Fig. 2: Schematic illustration of deep learning frameworks for in vivo disposition modeling of ADCs.

A Graph Neural Network (GNN) for Molecular-Level Pharmacokinetic (PK) Prediction of ADCs. The ADC molecule, comprising a monoclonal antibody, a linker, and a cytotoxic payload, is initially represented as a graph where atoms are nodes and chemical bonds are edges. This graph is then processed by a GNN core architecture, utilizing message-passing mechanisms among nodes (e.g., from neighboring nodes ①-④ to a central Target Node) to learn complex molecular features. The GNN subsequently outputs predicted key PK parameters, such as clearance (CL), volume of distribution (Vd), and tissue accumulation profiles, thereby facilitating in silico ADC design optimization. B 3D U-Net Convolutional Neural Network (CNN) for Preclinical Biodistribution Imaging. Low-count 3D PET/CT images, representing raw biodistribution data, are fed into a 3D U-Net model. The U-Net’s encoder path (blue blocks) performs sequential 3D convolution and max pooling for hierarchical feature extraction and down-sampling, while the decoder path (green blocks) utilizes 3D up-convolution for feature reconstruction and up-sampling. Skip connections (dashed lines) bridge corresponding resolution levels between the encoder and decoder paths, integrating high-resolution spatial information. The primary function of this model here is denoising these 3D volumetric data, yielding a high-quality denoised 3D PET/CT image to enable more accurate quantification of ADC distribution and identification of potential off-target accumulation.