Evaluating machine learning models for predicting pesticides toxicity to honey bees

This repository contains code for building and evaluating machine learning models using ApisTox dataset, from publication in Scientific Data. ArXiv preprint is available here.

Dataset is shared under CC-BY-NC-4.0 license Code is shared under MIT license, except for huggingmolecules library code, which uses Apache 2.0 license.

Reproducing results

Install uv and set up the virtual environment with uv sync. All dependencies are pinned in pyproject.toml and uv.lock files. You can also use the provided Makefile and run make install.

Dataset files

All input data files from ApisTox dataset are in data directory.

Code for experiments

All relevant scripts are in src directory.

Models training & evaluation:

• baselines.py - atom counts, LTP and MOLTOP baselines
• fingerprints.py - molecular fingerprints
• graph_kernels.py - graph kernels
• gnns.py - graph neural networks trained from scratch
• embedding_models.py - pretrained neural models for molecular embeddings

Data and results analysis:

• analysis_datasets_features.py - analysis and comparison of dataset properties
• analysis_counterfactuals.py - counterfactual examples model interpretability
• analysis_uncertain_data.py - predictions for uncertain molecules

huggingmolecules directory contains huggingmolecules library code. It uses the original Apache 2.0 license. The only change to the original code was automatic reformatting with ruff.