Use LEGOLAS Online

LEGOLAS: Fast Machine Learning Prediction of Protein NMR Backbone Chemical Shifts

LEGOLAS is a fully open source, TorchANI-based neural network model that performs rapid prediction of protein backbone NMR chemical shifts from 3D structures. It was developed by Mikayla Darrows, Dimuthu Kodituwakku, Jinze Xue, Ignacio Pickering, Nicholas Terrel and Adrian Roitberg in the Department of Chemistry at the University of Florida, and was described in a ChemRxiv preprint posted on 3 January 2025 (CC-BY-NC 4.0, not yet peer reviewed). It predicts chemical shifts for the backbone atoms N, Ca, Cb, C', HN and Ha.

NMR chemical shifts are among the most widely measured observables for proteins, and comparing them with values computed from a structural model is a standard way to check whether that model agrees with experiment. The widely used SHIFTX2 program is the reference point for this task, but the authors designed LEGOLAS to be fast so that shifts can be predicted for a very large number of input structures, such as every frame of a molecular dynamics trajectory.

Speed matters because a single structure rarely represents a flexible protein. The authors report that LEGOLAS predicts chemical shifts at least one order of magnitude faster than SHIFTX2, which makes ensemble-level comparison with experiment practical.

Why Fast NMR Chemical Shift Prediction Matters

When a structure or simulation is compared with NMR data, the prediction step is repeated for every conformation considered. A slow predictor limits that comparison to a handful of structures. The authors highlight two uses that a faster predictor opens up:

• Molecular dynamics ensembles: in a simulation of the E. coli protein BBL, averaging predicted shifts over a set of trajectory frames agreed with experiment substantially better than using a single frame.

• Native-state recognition: the authors show that LEGOLAS can be applied successfully to recognizing the native state of a protein among a set of decoys.

How LEGOLAS Works and What It Was Trained On

The available source material is the preprint abstract and supplementary tables, so this page does not describe the network architecture in detail. What the source does state is that the model builds on TorchANI, is open source, and was trained and evaluated on the SHIFTX2 dataset. Accuracy was assessed in two ways:

1. 5-fold cross validation: averaged validation RMSE over 141,472 chemical shifts across the six backbone atom types.

2. Independent test set: RMSE on 55,839 held-out chemical shifts from the SHIFTX2 test set.

LEGOLAS Accuracy Compared with SHIFTX2 and SHIFTX+

The supplement compares LEGOLAS with SHIFTX2 and SHIFTX+ on the SHIFTX2 test set, using the downloadable models run by the authors. RMSE is in ppm and lower is better.

Read honestly, SHIFTX2 is the more accurate predictor by RMSE for most atom types in this comparison. LEGOLAS is better on Cb (1.40 versus 1.89 ppm, with a higher correlation), and its main advantage is speed, at least ten times faster. The authors also report SHIFTX2 values from its original publication, which are lower still for several atoms, so the choice depends on whether throughput or maximum per-structure accuracy matters more for your project.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with LEGOLAS on Tamarind Bio

• Checking simulations against NMR data: predict shifts for many frames of a molecular dynamics trajectory and average them to compare with experimental values.

• Ranking candidate structures: use predicted shifts to help recognize native-like structures among decoys.

• Screening large structure sets: because the model is at least ten times faster than SHIFTX2, many models can be evaluated in one run.

• Quick backbone shift estimates: obtain N, Ca, Cb, C', HN and Ha predictions for a single structure before an experiment or assignment.

How to Use LEGOLAS on Tamarind Bio

1. Log in: sign in at tamarind.bio and open the LEGOLAS tool.

2. Provide a protein structure: upload the 3D structure whose backbone shifts you want to predict, for example a PDB file or a set of frames, if exposed.

3. Select options: adjust any prediction settings the tool exposes; the source does not define user parameters.

4. Run the job: submit it and wait for the predictions to complete.

5. Download results: retrieve the predicted chemical shifts for the backbone atoms.

6. Interpret: compare predictions with experimental shifts, keeping the per-atom RMSE values above in mind as typical error.

7. Use downstream: average predictions across frames or compare candidate structures against your data.

Things to Keep in Mind

• The work is a preprint and has not been peer reviewed; the data may be preliminary.

• On the SHIFTX2 test set, SHIFTX2 gives lower RMSE than LEGOLAS for most atom types; LEGOLAS gains are in speed and in Cb.

• Test RMSE is higher than cross-validation RMSE for every atom type, most visibly for Cb (1.40 versus 1.08 ppm).

• The model covers backbone atoms only (N, Ca, Cb, C', HN, Ha).

• The reported ensemble-averaging result comes from one protein, BBL, so results for other systems may differ.

Source: Darrows M., Kodituwakku D., Xue J., Pickering I., Terrel N., Roitberg A. LEGOLAS: a Machine Learning method for rapid and accurate predictions of protein NMR chemical shifts. ChemRxiv preprint, 3 January 2025 (CC-BY-NC 4.0, not peer reviewed). DOI: 10.26434/chemrxiv-2025-w2qn8.