Related tools
MD Trajectory Analysis
Analyze molecular dynamics trajectories using a ProteinIQ wrapper pinned to MDAnalysis 2.9.0. Calculate RMSD, residue-aggregated RMSF, radius of gyration, distance tracking, and additional trajectory observables from standard topology and trajectory files.
SuperWater
Predict protein hydration sites from a structure using a diffusion model with ESM features and a confidence-filtering head.
AllMetal3D
Predict metal and water binding sites in protein structures using 3D convolutional neural networks (AllMetal3D + Water3D).
DR-BERT
DR-BERT is a compact protein language model that predicts intrinsically disordered regions (IDRs) in proteins. It outputs per-residue disorder probability scores (0–1) from amino acid sequences, enabling fast and accurate annotation of disordered regions without structural data.
Aggrescan3D
Faithful static-mode Aggrescan3D wrapper for per-residue aggregation propensity analysis from a single protein structure.
MDGen
MDGen is a generative AI model for molecular dynamics trajectory generation. Generate physically plausible conformational ensembles from a single protein structure, enabling rapid exploration of protein dynamics without expensive MD simulations.
PROPKA 3
Predict pKa values of ionizable groups in proteins and protein-ligand complexes from 3D structure. PROPKA calculates environment-driven pKa shifts for standard ionizable residues, terminal groups, and supported ligand atom types.
CleaveNet
Official CleaveNet wrapper for matrix metalloproteinase cleavage prediction and peptide generation. Predict cleavage z-scores plus uncertainty across 17 MMP variants, evaluate against truth z-scores, or generate candidate peptides unconditionally or from MMP z-score profiles.
DeepEMhancer
DeepEMhancer is a deep learning-based post-processing tool for cryo-EM maps. It performs automatic sharpening, masking, and denoising in a single step without requiring an atomic model. Supports half-map inputs for improved local mask estimation.
DLKcat
DLKcat predicts enzyme turnover numbers (kcat values) from protein sequences and substrate structures using deep learning. Combines CNN and GNN architectures for accurate kinetic parameter prediction.
What is ORB v3?
ORB v3 is a universal interatomic potential developed by Orbital Materials. It uses a machine learning force field to predict energies, atomic forces, and stress tensors for arbitrary atomic systems — proteins, small molecules, crystals, and interfaces alike. Given a 3D structure, ORB v3 can relax it to a local energy minimum through geometry optimization, producing a physically refined structure along with per-step energy and force data.
Traditional force fields like AMBER or CHARMM rely on hand-tuned parameters for specific atom types. Machine learning force fields like ORB learn interaction potentials directly from quantum mechanical calculations (density functional theory), capturing complex many-body effects that classical force fields approximate poorly. The result is near-DFT accuracy at a fraction of the computational cost.
ORB v3 expands the performance-speed-memory Pareto frontier relative to earlier versions, achieving near state-of-the-art accuracy with >10x lower latency and >8x lower memory usage. This makes it practical for high-throughput screening and larger systems that would be prohibitively expensive with DFT.
How does ORB v3 work?
The model architecture is a Graph Network-based Simulator (GNS) augmented with smoothed graph attention. Atoms are represented as nodes in a graph, connected by edges based on spatial proximity. Through iterative message passing, each atom's representation is updated based on its neighbors — early iterations capture local bonding interactions, and deeper layers compose these into larger structural features.
Training proceeds in two stages. First, a denoising diffusion pretraining phase teaches the model to recover atomic positions from noisy inputs, which is mathematically equivalent to learning a force field. Second, the model is fine-tuned on mixed datasets of DFT calculations to predict per-atom forces, total system energy, and unit cell stress.
Conservative vs. direct models
Conservative models compute forces as the negative gradient of the predicted energy, guaranteeing energy conservation by construction. This is required for certain applications like NVE molecular dynamics where total energy must be preserved. Direct models predict forces independently from energy, which is significantly faster and uses less memory. For geometry optimization — the primary use case on ProteinIQ — both approaches produce comparable results.
Materials vs. molecules variants
The materials variants (omat) are trained on bulk crystal datasets and generalize to non-periodic systems like proteins. The molecules variants (omol) are trained on the Open Molecules 2025 dataset of over 100 million DFT calculations at the B97M-V/def2-TZVPD level, and are purpose-built for isolated molecular systems. For protein structures, either family works; for small organic molecules, the omol variants may be more accurate.
How to use ORB v3 online
ProteinIQ runs ORB v3 on GPU infrastructure with no installation or Python environment setup required. Upload a structure and receive an optimized geometry in minutes.
Input
| Input | Description |
|---|---|
Structure | PDB or CIF file, or a 4-character PDB ID fetched from RCSB. |
Settings
Model settings
| Setting | Description |
|---|---|
Model variant | Which ORB v3 model to use. Conservative - Materials (default) is the safest general-purpose choice. Direct variants are faster. Molecules variants are trained specifically on molecular data. |
Precision | Float32 High (default) is sufficient for most cases. Float32 Highest improves numerical accuracy at the cost of speed. |
Optimization settings
| Setting | Description |
|---|---|
Convergence threshold (fmax) | Maximum allowed force on any atom for the optimization to be considered converged, in eV/Å. Default: 0.01. Lower values produce tighter geometries but require more steps. |
Max optimization steps | Upper limit on BFGS optimization iterations. Default: 500. Increase for large or strained structures. |
Output
The tool produces two files:
| File | Description |
|---|---|
optimized_structure.pdb | The geometry-optimized structure, viewable in the 3D viewer. |
optimization_trajectory.csv | Energy and maximum force at each optimization step, useful for assessing convergence. |
The job summary reports initial and final energies (eV), energy change, maximum forces before and after optimization, number of steps taken, and whether the optimization converged within the step limit.
Interpreting results
Convergence: If the final maximum force is below the fmax threshold, the structure reached a local energy minimum. If not, the optimization hit the step limit — try increasing Max optimization steps or loosening fmax.
Energy change: A large negative energy change (tens to hundreds of eV) typically indicates the input structure had significant steric clashes or was far from equilibrium. Structures already near a minimum will show small energy changes.
Trajectory CSV: Plotting energy_eV against step should show a monotonically decreasing curve that flattens as the structure relaxes. The fmax_eV_A column shows how the worst-case atomic force decreases — a smooth decline indicates well-behaved optimization, while oscillations may suggest a rugged energy surface.
Limitations
- ORB v3 performs geometry optimization (energy minimization), not molecular dynamics. For time-dependent simulations, use OpenMM.
- The force field does not model explicit solvent, ions, or membrane environments. Structures are optimized in vacuum.
- Accuracy depends on how well the training data covers the chemistry of the input system. Unusual metal coordination or exotic ligands may be poorly represented.
- Very large structures (>5,000 atoms) may require extended optimization times.