Multilingual translation for zero-shot bio-medical classification using BioTranslator

Existing annotation paradigms for biological data rely on controlled vocabularies, limiting analysis to known concepts. This paper introduces BioTranslator, a novel multilingual translation method addressing this limitation. BioTranslator translates user-provided textual descriptions of new concepts into non-textual biological data instances. It achieves this by developing a multilingual translation framework mapping various biological data modalities to a shared text-based latent space. The study demonstrates BioTranslator's application in identifying novel cell types, protein function prediction, and drug target identification, freeing scientists from the constraints of predefined vocabularies and enabling free-text interaction with biological data.

Hanwen Xu, Addie Woicik, Hoifung Poon, Russ B. Altman, Sheng Wang