Inferring transporter specificity through language model–guided graph learning

Membrane transporters orchestrate chemical exchange across cells, playing a central role in cellular metabolism and serving as prime targets for disease intervention. Yet, the specific substrates of most transport proteins remain unknown, largely because experimental characterization is hindered by the instability of membrane proteins in experimental assays. Existing computational tools are often limited to narrow substrate classes or fail to generalize to distant homologs. Here, we present TraSPIN, a geometric deep learning framework that jointly models paired representations of transporter proteins and small-molecule substrates. TraSPIN consistently outperforms existing predictors, demonstrating excellent generalization to both novel substrates and sequence-dissimilar transporters. The model moves beyond mere prediction to reveal the underlying structural mechanisms and motifs governing transporter-substrate recognition. Through large-scale inference across diverse species and successful de-orphanization of uncharacterized solute carrier proteins, we highlight TraSPIN’s broad applicability. This powerful framework establishes a robust foundation for deciphering membrane transport, with immediate implications for drug discovery and systems medicine.