Computational Pharmacokinetic Profiling via Hybrid Multimodal Using Deep Learning Architectures

Abstract

Accurate Prediction of Pharmacokinetic behavior Plays a critical role in early drug discovery, where rapid screening of Molecular candidates can significantly reduce Laboratory workload and development cost. Traditional ADME Modeling Approaches rely heavily on handcrafted descriptors or Rule-Based filters, making them limited in Generalization and often Unsuitable for structurally diverse compounds. Recent Advances in Deep Learning have introduced architectures capable of learning directly from molecular structure, yet single-Modality Models still struggle to capture the full spectrum of chemical information. This paper presents a Hybrid Multimodal Deep Learning framework that integrates three complementary molecular Representations: Engineered chemical descriptors, Atom-Bond Graph Structures processed through graph neural networks (GNNS) and smiles sequences encoded using Transformer architectures Such as chemberta. By combining these modalities through a unified fusion network the system capture Global physicochemical trends, local structural patterns, and contextual sequence information simultaneously. The Framework includes explain ability modules based on snap values and graph-Attribution Techniques to Assist chemists in understanding model decisions. A Stream Lit-Based Interface and API enable practical development for Real-time ADME prediction. Experimental Evaluation on Publicly Available datasets demonstrates that the Multimodal Architecture consistently outperforms descriptors-only,GNN-only, and transformer only baselines across multiple ADME endpoints . The Results Highlight the potential of Multimodal Learning as a Robust, Interpretable, and User-Friendly Computational toll for early Pharmacokinetic Profiling.