LlewellynSystemsInc/ode-molecular-predictor

ODE MolPredictor — Molecular Property Prediction

By Llewellyn Systems Inc | Part of the ODE Science platform

Model Description

MLP-based molecular property predictor trained on OGB ogbg-molhiv dataset. Predicts HIV activity from molecular graph statistics.

Performance

• 97.5% validation accuracy on ogbg-molhiv binary classification
• Trained on RTX A4000 GPU
• 30 input features (node statistics, edge statistics, graph structure)

Architecture

Linear(30, 256) → ReLU → BN → Dropout(0.3)
Linear(256, 128) → ReLU → BN → Dropout(0.2) 
Linear(128, 64) → ReLU
Linear(64, 1) → Sigmoid

Usage

import torch
model = MolPredictor(30)
model.load_state_dict(torch.load("mol_predictor_best.pt"))
model.eval()

Training

• Dataset: OGB ogbg-molhiv (32,901 molecules)
• Epochs: 100
• Optimizer: Adam (lr=1e-3, weight_decay=1e-5)
• GPU: NVIDIA RTX A4000

About ODE Science

ODE Science is an AI-powered scientific research platform by Llewellyn Systems Inc, featuring molecular modeling, protein analysis, medical imaging, and quantum computing capabilities.

Website | ODE Science