VSRAJHAVEL/BioGraph-AI-Models

BioGraph AI: Deep Graph Neural Networks for Molecular Property Prediction & ADMET Profiling

BioGraph AI is an enterprise-grade Computer-Aided Drug Design (CADD) and Molecular Machine Learning platform. This repository hosts pre-trained model weights for 15 state-of-the-art graph architectures and classical baselines evaluated under strict Bemis-Murcko scaffold splitting across three canonical MoleculeNet benchmarks (ESOL, BBBP, Tox21).

Handcrafted by RAJHAVEL V S
Portfolio: vsrajhavel.github.io | LinkedIn: in/rajhavelvs | GitHub: @VSRAJHAVEL

---

🔬 Benchmark Performance (Bemis-Murcko Scaffold Split 80/10/10)

All models were trained, validated, and evaluated on local Intel CPU physical threads. Zero scaffold leakage was mathematically confirmed across all training and test partitions.

1. Aqueous Solubility (ESOL — Delaney Regression, $\log S$ in $\text{mol/L}$)

Model Architecture	Test RMSE ↓	Test MAE ↓	Test $R^2$ ↑	Checkpoint File
MPNN (Edge-Aware)	1.010	0.777	0.784	checkpoints/esol_mpnn_best.pt
GIN (Weisfeiler-Lehman)	1.054	0.812	0.764	checkpoints/esol_gin_best.pt
GCN (Kipf-Welling)	1.080	0.811	0.753	checkpoints/esol_gcn_best.pt
Baseline (RF + Morgan)	1.109	0.859	0.739	checkpoints/esol_baseline_best.pkl
GAT (Multi-Head)	1.208	0.938	0.691	checkpoints/esol_gat_best.pt

2. Blood-Brain Barrier Penetration (BBBP — Binary Classification)

Model Architecture	Test ROC-AUC ↑	Test PR-AUC ↑	Checkpoint File
Baseline (RF + Morgan)	0.853	0.978	checkpoints/bbbp_baseline_best.pkl
GIN	0.723	0.936	checkpoints/bbbp_gin_best.pt
GCN	0.719	0.951	checkpoints/bbbp_gcn_best.pt
MPNN	0.665	0.944	checkpoints/bbbp_mpnn_best.pt
GAT	0.624	0.928	checkpoints/bbbp_gat_best.pt

3. Toxicity Profiling (Tox21 — 12 Multi-Task Toxicity Pathways)

Evaluated across 12 nuclear receptor (NR) and stress response (SR) pathways with missing label masking.

Model Architecture	Test ROC-AUC ↑	Test PR-AUC ↑	Checkpoint File
MPNN (Edge-Aware)	0.757	0.304	checkpoints/tox21_mpnn_best.pt
GCN	0.754	0.310	checkpoints/tox21_gcn_best.pt
GIN	0.747	0.306	checkpoints/tox21_gin_best.pt
Baseline (RF)	0.730	0.349	checkpoints/tox21_baseline_best.pkl
GAT	0.711	0.264	checkpoints/tox21_gat_best.pt

---

🚀 Quickstart: Loading Checkpoints in Python

Installation

pip install torch torch-geometric rdkit huggingface_hub

Download Pretrained Weights

from huggingface_hub import hf_hub_download
import torch

# Download best ESOL MPNN checkpoint
checkpoint_path = hf_hub_download(
    repo_id="VSRAJHAVEL/BioGraph-AI-Models",
    filename="checkpoints/esol_mpnn_best.pt"
)

state_dict = torch.load(checkpoint_path, map_location="cpu")
print("Successfully loaded BioGraph AI MPNN checkpoint:", len(state_dict), "weight tensors")

---

🧬 Architectural Overview

1. Message Passing Neural Networks (MPNN): Formulated with edge-network conditioned message functions $M(h_v, h_w, e_{vw})$ incorporating bond stereo, conjugation, and aromaticity. Top performer for steric-dependent thermodynamic properties.
2. Graph Isomorphism Network (GIN): Employs sum-aggregation with multi-layer perceptrons approximating the 1-Weisfeiler-Lehman topological graph isomorphism test.
3. Graph Convolutional Networks (GCN): First-order localized spectral convolutions with symmetric normalized Laplacian aggregations.
4. Graph Attention Networks (GAT): LeakyReLU-parameterized multi-head attention over bonded neighborhoods.
5. Random Forest Baselines: 500-estimator ensembles trained on 2048-bit ECFP4 (Morgan radius 2) fingerprints.

---

🔗 Project Resources

• Full Application Repository: https://github.com/VSRAJHAVEL/BioGraph-AI
• Interactive Web App: Includes 3D Mol* Conformer Studio, ADMET Spider Radar, MMP Bioisosteric Optimization, PCA Chemical Space Explorer, and GNNExplainer Attribution.
• License: MIT License (Copyright 2026 RAJHAVEL V S)