Context-aware multi-property antibody predictor: a novel framework integrating text and protein language models

Area of Science:

• Computational Biology
• Machine Learning in Drug Discovery
• Protein Engineering

Background:

• Machine learning (ML) accelerates antibody development via in silico models for therapeutic candidate identification.
• Challenges include adapting property predictors to specific assays with limited data, batch effects causing bias, and high assay costs.
• Efficient prediction of unseen properties is crucial for cost-effective antibody development.

Purpose of the Study:

• To introduce a novel multimodal architecture for context-conditioned multi-property prediction in antibody development.
• To enable efficient adaptation to laboratory-specific assays and mitigate batch effects without retraining.
• To improve the accuracy and efficiency of predicting antibody developability properties.

Main Methods:

• Developed a novel multimodal architecture with specialized tokenization and embedding projection.
• Integrated text and protein language models (pLM) with a context-conditioned learning strategy.
• Utilized 876,898 antibody heavy chain sequences, including simulated batch effects, for training and validation.

Main Results:

• Achieved Spearman's ρ > 0.8 across multiple developability properties, demonstrating high prediction accuracy.
• Significantly outperformed fine-tuned multimodal large language models (LLMs).
• Demonstrated the model's ability to leverage correlations between properties for enhanced prediction.

Conclusions:

• The proposed multimodal architecture effectively addresses key challenges in antibody development, including data scarcity and batch effects.
• The framework enables context-conditioned multi-property prediction and inference-time adaptation without retraining.
• This approach offers a powerful tool for accelerating the identification and optimization of therapeutic antibody candidates.