CrossDDI: Cross-Source Evidence-Grounded Drug-Drug Interaction Verification

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

CrossDDI is a novel verification-first framework designed to improve the auditability of LLM-based drug–drug interaction (DDI) assessments by explicitly linking predictions to evidence. Unlike traditional retrieval-augmented generation (RAG) approaches, CrossDDI separates LLM-based evidence extraction from a deterministic, LLM-free arbitration process utilizing DrugBank and PubMed. This framework mandates that all positive DDI predictions are tied to explicit supporting evidence. Evaluated on 1,000 DDInter 2.0 pairs in a positive–unlabeled setting, CrossDDI achieved a recall of 0.576–0.593 for confirmed positives, with interaction prediction rates comparable to RAG, while significantly reducing cross-backbone variation from 0.066 to 0.018. Analysis revealed that literature evidence acquisition and attribution are key bottlenecks, as PubMed retrieval covered only 40.5% of confirmed positives, and Path B-only evidence proved less reliable than structured sources.

Key takeaway

For NLP Engineers developing drug-drug interaction (DDI) prediction systems, consider adopting a verification-first architecture like CrossDDI to enhance auditability and reduce prediction variability. You should prioritize integrating explicit evidence linking mechanisms to ensure all positive DDI predictions are grounded. Focus on expanding reliable structured evidence sources, as current literature retrieval may only cover a fraction of confirmed interactions, impacting overall system coverage and trustworthiness.

Key insights

CrossDDI improves DDI assessment auditability by requiring explicit evidence for LLM predictions through a verification-first framework.

Principles

• Verification-first architectures enhance traceability.
• Explicit evidence linking improves auditability.
• Structured evidence is more reliable than literature-only.

Method

CrossDDI separates LLM-based evidence extraction from an LLM-free arbitration step. It uses DrugBank and PubMed to verify DDI predictions, requiring explicit supporting evidence for positive outcomes.

In practice

• Implement LLM-free arbitration for critical predictions.
• Prioritize structured data over unstructured literature.
• Evaluate evidence coverage for RAG systems.

Topics

• Drug-Drug Interaction
• LLM Verification
• Retrieval-Augmented Generation
• Evidence-Grounded AI
• DrugBank
• PubMed
• AI Auditability

Best for: AI Scientist, Research Scientist, NLP Engineer

Related on AIssential

• A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug-Target Interaction Prediction — TriMod-DTI uses triple-modal contrastive learning with 1D, 2D, and 3D features for superior drug-target interaction prediction.
• T2D-Bench: Evidence-Gated Evaluation of LLM Outputs for Type 2 Diabetes Using a Multi-Layer Clinical-Lifestyle Knowledge Graph — LLMs require evidence-gated evaluation against clinical knowledge graphs to ensure medical accuracy and guideline compliance.
• ‘An AlphaFold 4’ – scientists marvel at DeepMind drug spin-off’s exclusive new AI — Isomorphic Labs' proprietary IsoDDE model significantly advances AI drug discovery, outperforming open-source alternatives in binding affinity prediction.
• E-MRL: Cross-view Aligned Evidence-driven Multimodal Reinforcement Learning for Reliable 3D Tumor Analysis — Grounding Vision-Language Models in specific visual evidence via reinforcement learning significantly reduces hallucinations in medical report generation.
• DeepER-Med: Advancing Deep Evidence-Based Research in Medicine Through Agentic AI — DeepER-Med enhances medical research trustworthiness via agentic AI, explicit evidence appraisal, and a structured workflow.
• GraphDiffMed: Knowledge-Constrained Differential Attention with Pharmacological Graph Priors for Medication Recommendation — Combining dual-scale differential attention with pharmacological graph priors significantly enhances medication recommendation safety and quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.