Research map / Large language models / Language, safety and interpretability
ExDDI: Explaining Drug-Drug Interaction Predictions with Natural Language
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ExDDI, a family of models that generate natural-language explanations (pharmacodynamic/pharmacokinetic mechanisms) for drug-drug interaction predictions from SMILES, using MolT5 fine-tuning, MACCS-based retrieval, and ChatGPT in-context prompting, evaluated on DDInter and DrugBank.
Problem
Existing DDI prediction models give binary or coarse DDI-type outputs without explanations, which limits user trust; the paper targets generating natural-language explanations of DDI mechanisms for both known and unseen drugs.
Approach
Formulates DDI explanation generation as text generation from drug SMILES and trains MolT5 with seq-to-seq, multi-task (generation plus bilinear classifier), and multi-task with classifier-constrained staged decoding; compares against fingerprint retrieval and in-context ChatGPT, and shows that training on detailed DDInter explanations also improves prediction.
Key results
- On DrugBank transductive, ExDDI-MT reaches BLEU 0.9447 and ExDDI-S2S ROUGE-2 0.9109, while ExDDI-IC (ChatGPT) scores BLEU 0.1187, showing poor performance of in-context prompting.
- In inductive S1 generation, ExDDI-MTS (DrugBank) gets ROUGE-L 0.4023 versus 0.3906 for retrieval, so fine-tuning has little advantage over retrieval on unseen drug pairs.
- On DDInter binary prediction, ExDDI-MTS reaches accuracy 0.9778 transductive and 0.8413 inductive S2, comparable to DSN-DDI (0.9694 transductive, 0.8192 S2 accuracy), but below DSN-DDI in inductive S1 (0.7115 vs 0.7342).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train fine-tuning models using detailed DDInter explanations rather than brief DrugBank explanations | ExDDI fine-tuning models · DDI binary classification, all settings | F1 | Helped |
| Add explanation-generation training to binary prediction (ablation) | ExDDI-MT and ExDDI-MTS · DDI binary classification on DDInter | F1: 0.9701{$\pm.0020$} -> 0.9783{$\pm.0006$} | Helped |
| Add explanation-generation training to binary prediction (ablation) | ExDDI-MT and ExDDI-MTS · DDI binary classification on DrugBank | F1: 0.9701{$\pm.0020$} -> 0.9749{$\pm.0054$} | Mixed |
| Use a multi-task classifier module instead of sequence-to-sequence generation alone (ablation) | ExDDI · DDI binary classification on DrugBank | F1: 0.9656{$\pm.0066$} -> 0.9749{$\pm.0054$} | Helped |
| Use multi-task training with staged inference instead of multi-task training (ablation) | ExDDI · DDI explanation generation | BLEU | No clear effect |
| Train a fine-tuning model rather than retrieve similar drug pairs | ExDDI · DDInter transductive | BLEU: 0.5701 -> 0.9392 | Helped |
| Train a fine-tuning model rather than retrieve similar drug pairs | ExDDI · DrugBank inductive S1 | BLEU: 0.2069 -> 0.2197 | No clear effect |
| Use DDInter explanation supervision rather than DrugBank explanation supervision | ExDDI fine-tuning models · DDI explanation generation | BLEU | No clear effect |
Surprising findings
- Models trained on DDInter's longer, richer explanations predict DDIs better than those trained on DrugBank's brief explanations (p<0.05).
- Fingerprint-based retrieval matches fine-tuned models when both query drugs are unseen.
Limitations
- Fine-tuned models generalise poorly to unseen drugs, likely due to linearised SMILES representation.
- General LLM (ChatGPT) performs very poorly given SMILES molecular representations.
- Multi-class DDI-type comparison is not like-for-like with prior work due to differing label sets (257 vs 65) and extra inputs (targets, enzymes) used by baselines.
Experimental setup
Datasets: DDInter DrugBank v5.1.10
Benchmarks: DDInter DrugBank
Metrics: BLEU ROUGE-1 ROUGE-2 ROUGE-L accuracy F1 precision recall
Baselines: GMPNN-CS DSN-DDI DDIMDL NMDADNN ExDDI-RV ExDDI-IC
Builds on: MolT5 MACCS keys DDInter
Closest papers by meaning
- MKG-FENN: A Multimodal Knowledge Graph Fused End-to-End Neural Network for Accurate Drug–Drug Interaction Prediction (AAAI 2024 · similarity 0.87)
- Blend the Separated: Mixture of Synergistic Experts for Data-Scarcity Drug-Target Interaction Prediction (AAAI 2025 · similarity 0.87)
- Dual-Channel Learning Framework for Drug-Drug Interaction Prediction via Relation-Aware Heterogeneous Graph Transformer (AAAI 2024 · similarity 0.87)
- Customized Subgraph Selection and Encoding for Drug-drug Interaction Prediction (NeurIPS 2024 · similarity 0.86)
- Motif-Oriented Representation Learning with Topology Refinement for Drug-Drug Interaction Prediction (AAAI 2025 · similarity 0.85)
- Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.