molecular-machine-learning

Configuration-driven, reproducible ML workflows for molecules and materials, with DOE, auditable artifacts and agent-assisted execution.

Code for the paper: M. Haeberle, P. van Gerwen, R. Laplaza, K. R. Briling, J. Weinreich, F. Eisenbrand, C. Corminboeuf, “Integer linear programming for unsupervised training set selection in molecular machine learning” Mach. Learn.: Sci. Technol. 6 025030 (2025)

Graph Neural Networks for Molecular Property Prediction using the QM9 Quantum Chemistry Dataset.

Open molecular discovery environment for interactive design, scientific computation, molecular ML, and traceable scientific workflows.

Ensemble Deep Learning for Asymmetric Catalysis

Leakage-aware scaffold, random, and hash-based train/val/test splitting for molecular machine learning.

Kaggle Bronze Medal solution for NeurIPS Open Polymer Prediction 2025 with molecular descriptors, GNN, CatBoost/XGBoost ensemble.

A hands-on tutorial implementing Graph Convolutional Networks (GCNs) for molecular property prediction using PyTorch Geometric. Predicts water solubility from chemical structures with complete pipeline from SMILES to predictions.

Quantum Kernel Machine Learning for Drug Design A rigorous, end-to-end Qiskit implementation of quantum kernel SVMs for predicting blood-brain barrier permeability (BBBP) — a core ADMET property in CNS drug discovery — with three controlled experiments that actually test whether the quantum part is doing anything useful.

Molecular property prediction on QM9 using Graph Neural Networks with PyTorch Geometric, including GCN and bond-aware NNConv models.

Predicting when pair-specific interaction effects transfer to unseen scientific entities across drug and chemical systems.

Reproducible molecular solubility benchmarks with PyTorch graph learning, RDKit baselines, conformal uncertainty, and CPU profiling.

How much of molecular ML survives an honest split? Random splits leak 67% of scaffolds; closing the leak costs every model ~0.13 RMSE, and no deep advantage was established.

AI-based prediction of BBB permeability using molecular features, machine learning, and graph neural networks.

Physics-distilled, structure-only hydration free-energy prediction without simulation at inference

Distribution-free uncertainty for cancer drug response: a GNN predictor with split conformal prediction on GDSC1.

E(3)-equivariant GNNs for protein-ligand binding-affinity prediction (EGNN + e3nn tensor-product), with a machine-precision invariance test suite.

A balance-aware approximation for scaling leakage-aware dataset splitting, evaluated on four molecular ML benchmarks.

Visual latent structural reasoning for molecular properties and edits

Generate the full Tox21 multi-task toxicity dataset from public PubChem BioAssay records