hlgap-gnn3d-transformer-pcqm4mv2-v1

HOMO–LUMO Gap Predictor - Hybrid GNN + 3D + Transformer (PCQM4Mv2)

Task: Regression (HOMO–LUMO energy gap, eV)
Domain: General organic molecules (PCQM4Mv2, 3,746,620 molecules)
Model: GINE + 3D distances + Triplet interaction block + Transformer attention + RDKit globals
Checkpoint: hlgap_hybrid_pcqm4mv2_v1.pt (PyTorch)

Why this model

Density Functional Theory (DFT) is accurate but slow. This model targets fast, scalable gap prediction for large screening pipelines while retaining strong accuracy.

Reported results (held-out set)

• MAE: 0.0988 eV
• RMSE: 0.1641 eV
• R²: 0.9801
• Final MSE: 0.0269

Trained/evaluated with OGB-compliant PCQM4Mv2 splits.

Architecture (high level)

• Backbone: GINEConv layers with bond/edge encoders
• 3D cues: Euclidean distances from MMFF/UFF conformers
• TripletBlock: three-body/angular interactions
• Global context: Custom Transformer encoder (self-attention over atoms + CLS)
• Late fusion: Concatenate attention-pooled graph embedding with 6 global RDKit descriptors
• Head: MLP → scalar gap (eV)

Inputs

• SMILES (required)
• Optional 3D: one minimized conformer per molecule (if not provided, the loader can generate via RDKit; accuracy is best with consistent 3D)
• RDKit globals (6): MolWt, MolMR, MolLogP, TPSA, FractionCSP3, RotatableBonds

Usage

import torch
from rdkit import Chem
##### to be filled out

device = "cuda" if torch.cuda.is_available() else "cpu"
ckpt = "hlgap-gnn3d-transformer-pcqm4mv2-v1.pt"  

# 1) load model
# from model_def import HLGapHybrid  # same module you export here
model = HLGapHybrid(...)
model.load_state_dict(torch.load(ckpt, map_location=device))
model.to(device).eval()

# 2) featurize
smiles = "CCOc1ccc2nc(S(N)(=O)=O)sc2c1"
g, globals6 = build_graph_from_smiles(smiles, with_3d=True)  # tensors

# 3) predict (eV)
with torch.no_grad():
    y_pred = model(g.to(device), globals6.to(device)).item()
print("Predicted gap (eV):", y_pred)