NeSyDPP4-QSAR: Discovering DPP-4 Inhibitors for Diabetes Treatment with a Neuro-symbolic AI Approach

Delower Hossain¹Ehsan Saghapour²Jake Y Chen^1*

• ¹Department of Computer Science, College of Arts and Sciences, University of Alabama at Birmingham, Birmingham, United States
• ²School of Graduate Biomedical Sciences, The University of Alabama at Birmingham, Birmingham, Alabama, United States

Diabetes Mellitus (DM) is a global epidemic and among the top ten leading causes of mortality (WHO, 2019), projected to rank seventh by 2030. The US National Diabetes Statistics Report (2021) states that 38.4 million Americans have diabetes. Dipeptidyl Peptidase-4 (DPP-4) is an FDA-approved target for type 2 diabetes mellitus (T2DM) treatment. However, current DPP-4 inhibitors are associated with some adverse effects, including gastrointestinal issues, severe joint pain (FDA safety warning report), nasopharyngitis, hypersensitivity, and nausea. Identifying novel inhibitors is essential. Moreover, direct in vivo DPP-4 inhibition assessment is costly and impractical; conducting in-silico IC50 prediction is a viable alternative to assess the efficacy. Quantitative Structure-Activity Relationship (QSAR) modeling is a widely used computational approach for chemical substance assessment. To build DPP4 QSAR, we employ LTN, a neuro-symbolic approach, alongside DNN and Transformers as baselines to compare the developed model performance. After deduplication and thresholding, DPP-4 related data was sourced from PubChem, ChEMBL, BindingDB, and GTP, comprising 6,563 bioactivity records (SMILES-based compounds with IC50 values). A diverse set of features, including descriptors (CDK Extended-PaDEL), fingerprints (Morgan), chemical language model embeddings (ChemBERTa2), LLaMa 3.2 embedding features, and physicochemical properties, are used to train the NeSyDPP4-QSAR model. Our model yielded the highest Accuracy, incorporating CDKextended and Morgan fingerprints, with an accuracy of 0.9725, an F1-score of 0.9723, an ROC AUC of 0.9719, and an MCC of 0.9446. The performance was benchmarked against two standard baseline models: a deep neural network and a transformer, as well as prior SOTA models. We conducted an external evaluation using the DTC dataset to ensure fair comparisons and assess model robustness. NeSyDPP4 achieved a stark performance, with an accuracy score of 0.9579, a ROC-AUC of 0.9565, a Matthews correlation coefficient (MCC) of 0.9171, and an F1 score of 0.9577. Overall, findings suggest that integrating Neuro-symbolic strategy (neural network-based learning and symbolic reasoning) holds immense potential for discovering drugs that can inhibit diabetes mellitus and classifying biological activities that inhibit it.