TraSPIN: inferring transporter specificity through language model-guided graph learning

Membrane transporters mediate cellular chemical exchange, yet their substrate specificities remain poorly characterized, with experimental studies constrained by the instability of membrane proteins. Existing computational tools often target narrow substrate classes or fail to generalize to distant homologs. We present TraSPIN, a geometric deep learning framework that jointly models transporter proteins and small-molecule substrates. TraSPIN outperforms existing methods, generalizes to unseen substrates and sequence-dissimilar transporters, and reveals structural determinants and sequence motifs of transporter–substrate recognition. Cross-species inference and de-orphanization of uncharacterized solute carrier proteins further demonstrate its utility for investigating transporter specificity and accelerating functional annotation.