orbmol / app.py
annabossler's picture
Update app.py
ed6967d verified
import os
os.environ["TORCH_DYNAMO_DISABLE"] = "1"
import tempfile
import numpy as np
import gradio as gr
from ase.io import read, write
from ase.io.trajectory import Trajectory
import subprocess, sys
from pathlib import Path
# === BUILD AND INSTALL LOCAL gradio_molecule3d ===
try:
print("🔧 Building and installing local gradio_molecule3d fork...")
base_path = Path(__file__).parent
local_pkg = base_path / "gradio_molecule3d"
# Step 1 — gradio cc install
subprocess.call(["gradio", "cc", "install"], cwd=local_pkg)
# Step 2 — gradio cc build
subprocess.call(["gradio", "cc", "build"], cwd=local_pkg)
# Step 3 — pip install generated wheel
wheel_path = local_pkg / "dist" / "gradio_molecule3d-0.0.7-py3-none-any.whl"
if not wheel_path.exists():
print("Wheel not found, listing dist contents:")
subprocess.call(["ls", "-R", str(local_pkg / "dist")])
subprocess.call(
[
sys.executable,
"-m",
"pip",
"install",
str(wheel_path),
],
cwd=base_path.parent,
)
print("gradio_molecule3d built and installed successfully!")
except Exception as e:
print(f"Error building gradio_molecule3d: {e}")
# === Import only after it's installed ===
from gradio_molecule3d import Molecule3D
from gradio_molecule3d import Molecule3D
from simulation_scripts_orbmol import load_orbmol_model, run_md_simulation, run_relaxation_simulation
import hashlib
# ==== Configuración Molecule3D ====
DEFAULT_MOLECULAR_REPRESENTATIONS = [
{
"model": 0,
"chain": "",
"resname": "",
"style": "sphere",
"color": "Jmol",
"around": 0,
"byres": False,
"scale": 0.3,
},
{
"model": 0,
"chain": "",
"resname": "",
"style": "stick",
"color": "Jmol",
"around": 0,
"byres": False,
"scale": 0.2,
},
]
DEFAULT_MOLECULAR_SETTINGS = {
"backgroundColor": "white",
"orthographic": False,
"disableFog": False,
}
# ==== Conversión a PDB para Molecule3D ====
def convert_to_pdb_for_viewer(file_path):
"""Convierte cualquier archivo a PDB para Molecule3D"""
if not file_path or not os.path.exists(file_path):
return None
try:
atoms = read(file_path)
cache_dir = os.path.join(tempfile.gettempdir(), "gradio")
os.makedirs(cache_dir, exist_ok=True)
pdb_path = os.path.join(cache_dir, f"mol_{hashlib.md5(file_path.encode()).hexdigest()[:12]}.pdb")
write(pdb_path, atoms, format="proteindatabank")
return pdb_path
except Exception as e:
print(f"Error converting to PDB: {e}")
return None
# ==== OrbMol SPE ====
def predict_molecule(structure_file, task_name, charge=0, spin_multiplicity=1):
"""Single Point Energy + fuerzas (OrbMol)"""
try:
calc = load_orbmol_model(task_name)
if not structure_file:
return "Error: Please upload a structure file", "Error", None
file_path = structure_file
if not os.path.exists(file_path):
return f"Error: File not found: {file_path}", "Error", None
if os.path.getsize(file_path) == 0:
return f"Error: Empty file: {file_path}", "Error", None
atoms = read(file_path)
if task_name in ["OMol", "OMol-Direct"]:
atoms.info = {"charge": int(charge), "spin": int(spin_multiplicity)}
atoms.calc = calc
energy = atoms.get_potential_energy()
forces = atoms.get_forces()
lines = [
f"Model: {task_name}",
f"Total Energy: {energy:.6f} eV",
"",
"Atomic Forces:"
]
for i, fc in enumerate(forces):
lines.append(f"Atom {i+1}: [{fc[0]:.4f}, {fc[1]:.4f}, {fc[2]:.4f}] eV/Å")
max_force = float(np.max(np.linalg.norm(forces, axis=1)))
lines += ["", f"Max Force: {max_force:.4f} eV/Å"]
pdb_file = convert_to_pdb_for_viewer(file_path)
return "\n".join(lines), f"Calculation completed with {task_name}", pdb_file
except Exception as e:
import traceback
traceback.print_exc()
return f"Error during calculation: {e}", "Error", None
# ==== Wrappers MD y Relax ====
def md_wrapper(structure_file, task_name, charge, spin, steps, tempK, timestep_fs, ensemble):
try:
if not structure_file:
return ("Error: Please upload a structure file", None, "", "", "", None)
traj_path, log_text, script_text, explanation = run_md_simulation(
structure_file,
int(steps),
20,
float(timestep_fs),
float(tempK),
"NVT" if ensemble == "NVT" else "NVE",
str(task_name),
int(charge),
int(spin),
)
status = f"MD completed: {int(steps)} steps at {int(tempK)} K ({ensemble})"
pdb_file = convert_to_pdb_for_viewer(traj_path)
return (status, traj_path, log_text, script_text, explanation, pdb_file)
except Exception as e:
import traceback
traceback.print_exc()
return (f"Error: {e}", None, "", "", "", None)
def relax_wrapper(structure_file, task_name, steps, fmax, charge, spin, relax_cell):
try:
if not structure_file:
return ("Error: Please upload a structure file", None, "", "", "", None)
traj_path, log_text, script_text, explanation = run_relaxation_simulation(
structure_file,
int(steps),
float(fmax),
str(task_name),
int(charge),
int(spin),
bool(relax_cell),
)
status = f"Relaxation finished (<={int(steps)} steps, fmax={float(fmax)} eV/Å)"
pdb_file = convert_to_pdb_for_viewer(traj_path)
return (status, traj_path, log_text, script_text, explanation, pdb_file)
except Exception as e:
import traceback
traceback.print_exc()
return (f"Error: {e}", None, "", "", "", None)
# ==== UI ====
with gr.Blocks(theme=gr.themes.Ocean(), title="OrbMol Demo") as demo:
with gr.Tabs():
# -------- HOME TAB --------
with gr.Tab("Home"):
with gr.Row():
# Columna izquierda con acordeones
with gr.Column(scale=1):
gr.Markdown("## Learn more about OrbMol")
with gr.Accordion("What is OrbMol?", open=False):
gr.Markdown("""
OrbMol is a suite of quantum-accurate machine learning models for molecular predictions. Built on the **Orb-v3 architecture**, OrbMol provides fast and accurate calculations of energies, forces, and molecular properties at the level of advanced quantum chemistry methods.
The models combine the transferability of universal potentials with quantum-level accuracy, making them suitable for a wide range of applications in chemistry, materials science, and drug discovery.
""")
with gr.Accordion("Available Models", open=False):
gr.Markdown("""
**OMol** and **OMol-Direct**
- **Training dataset**: OMol25 (>100M calculations on small molecules, biomolecules, metal complexes, and electrolytes)
- **Level of theory**: ωB97M-V/def2-TZVPD with non-local dispersion; solvation treated explicitly
- **Inputs**: total charge & spin multiplicity
- **Applications**: biology, organic chemistry, protein folding, small-molecule drugs, organic liquids, homogeneous catalysis
- **Caveats**: trained only on aperiodic systems → periodic/inorganic cases may not work well
- **Difference**: OMol enforces energy–force consistency; OMol-Direct relaxes this for efficiency
**OMat**
- **Training dataset**: OMat24 (>100M inorganic calculations, from Materials Project, Alexandria, and far-from-equilibrium samples)
- **Level of theory**: PBE/PBE+U with Materials Project settings; VASP 54 pseudopotentials; no dispersion
- **Inputs**: No support for spin and charge. Spin polarization included but magnetic state cannot be selected
- **Applications**: inorganic discovery, photovoltaics, alloys, superconductors, electronic/optical materials
- **Caveats**: magnetic effects may be incompletely captured
""")
with gr.Accordion("Supported File Formats", open=False):
gr.Markdown("""
OrbMol supports the following molecular structure formats:
- `.xyz` - XYZ coordinate files
- `.pdb` - Protein Data Bank format
- `.cif` - Crystallographic Information File
- `.traj` - ASE trajectory format
- `.mol` - MDL Molfile
- `.sdf` - Structure Data File
All formats are automatically converted internally for processing.
""")
with gr.Accordion("How to Use", open=False):
gr.Markdown("""
**Single Point Energy**: Upload a molecular structure and select a model to calculate energies and forces.
**Molecular Dynamics**: Run time-dependent simulations to observe molecular behavior at different temperatures and conditions.
**Relaxation/Optimization**: Find the minimum-energy configuration of your molecular structure.
Each tab provides specific parameters you can adjust to customize your calculations.
""")
with gr.Accordion("Technical Foundation", open=False):
gr.Markdown("""
All models are based on the **Orb-v3 architecture**, the latest generation of Orb universal interatomic potentials.
Key features:
- Graph neural network architecture
- Equivariant message passing
- Multi-task learning across different quantum chemistry methods
- Billions of training examples across diverse chemical spaces
- Sub-kcal/mol accuracy on test sets
""")
with gr.Accordion("Resources & Support", open=False):
gr.Markdown("""
- [Orb-v3 paper](https://arxiv.org/abs/2504.06231)
- [Orb-Models GitHub repository](https://github.com/orbital-materials/orb-models)
- For issues/questions, please open a GitHub issue or contact the developers
**Citation**: If you use OrbMol in your research, please cite the Orb-v3 paper and the relevant dataset papers (OMol25/OMat24).
""")
# Columna derecha con contenido principal
with gr.Column(scale=2):
gr.Image("logo_color_text.png",
show_share_button=False,
show_download_button=False,
show_label=False,
show_fullscreen_button=False)
gr.Markdown("# OrbMol — Quantum-Accurate Molecular Predictions")
gr.Markdown("""
Welcome to the OrbMol demo! This interactive platform allows you to explore the capabilities of our quantum-accurate machine learning models for molecular simulations.
## Quick Start
Use the tabs above to access different functionalities:
1. **Single Point Energy**: Calculate energies and forces for a given molecular structure
2. **Molecular Dynamics**: Run MD simulations using OrbMol-trained potentials
3. **Relaxation / Optimization**: Optimize molecular structures to their minimum-energy configurations
Simply upload a molecular structure file in any supported format (`.xyz`, `.pdb`, `.cif`, `.traj`, `.mol`, `.sdf`) and select the appropriate model for your system.
## Model Selection Guide
**Choose OMol/OMol-Direct for:**
- Organic molecules and biomolecules
- Drug-like compounds
- Metal-organic complexes
- Molecules in solution
- Systems where you need to specify charge and spin
**Choose OMat for:**
- Inorganic crystals and materials
- Periodic systems
- Bulk materials and alloys
- Solid-state compounds
Explore the accordions on the left to learn more about each model's capabilities, training data, and limitations.
""")
gr.Markdown("## Try an Example")
gr.Markdown("""
To get started quickly, navigate to any of the calculation tabs above and try one of these examples:
- **Single Point Energy**: Upload a small molecule to see energy and force predictions
- **Molecular Dynamics**: Run a short simulation at 300K to observe thermal motion
- **Relaxation**: Optimize a distorted structure to find its equilibrium geometry
""")
# -------- SPE --------
with gr.Tab("Single Point Energy"):
with gr.Row():
with gr.Column(scale=2):
gr.Markdown("# OrbMol — Quantum-Accurate Molecular Predictions")
gr.Markdown("**Supported formats:** .xyz, .pdb, .cif, .traj, .mol, .sdf")
xyz_input = gr.File(
label="Upload Structure File",
file_types=[".xyz", ".pdb", ".cif", ".traj", ".mol", ".sdf"],
file_count="single"
)
task_name_spe = gr.Radio(
["OMol", "OMat", "OMol-Direct"],
value="OMol",
label="Model Type"
)
with gr.Row():
charge_input = gr.Slider(-10, 10, 0, step=1, label="Charge")
spin_input = gr.Slider(1, 11, 1, step=1, label="Spin Multiplicity")
run_spe = gr.Button("Run OrbMol Prediction", variant="primary")
with gr.Column(variant="panel", min_width=500):
spe_out = gr.Textbox(label="Energy & Forces", lines=15, interactive=False)
spe_status = gr.Textbox(label="Status", interactive=False)
spe_viewer = Molecule3D(
label="Input Structure Viewer",
reps=DEFAULT_MOLECULAR_REPRESENTATIONS,
config=DEFAULT_MOLECULAR_SETTINGS
)
task_name_spe.change(
lambda x: (
gr.update(visible=x in ["OMol", "OMol-Direct"]),
gr.update(visible=x in ["OMol", "OMol-Direct"])
),
[task_name_spe],
[charge_input, spin_input]
)
run_spe.click(
predict_molecule,
[xyz_input, task_name_spe, charge_input, spin_input],
[spe_out, spe_status, spe_viewer]
)
# -------- MD --------
with gr.Tab("Molecular Dynamics"):
with gr.Row():
with gr.Column(scale=2):
gr.Markdown("## Molecular Dynamics Simulation")
xyz_md = gr.File(
label="Upload Structure File",
file_types=[".xyz", ".pdb", ".cif", ".traj", ".mol", ".sdf"],
file_count="single"
)
task_name_md = gr.Radio(
["OMol", "OMat", "OMol-Direct"],
value="OMol",
label="Model Type"
)
with gr.Row():
charge_md = gr.Slider(-10, 10, 0, step=1, label="Charge")
spin_md = gr.Slider(1, 11, 1, step=1, label="Spin Multiplicity")
with gr.Row():
steps_md = gr.Slider(10, 2000, 100, step=10, label="Steps")
temp_md = gr.Slider(10, 1500, 300, step=10, label="Temperature (K)")
with gr.Row():
timestep_md = gr.Slider(0.1, 5.0, 1.0, step=0.1, label="Timestep (fs)")
ensemble_md = gr.Radio(["NVE", "NVT"], value="NVE", label="Ensemble")
run_md_btn = gr.Button("Run MD Simulation", variant="primary")
with gr.Column(variant="panel", min_width=520):
md_status = gr.Textbox(label="MD Status", interactive=False)
md_traj = gr.File(label="Trajectory (.traj)", interactive=False)
md_viewer = Molecule3D(
label="MD Result Viewer",
reps=DEFAULT_MOLECULAR_REPRESENTATIONS,
config=DEFAULT_MOLECULAR_SETTINGS
)
md_log = gr.Textbox(label="Log", interactive=False, lines=15)
md_script = gr.Code(label="Reproduction Script", language="python", interactive=False, lines=20)
md_explain = gr.Markdown()
task_name_md.change(
lambda x: (
gr.update(visible=x in ["OMol", "OMol-Direct"]),
gr.update(visible=x in ["OMol", "OMol-Direct"])
),
[task_name_md],
[charge_md, spin_md]
)
run_md_btn.click(
md_wrapper,
[xyz_md, task_name_md, charge_md, spin_md, steps_md, temp_md, timestep_md, ensemble_md],
[md_status, md_traj, md_log, md_script, md_explain, md_viewer]
)
# -------- Relax --------
with gr.Tab("Relaxation / Optimization"):
with gr.Row():
with gr.Column(scale=2):
gr.Markdown("## Structure Relaxation/Optimization")
xyz_rlx = gr.File(
label="Upload Structure File",
file_types=[".xyz", ".pdb", ".cif", ".traj", ".mol", ".sdf"],
file_count="single"
)
task_name_rlx = gr.Radio(
["OMol", "OMat", "OMol-Direct"],
value="OMol",
label="Model Type"
)
with gr.Row():
steps_rlx = gr.Slider(1, 2000, 300, step=1, label="Max Steps")
fmax_rlx = gr.Slider(0.001, 0.5, 0.05, step=0.001, label="Fmax (eV/Å)")
with gr.Row():
charge_rlx = gr.Slider(-10, 10, 0, step=1, label="Charge")
spin_rlx = gr.Slider(1, 11, 1, step=1, label="Spin")
relax_cell = gr.Checkbox(False, label="Relax Unit Cell")
run_rlx_btn = gr.Button("Run Optimization", variant="primary")
with gr.Column(variant="panel", min_width=520):
rlx_status = gr.Textbox(label="Status", interactive=False)
rlx_traj = gr.File(label="Trajectory (.traj)", interactive=False)
rlx_viewer = Molecule3D(
label="Optimized Structure Viewer",
reps=DEFAULT_MOLECULAR_REPRESENTATIONS,
config=DEFAULT_MOLECULAR_SETTINGS
)
rlx_log = gr.Textbox(label="Log", interactive=False, lines=15)
rlx_script = gr.Code(label="Reproduction Script", language="python", interactive=False, lines=20)
rlx_explain = gr.Markdown()
task_name_rlx.change(
lambda x: (
gr.update(visible=x in ["OMol", "OMol-Direct"]),
gr.update(visible=x in ["OMol", "OMol-Direct"])
),
[task_name_rlx],
[charge_rlx, spin_rlx]
)
run_rlx_btn.click(
relax_wrapper,
[xyz_rlx, task_name_rlx, steps_rlx, fmax_rlx, charge_rlx, spin_rlx, relax_cell],
[rlx_status, rlx_traj, rlx_log, rlx_script, rlx_explain, rlx_viewer]
)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)