Open-source software for molecular property prediction, drug response, and chemical experimentation.
A Streamlit-based multi-agent application for designing, simulating, and evaluating chemical experiments. Built with LangGraph and demonstrated on the Big Kahuna robotic platform from Unchained Labs.
A molecular property prediction model with intrinsic interpretability. Pretrained on ~2 million drug-like molecules; finetuned checkpoints available for solubility and cancer drug response prediction.
A tool to identify regions of varying prediction accuracy for machine learning models using molecular descriptors. Compatible with any model that exposes a feature space.
Models supporting four drug encoding methods — molecular descriptors, Morgan fingerprints, Graph Neural Networks, and Transformers — for benchmarking drug representations in cancer response prediction.
A toolkit for building molecular property prediction models across multiple representations: GNNs (PyTorch Geometric), Feed Forward Networks, SchNet, and RNNs.