Recent advances in Machine Learning have transformed antibody development through in silico models, accelerating therapeutic candidate identification. However, challenges persist: rapid adaptation of property predictors to laboratory-specific assays with incomplete datasets; batch effects introducing systematic bias; assay costs necessitating efficient unseen property prediction. We introduce a novel multimodal architecture featuring specialized tokenization and embedding projection that integrates text and protein language models (pLM) and a learning strategy to enable context-conditioned multi-property prediction without learning shortcuts. Our framework enables prompting without dictionary merging across modalities, creating a compact model capable of context-conditioned learning for multi-property prediction. The orchestrating model avoids pLM-to-text projection while enabling inference-time adaptation without retraining. Using 876,898 antibody heavy chain sequences with batch effect simulation, our architecture achieved Spearman's ρ > 0.8 across multiple developability properties, significantly outperforming fine-tuned multimodal LLMs and showed the ability to leverage correlation between properties for prediction. This approach has the potential to address critical antibody development challenges.
Giancardo et al. (Sat,) studied this question.