Chebifier 2: An Ensemble for Chemistry

Chebifier 2: An Ensemble for Chemistry Simon Flügel (Osnabrück University) Based on: Simon Flügel, Martin Glauer, Janna Hastings, Till Mossakowski, Christopher J. Mungall, Charlotte Tumescheit and Fabian Neuhaus: Chebifier 2: An Ensemble for Chemistry (SEMANTiCS, 2025) Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 2 Ontological Classification tripeptide zwitterion dipeptide zwitterion peptide zwitterion dipolar compound chemical entity organic molecular entity nitrogen ylide ylide ? ??? Asp-Lys zwitterion (CHEBI:229953) Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 3 Chebifier 2 ELECTRA [1]: Transformer model GNN: Graph neural network ChemLog [2]: Logic-based model CCPs [3]: Chemical Classifier Programs (Generative AI-based) [1] Glauer, M., et al.: Chebifier: automating semantic classification in ChEBI to accelerate data-driven discovery. Digital Discovery 3(5), 896–907 (2024) [2] S. Flügel, M. Glauer, T. Mossakowski, F. Neuhaus, ChemLog: Making MSOL viable for ontological classification and learning, in: 5th International Joint Conference on Learning and Reasoning (accepted), preprint available on arXiv, 2025. URL: http://arxiv.org/abs/2507.13987. [3] C. J. Mungall, A. Malik, D. R. Korn, J. T. Reese, N. M. O’Boyle, J. Hastings, Chemical classification program synthesis using generative artificial intelligence, arXiv preprint arXiv:2505.18470 (2025) Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 4 Datasets for Deep Learning Extract label classes and molecules (=SMILES-annotated classes) from ChEBI Label classes are selected via threshold (min. 50 / 100 molecules) Optional: Filter for 3-STAR Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 5 ELECTRA 🤖 Deep Learning model (Transformer-based) Trained on SMILES strings –pretraining: PubChem –fine-tuning: ChEBI 61-10-10-10-10-53-735-10-7-54-19-8-52-761-8-10-10-7-54-19-867-8-10-7-54-19-8-67 Tokenise SMILES string Put SMILES into transformer Predict ontology class membership [NH3+]CCCC[C@H] (NC(=O)[C@@H] ([NH3+])CC(=O) [O-])C(=O)[O-] Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 6 Graph Neural Networks (GNNs) 🕸 2 architectures: Residual Gated Graph Convolutional Networks [1] and Graph Attention Networks [2] Graph representation of molecules + calculated properties Atom properties: Type, #bonds, charge, aromaticity, hybridisation, #Hs Bond properties: Type, inRing, aromaticity Global properties: RDKIt2DNormalized Calculate properties Initialise GNN Predict ontology class membership [NH3+]CCCC[C@H] (NC(=O)[C@@H] ([NH3+])CC(=O) [O-])C(=O)[O-] [1] X. Bresson, T. Laurent: Residual gated graph convnets, 2018. arxiv. URL: http://arxiv.org/abs/1711.07553 [2] Velickovic, Petar, et al.: Graph attention networks, 2019. ICLR 2018 Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 7 Augmented GNNs 🕸 3-layer architecture: atoms, functional groups, global node Figure from: Aditya Ganesh Khedekar: Integrating Chemical Knowledge into Graph Neural Networks, 2025. Master’s Thesis. Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 8 ChemLog 🧩 Rule-based classification with monadic second-order logic Covers 169 classes related to peptides or of the type “x molecular entity” / “organo-x compound” 128 classes are unique to ChemLog Classify with logical rules organooxygen Compound ⇔∃a,b:C ( a ) ∧O ( b ) ∧Bond ( a,b ) [NH3+]CCCC[C@H] (NC(=O)[C@@H] ([NH3+])CC(=O) [O-])C(=O)[O-] Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 9 Chemical Classifier Programs (CCPs) 📟 Given textual descriptions of ChEBI classes, use Large Language Models to generate Python scripts for classification Refine in an iterative process with samples from ChEBI Returns natural language explanations 338 classes covered (66 unique to CCPs) Generate programs [NH3+]CCCC[C@H](NC(=O) [C@@H]([NH3+])CC(=O) [O-])C(=O)[O-] amino-acid zwitterion: “The zwitterionic form of an amino acid having a negatively charged carboxyl group and a positively charged amino group.” def is_amino_acid_zwitterion(smiles: str): mol = Chem.MolFromSmiles(smiles) … return False, “Missing negatively charged carboxyl group” Classify with programs Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 16 Evaluation – Deep Learning 20 models (5 architectures x 4 datasets) Model selection strategies: Top-N, Reduce, Add Selected models (Reduce @ 4): GAT-aug (ChEBI50) RGGCN-aug (CHEBI50-3-STAR) GAT (ChEBI50) ELECTRA (ChEBI50-3-STAR) Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 17 Evaluation – Inconsistency Resolution Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 18 Evaluation – Adding ChemLog + C3Ps Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 19 Mismatches between ChEBI and ChemLog Example: methanol (CO, CHEBI:17790) is not classified as an organooxygen compound (CHEBI:36963, “containing at least one carbon-oxygen bond“) in ChEBI Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 20 https://myshare.uni-osnabrueck.de/f/b26e8848a4ae4acf9c28/ Live Demo C:\Users\sifluegel\Videos\Screen Recordings Chebifier 2: An Ensemble for Chemistry / Flügel et al. / 21 Conclusion No model is perfect and chemical classes are diverse Chebifier 2 combines logic + deep learning + generative AI Classand sample-specific weighting Total classes covered: 1,742 Explanations available for logic and generative AI models –Weak explanations available for transformers Accessible as a Python library or via web frontend chebifier.hastingslab.org github.com/ChEB-AI/python-chebifier Extra Slides