DynGeo-Pheno

A disease-associated phosphosite and a pathogenic missense mutation can land on the same residue of the same protein — 393 residues do so across the two benchmarks assembled here — yet the static features normally used to score them point in opposite directions. Phosphosites cluster in AlphaFold's least confident regions, at mean pLDDT around 30–45; pathogenic mutations are enriched in buried, ordered α-helices and β-strands. Static geometry describes two different populations, which is part of why the two are usually modeled separately. DynGeo-Pheno's premise is that what they share is dynamic: both preferentially occupy residues that are mechanically constrained and allosterically coupled to the rest of the fold.

The framework comes from the ComputeSuda group of Zhongjie Liang and Guang Hu at the Suzhou Medical College of Soochow University, with Gennady Verkhivker of Chapman University, and was posted to bioRxiv in August 2026. It combines two residue-level signals over a protein graph: embeddings from the pre-trained ESM-2 protein language model, and eigenvectors of an anisotropic network model (ANM) built on the AlphaFold structure.

DynGeo-Pheno is not itself pretrained; its footing as a foundation-model application is the frozen checkpoint beneath it. ESM-2 is used as released and never fine-tuned, the ANM channel is a deterministic elastic-network calculation, and what is trained is a compact graph head, shipped as two task-specific weight sets over one architecture rather than one checkpoint.

#Key Features

• Frozen language model backbone: ESM-2 (esm2_t33_650M_UR50D) supplies 1,280-dimensional residue embeddings, PCA-projected to 128 dimensions. For missense variants the feature is the wild-type minus mutant embedding difference.
• Dynamics from normal modes: ANMs are built on AlphaFold structures with a 15 Å cutoff; after the six rigid-body modes are dropped, the 30 lowest- and 30 highest-frequency modes give two 90-dimensional per-residue vectors.
• Adaptive Dynamic Fusion: a learnable feature-wise gate, trained end-to-end, weighs collective against localized motion per residue rather than concatenating the two channels.
• Parallel GCN–GAT graph learning: residue graphs with edges between Cα pairs under 10 Å pass through three layers of graph convolution and multi-head attention in parallel branches, then an MLP.
• Task-specific graph construction: phosphosite prediction adds a protein-level global graph to the residue-centered local graph; missense prediction uses only a mutation-centered local graph.
• Biophysical interpretation: flagged residues are mapped onto ligand-binding pockets and interaction interfaces and profiled with RMSF, cross-correlation, perturbation response scanning and mechanical stiffness.

#Technical Details

The phosphorylation benchmark draws on PTMD 2.0, PhosphoSitePlus, PTMint and iPTMnet intersected with PTMAtlas: 73,911 sites across 6,005 human proteins, 3,059 disease-associated. The missense benchmark combines ClinVar with the Rhapsody-2 dataset: 89,471 variants across 10,980 proteins, 30,724 pathogenic. Sequences were redundancy-reduced with CD-HIT at 30% identity, each set split 90/10, and phosphosite data undersampled to a 1:2 positive-to-negative ratio. Training used 10-fold stratified cross-validation with RMSprop, and test predictions come from a majority vote across the ten fold models.

On the held-out splits the ensembles reach AUROC 0.873 and AUPRC 0.772 for phosphosites, and AUROC 0.937 and AUPRC 0.896 for missense variants. Against published phosphosite predictors, AUROC rises from 0.692 (FuncPhos-SEQ) and 0.753 (FuncPhos-STR) to 0.864, and MCC from 0.265 and 0.361 to 0.572. On missense variants it records the highest recall at 0.885, with AUROC 0.914 and AUPRC 0.880 — comparable to AlphaMissense, which the authors treat as a complementary reference, and above Rhapsody-2. Ablation puts the dynamics channel's contribution at 0.7% AUROC for phosphosites and 0.4% for missense variants; ESM-2 carries most of the signal. The final configuration uses 48-residue neighborhoods and no pLDDT filtering, since stricter confidence thresholds consistently hurt.

#Applications

The immediate use is triage: ranking which of a protein's phosphorylation sites are disease-relevant, and which coding variants in a sequencing report are pathogenic. Inputs are a UniProt accession and a residue position, with structures taken from the AlphaFold Protein Structure Database, so it runs on any human protein that database covers. For wet-lab groups the interpretation layer matters as much as the score: a flagged residue arrives with its position relative to binding pockets and interfaces and its ANM profile, pointing at the experiment that would test the call.

#Impact

The independent test set is a random 10% split of the same curated benchmarks, not an external cohort, so the numbers describe held-out generalization within the training distribution, and the dynamics gain is consistent but small. The contribution beyond the two classifiers is the biophysical result they enable: phosphosites concentrate in flexible regulatory regions and pathogenic mutations in ordered elements, yet both converge on residues with enhanced long-range coupling, higher perturbation responsiveness and greater mechanical stiffness. That convergence argues for treating intrinsic dynamics as a determinant of pathogenicity alongside sequence conservation, testable independently of whether these checkpoints stay competitive. Source code and trained fold ensembles are released for both tasks; the preprint has not been peer reviewed.