TooTranslator: Zero-Shot Classification of Specific Substrates for Transport Proteins by Language Embedding Alignment of Proteins and Chemicals

Abstract

Transmembrane transport proteins mediate selective movement of ions and metabolites across membranes. Experimental characterization of their substrate specificity is limited. For novel class discovery with limited data, zero-shot learning aims to assign labels to test samples whose label has not been seen previously during training of the model. We introduce TooTranslator, a regression-based model that aligns embeddings from pre-trained protein (ProtBERT), chemical (ChemBERTa), and text (SciB-ERT) language models into a shared latent space, enabling substrate prediction by minimizing distances between protein and substrate embeddings. TooTranslator tackles zero-shot learning for the task of predicting the specific substrate of transmembrane transport proteins. Models using four loss functions are evaluated on protein test sets with seen and unseen substrates. We find no statistically significant difference in performance of the four loss functions. For tests with seen substrates, the models compare with the state-of-the art in performance. For tests with unseen (but known) substrate a top-k approach is needed as only 4.3% of test cases predict the correct label as the nearest label. For top-10 that rises to 17%; for top-50 rises to 54%, and top-100 reaches 80%. TooTranslator demonstrates the potential of multimodal embedding alignment for open-world protein function inference.