HierHGT-DTI Improves Prediction of Interactions With Unseen Protein Targets in Benchmarks
Published in Artificial Intelligence, Drugs development & discovery, Journals, Tools.
A study published in BMC Bioinformatics on September 14, 2026 introduces HierHGT-DTI, a graph-based model for predicting drug–target interactions when a drug or protein has no interaction labels in the training data. New media coverage appeared on October 4.
Zhangben Chen and Yaping Wan represent molecules and proteins at several scales, from atoms and residues to larger structures. A heterogeneous graph transformer combines these relationships to predict candidate interactions.
The model led all six tested DrugBank and BioSNAP settings. For unseen proteins, its AUROC exceeded the strongest comparator by 8.8 percentage points on DrugBank and 5.8 on BioSNAP. Results were more mixed on BindingDB: it led one unseen-protein metric, while a comparator led random and unseen-drug evaluations.
These are benchmark results, not experimentally verified binding or clinical efficacy. The work may help prioritize laboratory investigations. The publisher identifies the available article as an accepted, citable early version subject to further edits.
Source: Chen and Wan, BMC Bioinformatics, September 14, 2026. DOI: 10.1186/s12859-026-06637-3.