Week 1, July 2025: Computational Tools for Antibody-Drug Conjugate (ADC) Development

Despite decades of therapeutic innovation, cancer remains the second leading cause of death globally, accounting for approximately 10 million deaths annually. While conventional chemotherapeutics remain a mainstay of treatment, their nonspecific cytotoxicity often results in systemic toxicity, off-target effects, and the development of resistance. Antibody-drug conjugates (ADCs) combine the specificity of monoclonal antibodies (mAbs) with the potency of cytotoxic agents to enable targeted cancer cell eradication. There are 15 FDA-approved agents as of November 2024 and many more in development. Yet, the complexity of ADC architecture, including the need for precise target selection, optimized conjugation chemistry, and payload release kinetics, poses persistent challenges to rational design and efficient development. In response, researchers are looking into the integration of computational tools, from machine learning-driven antibody engineering to generative linker design and in silico pharmacokinetic modeling.

Artificial intelligence (AI) and machine learning (ML) are transforming the design and development of antibody-drug conjugates (ADCs), replacing empirical trial-and-error with predictive, data-driven strategies. Traditional approaches have been limited by incomplete structural information, inefficient linker-payload selection, and low-throughput experimental methods. In contrast, modern AI frameworks enable modeling of ADC components: predicting antibody structures, identifying optimal conjugation sites, and forecasting pharmacokinetics and toxicity profiles. Tools such as AlphaFold3 and DeepAb support high-resolution modeling of antibody-antigen complexes, glycosylation patterns, and linker conformations. Generative AI models, including GANs, VAEs, and diffusion-based architectures, now allow the de novo design of antibody variants and linkers tailored to specific functional requirements. These advances are further integrated with high-performance computing and experimental validation pipelines, enabling rapid iteration and optimization. 

Building on the advances in AI-guided design, the integration of physics-based computational modeling offers complementary insight into the structural complexities that define antibody-drug conjugates (ADCs). Despite clinical success, ADC development remains constrained by a persistent lack of high-resolution structural data, particularly for full-length antibodies, flexible linkers, and payload-conjugated assemblies. This structural opacity impedes rational design, making it difficult to predict conjugation site accessibility, linker strain, and payload orientation. Recent work has highlighted the critical role of 3D molecular modeling and dynamics simulations in addressing these challenges. These methods enable the visualization and prediction of conformational flexibility, solvent exposure, and intermolecular interactions across ADC components. Machine learning tools, when combined with physics-based simulations, further enhance the ability to identify optimal linker types, conjugation strategies, and drug-to-antibody ratios. By refining experimental hypotheses and reducing the number of necessary design iterations, these hybrid computational approaches bridge the gap between empirical assays and predictive design, supporting more efficient and rational ADC development. 

Expanding on the integration of structural and physics-based modeling, recent efforts have introduced deep learning (DL) architectures specifically tailored to decode the complex structure–activity relationships in antibody-drug conjugates (ADCs). Researchers have developed ADCNet, a unified neural network framework designed to predict the anticancer activity of ADCs by jointly modeling protein sequences, chemical structures, and drug–antibody ratio (DAR) values. By leveraging advanced representation learning from models such as ESM-2 and FG-BERT, ADCNet effectively captures the molecular features of antibodies, antigens, linkers, and payloads. Trained on a curated benchmark dataset and validated through cross-validation and external testing, ADCNet demonstrated high predictive performance across all evaluation metrics. Its robust architecture enables generalization to unseen ADCs, offering a valuable tool for early-stage candidate prioritization. Crucially, the accompanying web-based platform, DeepADC, provides open access to this predictive capability, facilitating broader adoption in the community.  

Meanwhile, researchers are also looking into the optimization of chemical linkers, which govern payload stability, release kinetics, and therapeutic index. Traditional linker types, cleavable or non-cleavable, remain limited in diversity and tunability. A new example is Linker-GPT, a transformer-based architecture trained on large-scale molecular datasets and fine-tuned for ADC-specific linker generation. Leveraging reinforcement learning (RL), the model iteratively optimizes drug-likeness and synthetic accessibility, producing structurally diverse linkers that surpass commercial analogs in key in silico metrics. Notably, the model-generated cleavable and non-cleavable linkers exhibit favorable QED and SAS scores, demonstrating its potential to expand the chemical design space for ADCs. However, the study remains computational; translating these in silico gains into clinical relevance will require experimental validation of stability, biological activity, and pharmacological performance. Future iterations of Linker-GPT aim to incorporate clinical performance metrics, plasma stability data, and immunogenicity predictors into RL reward functions.