Protein foundation models, particularly protein language models, have shown strong success in learning meaningful protein representations using transformer architectures pretrained on large-scale datasets through self-supervised learning. These representations have proven effective for downstream tasks such as predicting protein functions and properties. However, most existing models focus solely on amino acid sequences, overlooking other informative modalities such as 3D structures and literature text. While some recent efforts incorporate multiple modalities, they often suffer from limitations in modality coverage or training strategy. To address this gap, we propose a multimodal pretraining framework that integrates three complementary modalities-protein sequences, structures, and literature text. Our method uses the sequence modality as an anchor and aligns the other two modalities to it via contrastive learning, enabling the model to capture richer and more holistic protein representations. Across a diverse set of downstream tasks, ProteinAligner outperforms state-of-the-art foundation models in predicting protein functions and properties.
Zhang et al. (Wed,) studied this question.