Austin Liu

Published in Bioinformatics · 2 July 2026

3DICE: Interpretable 3D Cross-Modal Learning for Drug–Target Interaction Prediction and Large-Scale Drug Discovery

Second Award · Computational Biology and Bioinformatics · Regeneron ISEF 2026 (opens in a new tab)

3DICE is a structure-aware deep learning framework for predicting drug–target interactions at scale. It is the first approach to jointly employ 3D molecular and protein representations together through co-attention, improving efficiency and generalisation while also focusing on interpretability.

Austin Zi Rui Liu Nguyen Quoc Khanh Le Matthew Chin Heng Chua
DOI: 10.1093/bioinformatics/btag488 (opens in a new tab)

Overview

  1. 01 · Challenge

    Moving beyond sequences

    Many DTI models rely mainly on drug SMILES and protein FASTA sequences, limiting their access to the 3D structural information that shapes molecular binding.

  2. 02 · Approach

    3D cross-modal learning

    3DICE uses Uni-Mol for drugs and ESM-IF1 for proteins, then applies co-attention to fuse their structure-aware embeddings and model intermolecular relationships.

  3. 03 · Contribution

    Performance with insight

    Across DrugBank and KIBA, 3DICE outperformed state-of-the-art models on multiple metrics while its attention maps consistently highlighted decision-critical atoms and residues.

Why it matters

Drug–target interaction prediction is a high-throughput screening problem. Better predictions can help researchers narrow enormous candidate spaces before committing to more expensive experimental or physics-based evaluation. 3DICE is designed for this large-scale setting, combining structural information, cold-start generalisation, and interpretable attention within one framework.

Citation

Liu, A. Z. R., Le, N. Q. K., & Chua, M. C. H. (2026). 3DICE: Interpretable 3D Cross-Modal Learning for Drug–Target Interaction Prediction and Large-Scale Drug Discovery. Bioinformatics, btag488.

https://doi.org/10.1093/bioinformatics/btag488 (opens in a new tab)