Prediction of gut-bacterial drug biotransformation

Michael Zimmermann Group Prediction of gut-bacterial drug biotransformation Mahnoor Zulfiqar1, Ting-Hao Kuo1, Eleonora Mastrorilli1, Mariia Beliaeva1, Petar Scepanovic2,Michael Zimmermann1* 1Molecular Systems Biology Unit, European Molecular Biology Laboratory, Heidelberg, Germany. 2F. Hoffman-La Roche, Roche Pharmaceutical Research and Early Development pRED, Basel, Switzerland @zimmermannlab.bsky.social Motivation • Create and share ahigh-quality, AI-ready and FAIR-compliant dataset of gut bacterial drug biotransformation • Develop machine learning models to predict bacterial biotransformation from reactive functional groups • Validate the prediction model via active learning loop and to integrate gut microbial biotransformation prediction in the early drug development stage Aim •Prioritize feature subsets linked to drug–enzyme interactions. •Explore graph-based models and product-derived features. •Incorporate bacterial strain–specific insights for better predictions. •Address data gaps by enriching ChEMBL with microbial biotransformation data. zedmahnoor [email protected] zmahnoor14 EMBL Heidelberg Meyerhofstraße 1 · 69117 Heidelberg · Germany www.embl.org/groups/zimmermann Outlook and Future work Highlights Data Representation 527 amides within drugs 385142 Radius=13 Amide Biotransformation Classifier Machine Learning Workflow GLPG1837 Drugs not biotransformed by bacteria Drugs biotransformed by bacteria UMAP of current dataset within DrugBank Iproniazid Levosulpiride Sulfinpyrazone ~ 8 bacteria Acecainide ~ 2 bacteria Calpeptin ~ 39 bacteria Example amide containing drug structures and bacterial biotransformation UMAP 1 UMAP 2 DrugBank Biotransformation = 0 Biotransformation = 1 UMAP of amide containing drugs, labelled biotransformation label and data splitting At radius 13 around the carbonyl C of the amide, the whole molecule is covered and gives the highest MCC score of 0.54 using pattern fingerprints as the model features Example drug coverage at radius 13 around the carbonyl carbon from one direction Carbonyl C13th atom (Substructure) Pattern Fingerprint based Model performance Amide containing drug distribution Model results for Morgan Fingerprints + Physicochemical properties Model results for Morgan Fingerprints ROC AUC Balanced accuracy MCC ROC AUC Balanced accuracy MCC Scores The low MCC scores (average scores from cross validation mentioned in the bar plot above) from Morgan fingerprints and physicochemical properties indicate that drug biotransformation cannot be reliably predicted using the full chemical structure or the complete set of physicochemical descriptors. MCC = 0.54 ExtraTreesClassifier Logistic Regression Random Forest Random Forest + SMOTE Support Vector Machine XGBoost XGBoost + SMOTE ExtraTreesClassifier Logistic Regression Random Forest Random Forest + SMOTE Support Vector Machine XGBoost XGBoost + SMOTE 0.70.67 0.69 0.68 0.69 0.7 0.69 0.57 0.59 0.53 0.6 0.57 0.51 0.59 0.2 0.2 0.11 0.22 0.2 0.23 0.05 0.2 0.57 0.11 0.52 0.2 0.25 0.04 0.04 0.57 0.5 0.5 0.6 0.6 0.69 0.68 0.63 0.68 0.68 0.67 0.69 Extra Tree Classifier