DMPKformer: An Interpretable Multimodal Deep Learning Framework for Reliable ADMET Property Prediction

Abstract

Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties remains a critical challenge in drug discovery. Traditional single modality approaches often fail to capture the complex, multi-scale relationships governing molecular behavior across physicochemical, structural, and pharmacokinetic dimensions. In this work, we propose a multi-modal deep learning framework that integrates complementary molecular representations, MACCS fingerprints, molecular graphs, and physicochemical descriptors to achieve robust ADMET property prediction. Each modality is modeled using a specialized neural subnetwork tailored to its structural characteristics: a self-attention-based Transformer encoder for MACCS fingerprints, a Graph Attention Network (GAT) for molecular graph representations, and a tanh-activated multilayer perceptron for RDKit-, PaDEL-, and Mordred-derived descriptors. Each modality is independently trained for binary classification, and latent embeddings extracted from internal layers serve as transferable molecular representations. These embeddings are subsequently fused and fine-tuned via a tanh-activated dense network and shared prediction head to form a unified ADMET predictor. The proposed framework achieves competitive performance across multiple TDC ADMET benchmarks while providing enhanced interpretability through modality-specific attention mechanisms. In addition, the incorporation of latent-space out-of-distribution (OOD) confidence estimation enables identification of high-confidence operating regions, improving the reliability and practical applicability of the framework for molecular property prediction in drug discovery workflows.