Protein chemical shift prediction in web browsers: seven major programs ported, benchmarked, and studied for what they help to reveal about NMR structures, AlphaFold models and molecular simulations

Motivation. Predicting NMR chemical shifts from protein structures is routine in structure validation, resonance assignment and the analysis of molecular dynamics simulations, but the available programs come as platform-specific executables, aging software stacks or upload-based servers, and some are no longer distributed. Results. Assisted by Claude Code, I ported seven chemical shift predictors to run entirely in the web browser with no installation and no data leaving the user's computer: UCBShift-X, SHIFTX2, SPARTA+, PPM_One (with its MD-ensemble variant PPM), CamShift, LEGOLAS and the random-coil predictor POTENCI. The web apps share similar interfaces that work for single structures or ensembles (from NMR, MD simulations and emulations, docking poses, etc.), and each reproduces its original implementation to printed precision. On 153 proteins with BMRB assignments, UCBShift-X is the most accurate structure-based predictor (best for 71 % of proteins), followed by SHIFTX2, SPARTA+ and PPM_One, and then by LEGOLAS and CamShift. However, I found that the input structure matters more than the predictor: For every program, AlphaFold models and crystal structures give about 15 % lower errors than the deposited NMR structures! Averaging over MD trajectories also improves most programs, especially CamShift, LEGOLAS and PPM. The ranking holds on three recently determined NMR structures not yet released at the PDB/BMRB, and comparing predictions on the NMR and AlphaFold 3 models shows which parts of each structure the shifts support and which parts they flag as potential outliers or assignment errors. Finally, I ported SLAE-CS, a very recent neural network that predicts only backbone 15N shifts: on the deposited NMR structures it is about as accurate as UCBShift-X, but on AlphaFold models and crystal structures its 15N errors are 17–28 % lower. Availability. Static web applications at https://pdbms.altervista.org/web-csp/ within the PDB Manipulation Suite platform (https://pdbms.altervista.org). For a full report with all methods, tables and data, see the semi-automated, extended version of this preprint at https://pdbms.altervista.org/web-csp/predictions_report/.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-17
DOI
https://doi.org/10.26434/chemrxiv.15012400/v1
Primary Topic
Advanced NMR Techniques and Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Protein chemical shift prediction in web browsers: seven major programs ported, benchmarked, and studied for what they help to reveal about NMR structures, AlphaFold models and molecular simulations

Luciano A. Abriata
ChemRxiv
Advanced NMR Techniques and Applications
preprint

Protein chemical shift prediction in web browsers: seven major programs ported, benchmarked, and studied for what they help to reveal about NMR structures, AlphaFold models and molecular simulations

Luciano A. Abriata
preprint en

Abstract

Motivation. Predicting NMR chemical shifts from protein structures is routine in structure validation, resonance assignment and the analysis of molecular dynamics simulations, but the available programs come as platform-specific executables, aging software stacks or upload-based servers, and some are no longer distributed. Results. Assisted by Claude Code, I ported seven chemical shift predictors to run entirely in the web browser with no installation and no data leaving the user's computer: UCBShift-X, SHIFTX2, SPARTA+, PPM_One (with its MD-ensemble variant PPM), CamShift, LEGOLAS and the random-coil predictor POTENCI. The web apps share similar interfaces that work for single structures or ensembles (from NMR, MD simulations and emulations, docking poses, etc.), and each reproduces its original implementation to printed precision. On 153 proteins with BMRB assignments, UCBShift-X is the most accurate structure-based predictor (best for 71 % of proteins), followed by SHIFTX2, SPARTA+ and PPM_One, and then by LEGOLAS and CamShift. However, I found that the input structure matters more than the predictor: For every program, AlphaFold models and crystal structures give about 15 % lower errors than the deposited NMR structures! Averaging over MD trajectories also improves most programs, especially CamShift, LEGOLAS and PPM. The ranking holds on three recently determined NMR structures not yet released at the PDB/BMRB, and comparing predictions on the NMR and AlphaFold 3 models shows which parts of each structure the shifts support and which parts they flag as potential outliers or assignment errors. Finally, I ported SLAE-CS, a very recent neural network that predicts only backbone 15N shifts: on the deposited NMR structures it is about as accurate as UCBShift-X, but on AlphaFold models and crystal structures its 15N errors are 17–28 % lower. Availability. Static web applications at https://pdbms.altervista.org/web-csp/ within the PDB Manipulation Suite platform (https://pdbms.altervista.org). For a full report with all methods, tables and data, see the semi-automated, extended version of this preprint at https://pdbms.altervista.org/web-csp/predictions_report/.

ChemRxiv
École Polytechnique Fédérale de Lausanne (CH)
Advanced NMR Techniques and Applications
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