Binding Affinity Prediction for Pancreatic Ductal Adenocarcinoma Using Drug-Target Descriptors and Artificial Intelligence

Pancreatic ductal adenocarcinoma (PDAC) is the most common and aggressive form of pancreatic cancer, accounting for 90% of all pancreatic malignancies. This study addresses the gap in disease-specific binding affinity prediction by integrating PDAC-derived targets with diverse molecular descriptors and artificial intelligence (AI) models, enabling more accurate therapeutic profiling. Initially, we constructed a drug library using compounds from the DepMap database, targeting proteins such as LIFR, BTG2, EPHX2, and PAK3 identified as differentially expressed genes in a previous PDAC study. We employed descriptors such as Conjoint Triad, amino acid composition (AAC), and Quasi sequence order to represent the targets. Similarly, the drugs were described by Morgan, RDKit, and PubChem descriptors. We used AI algorithms like random forest regressor (RFR), extreme gradient boost regressor (XGBR), and one-dimensional convolutional neural network (1D-CNN) to predict the binding affinity. We also employed two benchmark datasets, DAVIS and BindingDB, to compare our models’ performance in binding affinity prediction. We achieved a mean square error (MSE) value of 1.5 using Morgan-RDKit-PubChem-Conjoint descriptors and 1D-CNN on the PDAC dataset. Similarly, 1D-CNN with PubChem-AAC descriptors produced an MSE of 0.27 on the DAVIS dataset. Further, the XGBR model using the PubChem-AAC descriptors produced an MSE of 0.69 on BindingDB. Our study demonstrates the potential of an AI-driven framework as an effective and scalable solution for disease-specific drug-target interaction prediction, with promising implications for drug repurposing in PDAC.