific prediction coverage. Truth labels and returned predictions are mean counts per eligible example; recall, IA recall, and AP summarize prediction coverage and ranking.

Model / summary	Training seed	$F_{\max}$	$F_{\mathrm{wmax}}$
Main benchmark model	20260425	0.6045	0.5172
Independent initialization	20261001	0.6179	0.5422
Independent initialization	20261002	0.6172	0.5274
Three runs, mean $\pm$ SD	—	$0.613200\pm 0.007543$	$0.528933\pm 0.012570$

Appendix

Table 13: Training variance and reproduction of the original benchmark result. The three-run row summarizes the matched-budget evaluation.

Protein Function Prediction

See all 10 papers

ENSEMBITS: an alphabet of protein conformational ensembles
May 13, 2026Kaiwen Shi, Carlos OliverProtein Language Models Protein Representation Learning
Department of Computer Science, Vanderbilt University · Center for AI in Protein Dynamics, Vanderbilt University · Department of Molecular Physiology and Biophysics, Vanderbilt University
MT-ProtBERT: Multi-task Learning ProtBERT for Intrinsically Disordered Proteins Classification with Scarce Data
Sep 21, 2026Jian Sun, Kingshuk Ghosh, Lilianna Houston +1Protein Language Models Multi-Task Learning
Department of Computer Science, Ritchie School of Engineering and Computer Science, University of Denver, 2155 E Wesley Ave, Denver, CO 80210 · Department of Surgery, David Geffen School of Medicine at the University of California, Los Angeles, 10833 Le Conte Ave, Los Angeles, CA 90095 · Department of Physics and Astronomy, at the University of Denver, CO 80210 +1
DynaPPI: A Large-scale Dynamic Protein Dataset for AI-driven Advances in Protein Interactomics
Aug 11, 2026Jiabao Wei, Zilong Geng, Yuze Wang +5Protein Structure Prediction Protein-Protein Interaction Prediction
1BIT · 3SJTU · 2HUST +2

May 13, 2026J/K move · Enter open · S save

ENSEMBITS: an alphabet of protein conformational ensembles

Kaiwen Shi, Carlos Oliver

Department of Computer Science, Vanderbilt University · Center for AI in Protein Dynamics, Vanderbilt University · Department of Molecular Physiology and Biophysics, Vanderbilt University

Protein Language Models Protein Representation Learning Molecular Representation Learning

Open paper AskPDF arXiv

Protein structure tokenizers (PSTs) are workhorses in protein language modeling, function prediction, and evolutionary analysis. However, existing PSTs only capture local geometry of static structures, and miss the correlated motions and alternative conformational states revealed by protein ensembles. Here we introduce Ensembits, the first tokenizer of protein conformational ensembles. Ensembits address challenges inherent to tokenizing dynamics: deriving informative geometric descriptors across conformations, permutation-invariance encoding of variable-size ensembles, and conquering sparsity in dynamics data. Trained with a Residual VQ-VAE using a frame distillation objective on a large molecular dynamics corpus, Ensembits outperforms all related methods on RMSF prediction, and is the strongest standalone structural tokenizer on an token-conditioned ANOVA test on per-residue motion amplitude. Ensembits further matches or exceeds static tokenizers on EC, GO, binding site/affinity prediction, and zero-shot mutation-effect prediction despite using far less pretraining data. Notably, the distillation objective enables Ensembits to predict dynamics token from one single predicted structure, which alleviates dynamics data sparsity. As the field moves from static structure prediction toward ensemble generation, Ensembits offer the discrete vocabulary needed to bring dynamics into protein language modeling and design.

Abstract

Protein Function Prediction

Explore similar work

ENSEMBITS: an alphabet of protein conformational ensembles

MT-ProtBERT: Multi-task Learning ProtBERT for Intrinsically Disordered Proteins Classification with Scarce Data

DynaPPI: A Large-scale Dynamic Protein Dataset for AI-driven Advances in Protein Interactomics