Uncertainty-aware deep learning for drug-target interaction prediction
Results in Engineering · 2026-12 · 32 · 113371
DOI: 10.1016/j.rineng.2026.113371 ↗Kurz gefasst
Fachartikel in der Zeitschrift Results in Engineering, 2026-12. Verfasst von Ruojia You und 2 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Science and Technology Council.
Abstract (Originalsprache)
Deep learning has accelerated the prediction of drug-target interactions (DTI), but existing models often produce overconfident predictions, limiting their reliability in high-throughput virtual screening. Evidential deep learning (EDL) can simultaneously output predictions and uncertainty estimates in a single forward pass, yet it remains unclear whether such uncertainty measures reliably distinguish correct from incorrect DTI predictions. Rather than proposing a new architecture, we used an existing evidential multimodal DTI framework (EviDTI), which combines protein language model embeddings with two-dimensional topological and three-dimensional geometric representations of drug molecules, serving as a testing platform to analyze its uncertainty behavior on the DrugBank dataset. Under optimized training conditions, the framework achieved an accuracy of 75.63% ± 2.99 and an AUPR of 0.8214 ± 0.0212. Uncertainty scores were higher for false positives than for correctly classified samples, and accuracy generally decreased as uncertainty increased, indicating that evidential uncertainty provides a useful relative ranking of prediction reliability. Ablation studies showed only minor performance drops when removing either 2D or 3D representations; furthermore, in cold split evaluation, uncertainty remained informative for unseen targets, but is close to uninformative on average for unseen drugs. These findings clarify both the practical value and current limitations of evidential uncertainty in multimodal DTI prediction.
Förderung
- National Science and Technology Council
Zitieren
APA
You, R., Le, N. Q. K., & Chua, M. C. H. (2026). Uncertainty-aware deep learning for drug-target interaction prediction. Results in Engineering, 32, 113371. https://doi.org/10.1016/j.rineng.2026.113371
Vancouver
You R, Le NQK, Chua MCH. Uncertainty-aware deep learning for drug-target interaction prediction. Results in Engineering. 2026;32:113371. doi:10.1016/j.rineng.2026.113371
BibTeX
@article{You2026,
author = {You, Ruojia and Le, Nguyen Quoc Khanh and Chua, Matthew Chin Heng},
title = {Uncertainty-aware deep learning for drug-target interaction prediction},
journal = {Results in Engineering},
year = {2026},
volume = {32},
pages = {113371},
doi = {10.1016/j.rineng.2026.113371}
}
EUVIMED