MultiADC: Advanced Antibody-Drug Conjugate Activity Prediction Through Multi-scale Feature Fusion

Targeted drug discovery is crucial for developing selective and effective anticancer therapies. Antibody–drug conjugates (ADCs) have emerged as a promising modality due to their high tumor specificity and reduced off-target toxicity. However, ADC discovery remains time-consuming and costly, and computational methods specifically designed for ADC activity prediction are still limited. Therefore, efficient and accurate computational strategies are urgently needed to improve screening efficiency and reduce development costs. In this paper, we propose MultiADC, a novel framework for ADC activity prediction that integrates global structural features with fine-grained component-specific features via a multi-scale fusion strategy. Fine-grained features of each component are extracted using a feedforward neural network with dimensionality reduction, while virtual graphs model inter-component interactions to capture global structural information. The fused multi-scale representations are then processed by a multilayer perceptron to predict ADC activity. Experimental results demonstrate that MultiADC consistently outperforms existing methods. Availability and implementation. https://github.com/TAI-Medical-Lab/MultiADC .