SemMol

Augmenting molecular structure representation learning using semantic biomedical knowledge

Requirements

• Python ≥ 3.8.0;
• requirements.txt contains the Python packages requirements.

Data

The data used are made available through the following box folder, where you can find:

• data/ contains pretraining dataset, the Knowledge Graphs created with the relative dictionary of entities and their ids, the classification datasets (datasets_valid_and_splits/ contais for each assay the tabular dataset with Smiles string, MACCS key, chemicals name and labels, and the training, validataion and test index for the 5 random runs), and tsne_2d_embeddings_all_chemicals_37tox21_emb.xlsx, that is a dataframe containing chemical names, MACCS keys, physical properties and the 2D t-SNE projections for all the n = 8541 chemicals that belong to the set of the 37 Tox21 assays considered;
• ckpt/ contains the pretrained GNN molecule encoder.

Models training for toxicology predictions

• Machine learning: baseline ML models can be trained by running ML.py. The results will be written in results/ML with a directory for each random runs (seed).

python ML.py

• Finetune MolCLR: MolCLR can be finetuned by running MolCLR/finetune.py. The results will be written in results/graph_structure_comptoxAI with a directory for each random runs (seed).

python MolCLR/finetune.py

• Finetune HiMol: HiMol can be finetuned by running HiMol/finetune/finetune.py. The results will be written in results/graph_structure_comptoxAI with a directory for each random runs (seed).

python HiMol/finetune/finetune.py

• Semantic GNN: Semantic GNN model can be trained by running semantic.py. The results will be written in results/semantic_gat with a directory for each random runs (seed).

python semantic.py

• SemMol: SemMol model can be trained by running SemMol.py. The results will be written in results/semantic_and_graph with a directory for each random runs (seed).

python SemMol.py

XAI

Explainability with LayerSaliency can be obtained for positive chemicals by running the explain.ipynb notebook. The results will be written in results/gnn_xai with a directory for each random runs.

Evaluation

The evaluation.py script contains code for:

• compute pretrained embeddings for all the chemcials involved in the Tox21 assays considered, project them in 2D with t-SNE and colour them according to chemical and physical properties of the molecules (extracted from ComptoxAI or through puchem API);
• process the classification results by computing the mean classification metrics for each model and for each assay, to create a dataframe than can be used to compute the violin plot with the mean results computedall the assay and the heatmap with the single assay results.
• process the xai results, by thresholding the number of edges to keep, and create the images with the molecule graph and most important subgraph identified a specific compound in input.

python evaluate.py