Graph Neural Networks for Molecular Property Prediction — ADMET prediction on MoleculeNet with scaffold-split evaluation and atom-level interpretability
Empirical analysis quantifying the 0.009 ROC-AUC gap between PyTorch GNNs (2D graphs) and OLMo-7B (1D SMILES) on Tox21. Exposes the Random Split illusion and provides the baseline justification for Topological State Machine (TSM) integration.
Optimizing Mistral-7B for BACE-1 inhibitor prediction using Low-Rank Adaptation (r=8) to capture peak generalization on scaffold-split molecular data. Achieved 0.8034 ROC-AUC. Built for DeepChem GSoC '26.
Molecular property prediction on BBBP using Mistral-7B. Features Nitrogen valence sanitization and property-aware prompting (MW/TPSA). Achieved 0.75+ ROC-AUC on scaffold splits for DeepChem GSoC '26.
Fine-tuning Mistral-7B with QLoRA & DeepChem for Clinical Toxicity Prediction (0.99 ROC-AUC).
Add a description, image, and links to the scaffold-split topic page so that developers can more easily learn about it.
To associate your repository with the scaffold-split topic, visit your repo's landing page and select "manage topics."