Multilingual translation for zero-shot biomedical classification using BioTranslator

Hanwen Xu, Addie Woicik, Hoifung Poon, Russ B. Altman, Sheng WangNature Communications 2023 · DOI

Abstract

Existing annotation paradigms rely on controlled vocabularies, where each data instance is classified into one term from a predefined set of controlled vocabularies. This paradigm restricts the analysis to concepts that are known and well-characterized. Here, we present the novel multilingual translation method BioTranslator to address this problem. BioTranslator takes a user-written textual description of a new concept and then translates this description to a non-text biological data instance. The key idea of BioTranslator is to develop a multilingual translation framework, where multiple modalities of biological data are all translated to text. We demonstrate how BioTranslator enables the identification of novel cell types using only a textual description and how BioTranslator can be further generalized to protein function prediction and drug target identification. Our tool frees scientists from limiting their analyses within predefined controlled vocabularies, enabling them to interact with biological data using free text.

Introduction

In contemporary biology, data organization typically relies on "controlled vocabularies" (CVs), such as the Gene Ontology (GO) or Cell Ontology. This hierarchical structure acts like a rigid directory where every biological entity must be assigned to a predefined category. While effective for cataloging established knowledge, this model struggles with the rapid pace of discovery. When researchers encounter a novel cell type or a protein function that has not yet been officially categorized, traditional tools—which are often "blind" to anything outside their training set—fail to provide accurate annotations. Creating new CV terms is a laborious, manual process requiring significant domain expertise. BioTranslator aims to transition from this "directory" model to a "search engine" model, allowing scientists to describe new biological concepts in natural language and automatically link them to corresponding non-text data.

The BioTranslator Framework

BioTranslator is a multilingual translation framework designed to map diverse biological modalities—including gene expression, protein sequences, and drug molecules—into a shared latent space anchored by natural language.

The BioTranslator framework translates user-written text into non-text biological data, such as gene expression, protein sequences, and pathways, using a shared embedding space.
  1. Text Encoding: The framework utilizes PubMedBERT, a language model pre-trained on a vast corpus of biomedical literature. The model was fine-tuned on over 2 million textual descriptions from 225 biomedical ontologies using contrastive learning. This ensures that semantically related concepts are positioned closely within the embedding space.
  2. Modality Projection: Non-textual data types are processed through specialized deep neural networks and projected into the same shared embedding space.
  3. Zero-Shot Classification: By aligning all data types within a single space, BioTranslator can perform zero-shot classification. A researcher can input a novel text description, and the system identifies the "nearest neighbors" among the biological data, enabling the classification of entities the model has never explicitly seen during training.

Semantic Similarity as a Biological Bridge

A fundamental discovery of this work is that natural language serves as a highly flexible and effective bridge for biological data. The authors demonstrated that text-based similarity between ontology terms strongly correlates with established biological metrics, such as functional similarity and graph distance in ontologies. This suggests that the nuanced descriptions written by human experts contain sufficient signal to augment or even replace rigid hierarchical structures. Furthermore, the framework is bidirectional; it can not only find data from text but also generate biologically meaningful descriptions for groups of genes, enhancing the interpretability of complex datasets.

Validation of the semantic bridge: Text-based similarity scores correlate strongly with annotation-based similarity, graph distance, and pathway similarity across various biological domains.

Key Results and Performance

The researchers validated BioTranslator across multiple domains, including cell biology, proteomics, and drug discovery, using massive datasets like Tabula Muris and SwissProt.

  • Novel Cell Type Identification: BioTranslator achieved an average AUROC of 0.90 in identifying unseen cell types based solely on short textual descriptions, significantly outperforming traditional classifiers like ACTINN.
  • Protein Function Prediction: In "few-shot" scenarios where functions have minimal known examples, BioTranslator substantially outperformed DeepGOPlus, the current state-of-the-art in protein function prediction.
  • Cross-Modal Discovery: The model successfully predicted drug-target interactions and phenotype-pathway associations without requiring paired data for those specific modalities during training, relying instead on their shared relationship with the text space.

Discussion and Future Directions

BioTranslator represents a significant step toward more flexible biological data analysis. While currently dependent on high-quality expert descriptions for training, future iterations aim to incorporate text normalization to handle more informal lab notes. By enabling a "Google Search" for the biological world, BioTranslator empowers scientists to move beyond predefined vocabularies and accelerate the discovery and classification of novel biological entities.

Summary updated

Abstract

Existing annotation paradigms rely on controlled vocabularies, where each data instance is classified into one term from a predefined set of controlled vocabularies. This paradigm restricts the analysis to concepts that are known and well-characterized. Here, we present the novel multilingual translation method BioTranslator to address this problem. BioTranslator takes a user-written textual description of a new concept and then translates this description to a non-text biological data instance. The key idea of BioTranslator is to develop a multilingual translation framework, where multiple modalities of biological data are all translated to text. We demonstrate how BioTranslator enables the identification of novel cell types using only a textual description and how BioTranslator can be further generalized to protein function prediction and drug target identification. Our tool frees scientists from limiting their analyses within predefined controlled vocabularies, enabling them to interact with biological data using free text.
BibTeX
@inproceedings{xu2023multilingual,
  author = {Hanwen Xu and Addie Woicik and Hoifung Poon and Russ B. Altman and Sheng Wang},
  title = {{Multilingual translation for zero-shot biomedical classification using BioTranslator}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-36476-2},
}
Figure