A Neuro-Symbolic AI Approach to Identifying Potent DPP-4 Inhibitors for Diabetes Treatment

EasyChair Preprint 12394

Abstract

Diabetes Mellitus (DM) is the most widespread category within metabolic disorders and finding a potential therapeutic Dipeptidyl peptidase-4 (DPP-4) inhibitor agent is crucial. This study aims to uncover the efficacy of DPP-4 inhibitors utilizing a Neuro-symbolic approach, a new branch of artificial intelligence, and RoBERTa (NLP-transformermodel). We employ the LTN (Logical Tensor Networks), a novel machine learning technique, procuring data from ChEMBL and BindingDB databases. After curation, each database consists of 3918 and 3285 for the classification task. We experimented with 14 molecular feature extraction approaches, including descriptors fingerprints such as AtomPairs2DCount, AtomPairs2D, EState CDKextended, CDK, CDKgraphonly, KlekotaRoth, KlekotaRothCount, MACCS, Substructure, PubChem, SubstructureCount, PubChemPy, Lipinski’s Rule (RDKit). The LTN model yields a groundbreaking Accuracy incorporating an CDKextended fingerprint of 0.978, an F1-score of 0.978, an ROC AUC of 0.966, and an MCC of 0.931. Conversely, RoBERTa resulted in 0.9493 Accuracy, F1 score of 0.9491, ROC AUC 0.9174, and MCC 0.8423. Our findings show that integrating the neuro-symbolic strategy (neural network-based learning and symbolic reasoning) has immense potential to discover the drugs that have the potential to inhibit diabetes mellitus and classify biological activities that inhibit it. Overall, the LTN model exhibits interpretable reasoning and learning, which enables the discovery of novel insights into type 2 diabetes mellitus inhibitors.

Keyphrases: Diabetes Mellitus (DM) Drug discovery and AI, Dipeptidyl peptidase-4 (DPP-4), Neuro-symbolic AI

@booklet{EasyChair:12394,
  author    = {Delower Hossain and Ehsan Saghapour and Jake Chen},
  title     = {A Neuro-Symbolic AI Approach to Identifying Potent DPP-4 Inhibitors for Diabetes Treatment},
  howpublished = {EasyChair Preprint 12394},
  year      = {EasyChair, 2024}}