(M1230-01-01) Generating Novel Linker Sequences in Antibody-Drug Conjugates with Diffusion Models

Purpose: Antibody-drug conjugates (ADCs) are precise drug delivery systems composed of an antibody, linker, and payload. Designing stable linkers remains a challenge due to their complex structure and susceptibility to premature release, which can cause systemic toxicity. The complex structures of many linkers means there are biological requirements to ensure stability, as well as requirements to account for reactive groups on cytotoxic drugs and derivative functional groups. This research aims to address these challenges by developing a diffusion-based generative model capable of producing novel linker sequences that are optimized for ADCs. The model encodes the entire ADC, including antibody sequences and payloads, to design linkers best adapted to these features. By leveraging curated datasets and advanced machine learning techniques, this model enables a systematic exploration of linker stability, selective payload release, and toxicity minimization.

Methods: A curated dataset of antibody-drug conjugates (ADCs) was obtained from ADCDB, encompassing linker, payload, and antibody components. Linkers and payloads, classified as small molecules, were represented using SMILES sequences, while antibodies, as proteins, were characterized by their heavy chain and light chain amino acid sequences. The dataset was standardized by removing duplicates and incomplete entries. In total, there were 352 complete ADCs with corresponding information, and 174 unique linker structures, which were split into training (80%), validation (10%), and testing (10%) sets to ensure robust model evaluation. For feature extraction, ESM-2, a transformer-based language model, was used to capture the structural and functional properties of antibodies. For linkers and payloads, a novel NLP-based approach was applied to analyze SMILES sequences as linguistic representations, isolating meaningful patterns and structural dependencies. The model architecture employed a diffusion-based generative framework designed to encode the entire ADC, including the antibody and payload, to optimize linker design. The diffusion model utilized an iterative denoising process. At each timestep, noise was added to the input embeddings, and the model was trained to predict and remove this noise, progressively generating a refined linker embedding. The final linker structure was decoded from embeddings into SMILES format using a neural network-based decoder that employed tokenization and context-sensitive reconstruction. To evaluate the generated linkers, tools such as ProTox-3.0 and SwissADME were used to assess toxicity profiles and pharmacokinetic properties, ensuring alignment with FDA-approved ADC characteristics.

Results: The diffusion model successfully generated novel linker sequences optimized for specific ADC components, including antibody sequences and payloads. Training and validation loss consistently decreased, indicating effective learning of the denoising process and context alignment. Noise and context embedding losses stabilized after an initial steep decline, reflecting the model’s ability to generalize from limited data while maintaining robust performance. Over 80% of substructures in the generated SMILES sequences were chemically reasonable, with fewer than 5 misplaced tokens per string on average, demonstrating high structural accuracy. Toxicity analysis using ProTox-3.0 revealed LD50 values comparable to FDA-approved ADC linkers, supporting favorable safety profiles. UMAP reduction analysis showed alignment between the generated linkers’ toxicity classes and their nearest neighbors, further validating their biological relevance. Pharmacokinetic assessments via SwissADME confirmed desirable solubility and stability properties, consistent with established linkers, highlighting their potential for clinical applications. However, it is noted that these values assess only the linkers and may vary when integrated into full ADCs, given the influence of the entire drug context. One promising candidate, Novel Linker 9, demonstrated a balance of moderate toxicity (LD50: 900 mg/kg), manageable synthetic accessibility (6.07), and acceptable bioavailability (0.17), outperforming its nearest neighbors. In contrast, CEN010-105, an existing ADC linker, exhibited high synthetic complexity, potentially limiting its practicality despite comparable toxicity. These results underscore the potential of the diffusion model to generate novel linkers with optimized properties for further refinement in ADC development.

Conclusion: This study demonstrates the potential of diffusion-based generative models to revolutionize the design of linker sequences for antibody-drug conjugates (ADCs). By encoding the entire ADC, including antibody sequences and payloads, the model effectively generates linkers that are structurally valid, pharmacologically stable, and adapted to the specific requirements of ADC components. The results highlight the model’s capacity to address key challenges in ADC development, such as optimizing linker stability, minimizing off-target toxicity, and enhancing overall therapeutic efficacy. Evaluation metrics, including chemical validity, toxicity profiles, and pharmacokinetic properties, affirmed the feasibility of using diffusion models for ADC linker generation. While challenges in SMILES string accuracy remain, iterative refinement strategies and advanced embedding techniques demonstrate promise for overcoming these limitations in future work. This research not only advances the computational methodology for ADC development but also provides a scalable framework for integrating machine learning into drug discovery. Future directions include expanding the dataset, improving model robustness, and applying the model to broader ADC design challenges. These advancements have the potential to accelerate preclinical and clinical testing, reducing costs and enabling more effective cancer therapies.

References: Sheyi, R., de la Torre, B. G., & Albericio, F. Linkers: An Assurance for Controlled Delivery of Antibody-Drug Conjugate. Pharmaceutics. 14, 396 (2022).
Shen, L. et al. ADCdb: the database of antibody-drug conjugates. Nucleic Acids Res., (2023).
Lin, Z. et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 379, 1123-1130 (2023).
Sharma, R., Saghapour, E., & Chen, J. Y. An NLP-based technique to extract meaningful features from drug SMILES. iScience. 27, 3 (2024)
Banerjee, P., Kemmler, E., Dunkel, M., & Preissner, R. ProTox 3.0: a webserver for the prediction of toxicity of chemicals. Nucleic Acids Res., (2024)
Daina, A., Michielin, O. & Zoete, V. SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Sci Rep 7, 42717 (2017).

Acknowledgements: There is no conflict of interest.


Top: Diffusion model training and validation results
Bottom: Decoder training and validation results


2D UMAP projection of linker embeddings with novel diffusion-model generated linkers


Table 1: Comparative Evaluation of Novel and Existing Linkers for ADC Development