GeoDSSOP: Geometry-aware prediction of protein backbone order parameters from sequence and structure
- 1State Key Laboratory of Magnetic Resonance Spectroscopy and Imaging
- 2National Center for Magnetic Resonance in Wuhan, Wuhan Institute of Physics and Mathematics, Innovation Academy for Precision Measurement Science and Technology, Chinese Academy of Sciences
- 3University of Chinese Academy of Sciences
Abstract
Backbone amide order parameters (S2) describe protein motions on the picosecond-to-nanosecond time scale at residue resolution. However, NMR measurement and calculation from atomistic molecular dynamics (MD) trajectories are time-consuming. We developed GeoDSSOP (Geometry-aware Deep Sequence–Structure Order-Parameter Predictor) to predict backbone N–H S2 from one amino-acid sequence and one corresponding structure. GeoDSSOP combines frozen ESM-2 residue embeddings with a geometric vector perceptron-style encoder of a residue graph. A gated residual module fuses the sequence and structure representations, and a dilated temporal convolutional network gives a bounded prediction for each residue. We trained these models on 3,490 proteins with MD-iRED labels or S-OPPE pseudo-labels. The two label sources were weighted equally, and homologous proteins were kept in the same data partition. The model gave a protein-macro PCC of 0.8250 on a protected and homology-separated test set of 136 MD-labelled proteins, and 0.8424 on a historical experimental NMR benchmark of ten proteins that was not used to train OPPE, S-OPPE, or GeoDSSOP. In a matched three-run ablation, adding the structure pathway increased the mean protein-macro PCC on MD-iRED validation from 0.7220 to 0.8026. This result shows that one protein structure provides useful information beyond the sequence pathway. Label-source ablations also showed domain-specific performance and asymmetric transfer. Balanced dual-source training gave the best overall performance on the mixed-domain development set. GeoDSSOP can rapidly produce residue-level fast-dynamics profiles from a sequence and one available structure. It does not require a new MD simulation or a multi-conformer structure ensemble, but it remains a statistical predictor and does not replace experiment or trajectory-based analysis.
Related
- KAT-Net: a hyperbolic routing framework for joint lysine acylation site prediction with shared sequence–structure representations
- Beyond structure prediction: Translating D‐I‐TASSER models into functional and mechanistic hypotheses
- Representation learning for low-data protein fitness prediction using sequence–structure embeddings and graph neural networks
- CBAT-Net: a unified CNN–BiLSTM–MultiHead transformer architecture for joint Q3/Q8 protein secondary structure prediction
- Rapid classification of β-lactamase activity through AI-based structure prediction and electric field analysis
BibTeX
@misc{tang2026geodssop,
title = {{GeoDSSOP: Geometry-aware prediction of protein backbone order parameters from sequence and structure}},
author = {Wanjia Tang and Qianqian Wang and Xiongjie Xiao and Xu Zhang and Bin Jiang and Maili Liu},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008246/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008246/v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - GeoDSSOP: Geometry-aware prediction of protein backbone order parameters from sequence and structure
AU - Wanjia Tang
AU - Qianqian Wang
AU - Xiongjie Xiao
AU - Xu Zhang
AU - Bin Jiang
AU - Maili Liu
PY - 2026
DA - 2026/09/03
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008246/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008246/v1
AB - Backbone amide order parameters (S2) describe protein motions on the picosecond-to-nanosecond time scale at residue resolution. However, NMR measurement and calculation from atomistic molecular dynamics (MD) trajectories are time-consuming. We developed GeoDSSOP (Geometry-aware Deep Sequence–Structure Order-Parameter Predictor) to predict backbone N–H S2 from one amino-acid sequence and one corresponding structure. GeoDSSOP combines frozen ESM-2 residue embeddings with a geometric vector perceptron-style encoder of a residue graph. A gated residual module fuses the sequence and structure representations, and a dilated temporal convolutional network gives a bounded prediction for each residue. We trained these models on 3,490 proteins with MD-iRED labels or S-OPPE pseudo-labels. The two label sources were weighted equally, and homologous proteins were kept in the same data partition. The model gave a protein-macro PCC of 0.8250 on a protected and homology-separated test set of 136 MD-labelled proteins, and 0.8424 on a historical experimental NMR benchmark of ten proteins that was not used to train OPPE, S-OPPE, or GeoDSSOP. In a matched three-run ablation, adding the structure pathway increased the mean protein-macro PCC on MD-iRED validation from 0.7220 to 0.8026. This result shows that one protein structure provides useful information beyond the sequence pathway. Label-source ablations also showed domain-specific performance and asymmetric transfer. Balanced dual-source training gave the best overall performance on the mixed-domain development set. GeoDSSOP can rapidly produce residue-level fast-dynamics profiles from a sequence and one available structure. It does not require a new MD simulation or a multi-conformer structure ensemble, but it remains a statistical predictor and does not replace experiment or trajectory-based analysis.
ER -