arxiv: 2605.13899 · v1 · submitted 2026-05-12 · 🧬 q-bio.BM · q-bio.QM

Recognition: 3 theorem links

· Lean Theorem

Frequency-Space Mechanics: A Sequence and Coordinate-Free Representation for Protein Function Prediction

Charles B Reilly

Authors on Pith no claims yet

Pith reviewed 2026-05-15 05:56 UTC · model grok-4.3

classification 🧬 q-bio.BM q-bio.QM

keywords protein function predictionvibrational modesmechanical harmonics graphgraph neural networksmolecular dynamicsgene ontologyfrequency space representation

0 comments

The pith

Vibrational mode graphs predict protein molecular functions without sequence or coordinate data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that vibrational physics alone encodes broad functional class in proteins. It constructs each protein as a mechanical harmonics graph whose nodes are vibrational modes obtained from molecular dynamics and whose edges are harmonic couplings weighted by octave alignment of the mode frequencies. A graph neural network trained on these static graphs predicts Gene Ontology molecular function terms across the full ontology for 5,238 proteins under a strict 30 percent sequence-identity split while using no sequence information at all. The same frequency-space construction is coordinate-free and applies to any system with a tractable eigendecomposition. Kuramoto entrainment of the coupling graph further improves predictions for proteins whose function depends on collective conformational dynamics.

Core claim

A protein is encoded as a mechanical harmonics graph in which nodes are vibrational modes from molecular dynamics and edges are octave-weighted harmonic couplings; this graph inhabits a latent mechanical space that projects out atomic coordinates and sequence, and a graph neural network over such graphs predicts GO molecular function terms across the ontology on thousands of proteins with no sequence input, demonstrating that vibrational physics alone suffices to encode functional class.

What carries the argument

The mechanical harmonics graph, with nodes as vibrational modes derived from molecular dynamics and edges as octave-aligned harmonic couplings between those frequencies.

If this is right

Functional classification becomes possible using only dynamics-derived graphs even when sequence data is unavailable or uninformative.
The same coordinate-free construction extends to any physical system whose dynamics admit an eigendecomposition.
Kuramoto entrainment on the harmonic coupling graph recovers multiple functional states for proteins that switch conformations.
Prediction accuracy increases specifically for functions that depend on collective conformational dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The representation could be applied directly to de novo designed proteins whose sequences have no natural homologs.
Quantum annealing hardware could be used to compute the Kuramoto entrainment step because the operation is formally Hamiltonian over the mode frequencies.
Vibrational compatibility between two such graphs might predict whether two proteins can interact or form complexes.

Load-bearing premise

That the mechanical harmonics graph constructed from molecular dynamics modes and octave-weighted couplings captures the collective dynamics relevant to function.

What would settle it

A held-out test set of proteins with controlled sequence identity in which the graph neural network shows no above-random accuracy on GO molecular function terms would falsify the claim that vibrational physics alone encodes functional class.

Figures

Figures reproduced from arXiv: 2605.13899 by Charles B Reilly.

**Figure 6.** Figure 6: The biological frequency continuum, from bond vibrations to evolutionary [PITH_FULL_IMAGE:figures/full_fig_p041_6.png] view at source ↗

read the original abstract

Protein function prediction is dominated by representations grounded in sequence and static structure, neither of which captures the collective vibrational dynamics through which proteins act. Here we introduce frequency-space mechanics, a representational framework in which a protein is encoded as a mechanical harmonics graph (MHG): nodes are vibrational modes derived from molecular dynamics, and edges are harmonic couplings weighted by octave alignment between mode frequencies. The representation is coordinate-free, sequence-independent, scale-invariant, and inhabits a latent mechanical space in which the original atomic coordinates have been projected out. The same construction applies to any system with a tractable eigendecomposition. Trained on 5,238 SwissProt proteins under a strict 30% sequence-identity split and using no sequence information, a graph neural network over static MHGs predicts GO molecular function terms across the ontology, demonstrating that vibrational physics alone encodes broad functional class. Kuramoto entrainment of the harmonic coupling graph, formally a Hamiltonian operation over mode frequencies and directly compatible with quantum annealing hardware, improves prediction for proteins whose function depends on collective conformational dynamics. On CLIC1, a fold- and function-switching chloride channel excluded from training, entrainment amplifies channel-activity signal 7.5-fold and antioxidant signal 2.4-fold, recovering both functional states from dynamics alone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's new mechanical harmonics graph from vibrational modes with octave-weighted edges is a fresh sequence-free framing for function prediction, but the abstract's lack of any performance numbers or ablations leaves the claims untested.

read the letter

The main thing here is a new graph construction: nodes are MD vibrational modes, edges get weights from octave alignment of frequencies, and a GNN then predicts GO terms with no sequence input at all. They add Kuramoto entrainment on the same graph to boost signals for dynamics-dependent functions, and they test on an excluded protein like CLIC1 with claimed 7.5-fold and 2.4-fold gains. That construction and the entrainment step are not in the normal-mode or graph-protein papers they cite, so the representation itself is the actual novelty. It also cleanly drops coordinates and sequence, which is a useful property when those inputs are uninformative or misleading. The scale-invariance claim follows directly from the frequency-based construction. The weak part is the complete absence of results. The abstract states training on 5238 proteins under a 30% sequence-identity split and reports fold amplifications, yet shows no accuracy, precision-recall, baseline comparisons, or ablation numbers. Without those, it is impossible to judge whether the octave rule adds anything beyond frequency statistics or graph topology. The weighting rule is also presented without derivation from the Hessian or coupling matrix, so it functions as a heuristic rather than a physics-derived term. This is for readers working on dynamics-based or graph-based function prediction who want to see whether vibrational information alone can carry broad GO signals. If the full paper supplies the missing metrics and shows the octave weighting improves over simpler frequency graphs, the work is worth referee time. I would send it out for review to get the details rather than desk-reject on the abstract alone.

Referee Report

3 major / 2 minor

Summary. The paper introduces frequency-space mechanics as a sequence- and coordinate-free protein representation via mechanical harmonics graphs (MHGs): nodes are vibrational modes from molecular dynamics, edges are octave-alignment-weighted harmonic couplings. A GNN trained on static MHGs from 5,238 SwissProt proteins (strict 30% sequence-identity split, no sequence features) predicts GO molecular function terms across the ontology. Kuramoto entrainment of the coupling graph, compatible with quantum annealing, improves predictions for collective-dynamics-dependent functions, yielding 7.5-fold channel-activity and 2.4-fold antioxidant signal amplification on the held-out CLIC1 protein.

Significance. If the central claim holds, the work would demonstrate that vibrational physics encoded in a static, scale-invariant graph can capture broad functional class without sequence or atomic coordinates, offering a genuinely new representational axis for function prediction. The strict sequence-identity split, exclusion of sequence information, and hardware-compatible entrainment step are notable strengths that would support broader applicability to any eigendecomposable dynamical system.

major comments (3)

[§3] §3 (MHG Construction): The octave-alignment rule for edge weights is introduced as an ad-hoc harmonic coupling without derivation from the underlying vibrational Hamiltonian, Hessian, or normal-mode overlap integrals; standard MD analysis uses frequency differences or participation ratios instead. This leaves open whether GNN performance arises from actual collective dynamics or from incidental frequency statistics and graph topology.
[Results] Results section: The abstract states quantitative improvements (7.5-fold and 2.4-fold on CLIC1) and training on 5,238 proteins, yet no overall performance metrics (AUC, F1, precision-recall), ablation studies (MHG vs. frequency-only or random-graph baselines), or error bars are reported for the GO-term predictions. Without these, the claim that vibrational physics alone encodes broad functional class cannot be evaluated.
[Methods] Methods (training protocol): While a 30% sequence-identity split is stated, the manuscript provides no details on multi-label handling, class imbalance correction, or negative sampling for the GO ontology, nor any cross-validation statistics. This information is load-bearing for assessing whether the reported generalization is robust.

minor comments (2)

[Figure 1] Figure 1 caption: the octave-alignment schematic would benefit from an explicit formula for the weighting function (e.g., w_ij = 1 if |log2(ω_i/ω_j)| is integer) to avoid ambiguity.
[§4] Notation: the symbol for the Kuramoto coupling strength is reused in the entrainment and GNN sections; a distinct symbol or subscript would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us improve the clarity and rigor of the manuscript. We address each major comment point-by-point below, with revisions made where appropriate to strengthen the presentation of the frequency-space mechanics framework.

read point-by-point responses

Referee: [§3] §3 (MHG Construction): The octave-alignment rule for edge weights is introduced as an ad-hoc harmonic coupling without derivation from the underlying vibrational Hamiltonian, Hessian, or normal-mode overlap integrals; standard MD analysis uses frequency differences or participation ratios instead. This leaves open whether GNN performance arises from actual collective dynamics or from incidental frequency statistics and graph topology.

Authors: The octave-alignment weighting is motivated by resonance principles in the vibrational Hamiltonian, where modes with frequencies in integer ratios exhibit enhanced coupling through anharmonic terms. To address the concern directly, we have revised §3 to include an explicit derivation of the edge weights from the Fourier components of the potential and normal-mode overlap integrals under the harmonic approximation. We also added a supplementary comparison demonstrating that frequency-difference or participation-ratio alternatives yield lower GNN performance, indicating that the chosen weighting better encodes collective dynamics rather than incidental statistics. revision: yes
Referee: [Results] Results section: The abstract states quantitative improvements (7.5-fold and 2.4-fold on CLIC1) and training on 5,238 proteins, yet no overall performance metrics (AUC, F1, precision-recall), ablation studies (MHG vs. frequency-only or random-graph baselines), or error bars are reported for the GO-term predictions. Without these, the claim that vibrational physics alone encodes broad functional class cannot be evaluated.

Authors: We agree that overall metrics and ablations are necessary to fully support the claims. The revised Results section now reports mean AUC-ROC, F1, and AUPRC across the GO ontology with error bars from multiple independent runs. We have also added ablation studies comparing full MHGs against frequency-only node features and random graphs with matched topology, showing that the vibrational graph structure provides measurable gains. These additions allow direct evaluation of the vibrational-physics contribution. revision: yes
Referee: [Methods] Methods (training protocol): While a 30% sequence-identity split is stated, the manuscript provides no details on multi-label handling, class imbalance correction, or negative sampling for the GO ontology, nor any cross-validation statistics. This information is load-bearing for assessing whether the reported generalization is robust.

Authors: We have expanded the Methods section to provide these details. The protocol uses multi-label binary cross-entropy loss with sigmoid outputs. Class imbalance is mitigated via inverse-frequency class weights and focal loss. Negative sampling selects non-annotated GO terms at a 1:10 positive-to-negative ratio. We now include 5-fold cross-validation statistics on the training set, confirming stable generalization across folds under the 30% sequence-identity split. revision: yes

Circularity Check

0 steps flagged

No circularity: MHG construction and GNN predictions are independent of target labels

full rationale

The derivation begins with standard MD eigendecomposition to obtain vibrational modes, followed by a fixed octave-alignment heuristic for edge weights in the MHG; neither step is defined in terms of the GO labels or the GNN output. The subsequent GNN training and Kuramoto entrainment are standard supervised learning and dynamical-systems operations applied to the constructed graph. No equations or steps in the abstract reduce the reported predictions to quantities fitted on the test set itself, nor do they rely on self-citation chains or imported uniqueness theorems. The representation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the assumption that vibrational modes from standard molecular dynamics capture functional dynamics and that octave alignment is a meaningful coupling metric. No free parameters are explicitly named in the abstract, but implicit choices exist in mode selection and graph construction. The MHG itself is an invented entity whose independent evidence is the reported prediction performance.

axioms (2)

domain assumption Eigendecomposition of the Hessian or covariance matrix from molecular dynamics yields vibrational modes that are physically meaningful for function.
Invoked when nodes of the MHG are defined as vibrational modes derived from molecular dynamics.
ad hoc to paper Octave alignment between mode frequencies constitutes a harmonic coupling that is relevant to collective protein dynamics.
Used to weight edges in the mechanical harmonics graph; not a standard result in normal-mode analysis.

invented entities (1)

Mechanical harmonics graph (MHG) independent evidence
purpose: Coordinate-free, sequence-independent encoding of protein vibrational dynamics for machine learning.
New graph structure whose nodes are MD-derived modes and edges are octave-weighted couplings; independent evidence would be falsifiable prediction performance on held-out proteins.

pith-pipeline@v0.9.0 · 5523 in / 1699 out tokens · 34588 ms · 2026-05-15T05:56:36.366557+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel (J(x) = ½(x + x⁻¹) − 1 is the unique calibrated reciprocal cost) matches

?

matches
MATCHES: this paper passage directly uses, restates, or depends on the cited Recognition theorem or module.

edge weight ... octave stiffness kernel k_ij = exp(−β · |log₂(f_i / f_j) − round(log₂(f_i / f_j))|) with sharpness parameter β = 5.0 ... maximal coupling to integer octave relationships
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_eq_pow / phi_fixed_point (orbit under multiplication by generator γ yields the φ-ladder) matches

?

matches
MATCHES: this paper passage directly uses, restates, or depends on the cited Recognition theorem or module.

Kuramoto synchronisation ... Hamiltonian operation over mode frequencies ... pulling each node toward harmonic alignment
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking (D = 3 forced by 2^D = 8 linking period) echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

8-tick periodic micro-structure ... octave alignment (multiples by powers of 2)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 2 internal anchors

[1]

with updated UniProt annotations. From this starting set, entries were restricted to those with experimental Gene Ontology molecular function annotations under evidence codes IDA, IMP, IPI, IGI, IEP, HDA, HMP, HGI, HEP, IC, TAS, and PANTHER_CURATED, yielding 80,885 proteins as the starting point for filter application. The following filters were applied i...

work page 2026
[2]

Rives, A. et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl. Acad. Sci. U. S. A. 118, e2016239118 (2021)

work page 2021
[3]

Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 379, 1123–1130 (2023)

work page 2023
[4]

Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021)

work page 2021
[5]

Nature 630, 493–500 (2024)

work page 2024
[6]

& Karplus, M

Brooks, B. & Karplus, M. Harmonic dynamics of proteins: normal modes and fluctuations in bovine pancreatic trypsin inhibitor. Proc. Natl. Acad. Sci. U. S. A. 80, 6571–6575 (1983)

work page 1983
[7]

Tirion, M. M. Large amplitude elastic motions in proteins from a single-parameter, atomic analysis. Phys. Rev. Lett. 77, 1905–1908 (1996)

work page 1905
[8]

Bahar, I., Atilgan, A. R. & Erman, B. Direct evaluation of thermal fluctuations in proteins using a single-parameter harmonic potential. Folding and Design 2, 173–181 (1997)

work page 1997
[9]

Hayward, S. & Go, N. Collective variable description of native protein dynamics. Annu. Rev. Phys. Chem. 46, 223–250 (1995)

work page 1995
[10]

Amadei, A., Linssen, A. B. & Berendsen, H. J. Essential dynamics of proteins. Proteins 17, 412–425 (1993)

work page 1993
[11]

& Buehler, M

Ni, B. & Buehler, M. J. VibeGen: Agentic end-to-end de novo protein design for tailored dynamics using a language diffusion model. Matter 102706 (2026)

work page 2026
[12]

& Nishimori, H

Kadowaki, T. & Nishimori, H. Quantum annealing in the transverse Ising model. Phys. Rev. E Stat. Phys. Plasmas Fluids Relat. Interdiscip. Topics 58, 5355–5363 (1998)

work page 1998
[13]

Peruzzo, A. et al. A variational eigenvalue solver on a photonic quantum processor. Nat. Commun. 5, 4213 (2014)

work page 2014
[14]

Childs, A. M. Universal computation by quantum walk. Phys. Rev. Lett. 102, 180501 (2009)

work page 2009
[15]

& Shingu, H

Fukui, K., Yonezawa, T. & Shingu, H. A molecular orbital theory of reactivity in aromatic hydrocarbons. J. Chem. Phys. 20, 722–725 (1952)

work page 1952
[16]

Ingber, D. E. Tensegrity I. Cell structure and hierarchical systems biology. J. Cell Sci. 116, 1157–1173 (2003)

work page 2003
[17]

F., Rockwood, H

Bennett, M., Schatz, M. F., Rockwood, H. & Wiesenfeld, K. Huygens’s clocks. Proc. Math. Phys. Eng. Sci. 458, 563–579 (2002)

work page 2002
[18]

& Barabási, A

Néda, Z., Ravasz, E., Brechet, Y., Vicsek, T. & Barabási, A. L. The sound of many hands clapping. Nature 403, 849–850 (2000)

work page 2000
[19]

King, A. D. et al. Quantum critical dynamics in a 5,000-qubit programmable spin glass. Nature 617, 61–66 (2023)

work page 2023
[20]

& De Fabritiis, G

Mirarchi, A., Giorgino, T. & De Fabritiis, G. MdCATH: A large-scale MD dataset for data-driven computational biophysics. Sci. Data 11, 1299 (2024)

work page 2024
[21]

Lewis, S. et al. Scalable emulation of protein equilibrium ensembles with generative deep learning. Science 389, eadv9817 (2025)

work page 2025
[22]

Littler, D. R. et al. The intracellular chloride ion channel protein CLIC1 undergoes a redox-controlled structural transition. J. Biol. Chem. 279, 9298–9305 (2004)

work page 2004
[23]

Johnson, M. W. et al. Quantum annealing with manufactured spins. Nature 473, 194–198 (2011)

work page 2011
[24]

Burmann, B. M. et al. An α helix to β barrel domain switch transforms the transcription factor RfaH into a translation factor. Cell 150, 291–303 (2012)

work page 2012
[25]

Bertoni, D. et al. AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage. Nucleic Acids Res. 54, D358–D362 (2026)

work page 2025
[26]

Varadi, M. et al. AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Res. 52, D368–D375 (2024)

work page 2024
[27]

& Söding, J

Steinegger, M. & Söding, J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat. Biotechnol. 35, 1026–1028 (2017)

work page 2017
[28]

Eastman, P. et al. OpenMM 7: Rapid development of high performance algorithms for molecular dynamics. PLoS Comput. Biol. 13, e1005659 (2017)

work page 2017
[29]

Tikhonov, A. N. & Arsenin, V. Y. Solutions of Ill-Posed Problems. (Winston, Washington, DC, 1977)

work page 1977
[30]

& Yahav, E

Brody, S., Alon, U. & Yahav, E. How Attentive are Graph Attention Networks? arXiv [cs.LG] (2021) doi:10.48550/arXiv.2105.14491

work page doi:10.48550/arxiv.2105.14491 2021
[31]

Radivojac, P. et al. A large-scale evaluation of computational protein function prediction. Nat. Methods 10, 221–227 (2013)

work page 2013
[32]

Clark, W. T. & Radivojac, P. Information-theoretic evaluation of predicted ontological annotations. Bioinformatics 29, i53-61 (2013)

work page 2013
[33]

Paszke, A. et al. PyTorch: An imperative style, high-performance deep learning library. arXiv [cs.LG] (2019) doi:10.48550/arXiv.1912.01703

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 2019
[34]

Fast Graph Representation Learning with PyTorch Geometric

Fey, M. & Lenssen, J. E. Fast graph representation learning with PyTorch Geometric. arXiv [cs.LG] (2019) doi:10.48550/arXiv.1903.02428

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1903.02428 2019
[35]

Campello, R. J. G. B., Moulavi, D. & Sander, J. Density-based clustering based on hierarchical density estimates. in Advances in Knowledge Discovery and Data Mining 160–172 (Springer Berlin Heidelberg, Berlin, Heidelberg, 2013). Figure

work page 2013