Zero-shot biomedical text classification requires accurate assignment of biomedical text (e.g., scientific abstract) to previously unseen labels (or concepts). Existing methods often struggle to generalize to novel concepts such as new diseases, drugs, and genes. To address these unique challenges, we propose a framework that combines feature disentanglement with contrastive learning to address this limitation. It separates each abstract into a content representation relevant for classification and a variance representation. This disentanglement ensures that the content features are invariant to the writing style, improving generalization to unseen labels. A contrastive learning strategy further structures the latent space by encouraging semantic clustering and separation of categories. Moreover, we model intra-class variance as a shared distribution across labels to enable variational data augmentation, enhancing robustness. The framework uses a domain-specific biomedical language model for feature extraction and fixed label anchors for semantic alignment. Extensive experiments conducted on the largest available biomedical corpus achieve superior performance on zero-shot multi-label classification tasks by learning discriminative and style-invariant representations.

Biomedical text classification refers to the task of annotating a biomedical text with its relevant labels from a candidate label set. Most of the existing approach operate in a fully supervised setting and thus heavily rely on human-annotated training data which is both labor-intensive and monetarily expensive. To address this, we propose to formulate biomedical text classification under the zero-shot learning (ZSL) paradigm that does not require any labeled training data and only relies on label surface names for training and inference. Specifically, we propose a new context-aware contrastive learning technique for ZSL that fully exploits the context information present in the biomedical text to generate semantically enriched feature representations needed for accurate zero-shot biomedical text classification. Unlike existing contrastive learning approaches that typically employ random text segmentation strategies to generate contrastive pairs, our approach utilizes the context information inherently present in biomedical text to generate semantically meaningful contrastive pairs. Extensive experiments on the largest available biomedical corpus validates the effectiveness of the proposed approach.