chemprop

MoleculeNet benchmark dataset & MolMapNet dataset

Antibiotic discovery using graph deep learning, with Chemprop.

This repository contains code and documentation for participating in the OpenADMET + ExpansionRx Blind Challenge. The goal of this challenge is to develop machine learning models to predict various ADMET properties of small molecules using the provided dataset.

Hydrolysis reaction rate estimation from chemical structure

Multi-task learning (BBB + P-gp) for blood-brain barrier permeability prediction

AI-powered virtual screening platform: 3-tier ML + physics pipeline for drug discovery binding affinity prediction

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

This project is focused on the development and implementation of chemical molecular models following MolSSI's best practices for version control and Python package management. The project is intended for beginners who are interested in working with chemical molecular simulations.

AI-driven ADMET property prediction platform utilizing Chemprop-RDKit models trained on Therapeutics Data Commons (TDC) datasets. Features fast molecular property evaluation, radar visual summaries, drug candidate profiling, and interactive web visualization tools.

TADF ST Gap prediction with ChemProp D-MPNN, reproducible validation, and error analysis

Neural network for predicting Fukui indices using Kernel-based Attention Networks (KAN) with Chebyshev graph convolutions.

PROTACXtend — open-source agentic AI platform for PROTAC discovery & design: generative linker assembly, ternary complex simulation, degradation (DC50/Dmax) & ADMET prediction. By Saveena Solanki & Ahuja Lab, IIIT Delhi.

Can we combine d-MPNNs and Random Forests to outperform each of them individually?

Bayesian optimisation benchmark for N-heterocyclic carbene design on xTB binding free energies: FABO, PCA, PLS and OPLS against a vanilla GP and model-free controls on DFT descriptors, ChemBERTa-2 and Mordred, plus a tuned surrogate benchmark with Chemprop and tabular foundation models.

Comparing RDKit descriptors + LightGBM against a Chemprop graph neural network for blood-brain barrier permeability, evaluated across five scaffold splits with paired statistical testing.

How much of molecular ML survives an honest split? Random splits leak 67% of scaffolds. Closing the leak costs every model about 0.13 RMSE, and no deep-model advantage was established. 45 cells over 5 ChEMBL Ki targets and 20,255 curated compounds.