ACANet v3

This release provides the source code and resources associated with the manuscript:

Activity-cliff awareness enables robust graph learning for predicting molecular properties

ACANet is a graph-learning framework designed to improve molecular property prediction by explicitly incorporating activity-cliff awareness into model training. The released code includes implementations of the activity-cliff-aware learning strategy, model training and evaluation scripts, and analysis utilities used to reproduce the main computational experiments reported in the paper.

Contents

This release includes:

• Source code for training and evaluating ACANet models
• Implementation of the activity-cliff-aware loss strategy
• Scripts for molecular property prediction benchmarks
• Utilities for activity-cliff and non-activity-cliff analysis
• Code used for generating model comparison and visualization results
• Configuration files and examples for reproducing the reported experiments

Related manuscript

Activity-cliff awareness enables robust graph learning for predicting molecular properties

Data availability

The datasets used in this study, including the 9 small-sample narrow scaffold datasets, the 30 large-sample mixed scaffold datasets, the 3 MMP datasets, and the 10 ADMET datasets, are available from the Molecular Property Cliff Task collection in Therapeutics Data Commons. Additional dataset information is described in the Data Availability statement of the manuscript.

Code availability

This repository contains the source code and Jupyter notebooks used in the study. The code supports reproducible training, evaluation, and analysis of ACANet and baseline models for molecular property prediction.

Citation

If you use this code, please cite the associated manuscript and this archived software release.