Structural Context Collapse in Deep-Learning Predictors: A Benchmark of Outer Membrane Protein Loop Decontextualization and a Backbone-Centric Filter for Surface Epitope Identification

Abstract Deep-learning structure predictors (AlphaFold2, ESMFold) require global context. Modeling outer membrane protein (OMP) loops without transmembrane scaffolds creates an out-of-distribution challenge. Across 28 non-redundant OMP crystal structures (1.40–3.10 Å), isolated loops collapsed, inverting core solvent accessibility (native core RSA: 3.6 ± 0.7% inverted to 58.7 ± 4.5% [AF2] and 62.1 ± 4.8% [ESMFold] for buried positions; mean loop RSA: 32.1 ± 10.0%), yielding 0.00% true negative rate (0/28 retaining core burial; 95% Clopper–Pearson CI: [0.00%, 10.15%]) across RSA thresholds 5%–20%. Intact models preserved burial (AF2: 27/28 [96.43%], residue TNR 96.89% ± 5.6%; ESMFold: 26/28 [92.86%], residue TNR 93.21% ± 4.2%). Mean pLDDT fell from 83.0 ± 5.1 to 66.9 ± 4.8, correlating with backbone Cα RMSD (r = − 0.68, p < 0.001; 4.83 ± 1.92 Å); 27/28 intact proteins (96.4%) maintained pLDDT > 70. Mixed-effects modeling confirmed core RSA increases (β = +28.5%, p < 0.0001). A backbone-centric Cα contact density filter (CDα ≤ 10.17; ≤ 10 neighbors at rc = 8.0 Å) identified exposed loops, achieving pilot residue-level macro-averaged AUC of 0.80 ± 0.08 (pooled AUC = 0.96; n = 3 test proteins). Prioritized Pseudomonas aeruginosa epitopes were assembled into a 93-residue chimera. Explicit-solvent molecular dynamics (4 systems × 5 × 300 ns; 6.0 µs CHARMM36m + 6.0 µs Amber ff14SB = 12.0 µs total; 6.0 µs/force field) preserved epitope exposure (41.8 ± 1.2 vs. 28.3 ± 3.1 nm²; Cohen’s d = 4.8). Predictors require intact scaffolding for OMP loop modeling; CDα provides an exploratory filter for full-length prioritization.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11308549/v1
Primary Topic
Protein Structure and Dynamics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Structural Context Collapse in Deep-Learning Predictors: A Benchmark of Outer Membrane Protein Loop Decontextualization and a Backbone-Centric Filter for Surface Epitope Identification

Ehsan Kazemi Moghaddam, Mahsa Manafi Varkiani, Mohammad Reza Rahbar, Yasamin Nasiri Khonsari et al.
1 links
Research Square
Protein Structure and Dynamics
preprint

Structural Context Collapse in Deep-Learning Predictors: A Benchmark of Outer Membrane Protein Loop Decontextualization and a Backbone-Centric Filter for Surface Epitope Identification

Ehsan Kazemi Moghaddam, Mahsa Manafi Varkiani, Mohammad Reza Rahbar, Yasamin Nasiri Khonsari, Anahita Hessami, Safa Samadzadeh Etehadi
preprint en

Abstract

Abstract Deep-learning structure predictors (AlphaFold2, ESMFold) require global context. Modeling outer membrane protein (OMP) loops without transmembrane scaffolds creates an out-of-distribution challenge. Across 28 non-redundant OMP crystal structures (1.40–3.10 Å), isolated loops collapsed, inverting core solvent accessibility (native core RSA: 3.6 ± 0.7% inverted to 58.7 ± 4.5% [AF2] and 62.1 ± 4.8% [ESMFold] for buried positions; mean loop RSA: 32.1 ± 10.0%), yielding 0.00% true negative rate (0/28 retaining core burial; 95% Clopper–Pearson CI: [0.00%, 10.15%]) across RSA thresholds 5%–20%. Intact models preserved burial (AF2: 27/28 [96.43%], residue TNR 96.89% ± 5.6%; ESMFold: 26/28 [92.86%], residue TNR 93.21% ± 4.2%). Mean pLDDT fell from 83.0 ± 5.1 to 66.9 ± 4.8, correlating with backbone Cα RMSD (r = − 0.68, p < 0.001; 4.83 ± 1.92 Å); 27/28 intact proteins (96.4%) maintained pLDDT > 70. Mixed-effects modeling confirmed core RSA increases (β = +28.5%, p < 0.0001). A backbone-centric Cα contact density filter (CDα ≤ 10.17; ≤ 10 neighbors at rc = 8.0 Å) identified exposed loops, achieving pilot residue-level macro-averaged AUC of 0.80 ± 0.08 (pooled AUC = 0.96; n = 3 test proteins). Prioritized Pseudomonas aeruginosa epitopes were assembled into a 93-residue chimera. Explicit-solvent molecular dynamics (4 systems × 5 × 300 ns; 6.0 µs CHARMM36m + 6.0 µs Amber ff14SB = 12.0 µs total; 6.0 µs/force field) preserved epitope exposure (41.8 ± 1.2 vs. 28.3 ± 3.1 nm²; Cohen’s d = 4.8). Predictors require intact scaffolding for OMP loop modeling; CDα provides an exploratory filter for full-length prioritization.

Research Square
Shahrekord University (IR), Islamic Azad University, Mashhad (IR), Shiraz University of Medical Sciences (IR), Iran University of Science and Technology (IR), Tehran University of Medical Sciences (IR)
Protein Structure and Dynamics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon