Skip to content

DDIAgents: Mechanism-Conditioned Context Flow for Drug-Drug Interaction Prediction

🕒 Published (v1): 2026-06-30 03:19 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DDIAgents is a multi-agent framework for drug-drug interaction (DDI) prediction that replaces static knowledge routing with mechanism-conditioned, per-instance dynamic context flow: a planner agent selects which heterogeneous knowledge sources each expert agent receives, updated iteratively via conclusion-agent feedback. On DrugBank and TWOSIDES under realistic drug-split settings, it consistently outperforms feature-based, graph-based, LLM-based, and agent-based baselines.

Problem

Existing DDI prediction methods—whether feature-based, graph-based, or LLM-based—apply a fixed reasoning pattern and static evidence scope to all drug pairs. Different interaction mechanisms (pharmacokinetic, pharmacodynamic, chemically mediated) require different decisive evidence modalities, so static pipelines either dilute mechanism-critical signals with irrelevant context or miss the relevant modality entirely. Existing multi-agent systems compound this by using fixed context flow (1-to-1, 1-to-N, or N-to-N broadcast), which cannot adapt evidence routing to the governing mechanism of each drug pair.

Method

DDIAgents iterates up to \(T = 3\) rounds, each comprising three stages:

  1. Expert agent instantiation: A planner agent spawns \(m = 3\) expert agents with distinct roles. In iteration \(t = 1\), roles are fixed as Pharmacist, Pharmacokineticist, and Pharmacologist. In subsequent iterations, the planner analyzes the conclusion agent's feedback from \(t-1\) to identify unresolved gaps and instantiate new role profiles.

  2. Dynamic context flow (\(N(t) \to N(t)\)): The planner assigns each expert a mechanism-specific subset of three heterogeneous knowledge source types: (i) molecular structural knowledge (SMILES strings), (ii) biomedical relational knowledge (DDI training data, drug-target associations, side-effect profiles), and (iii) semantic contextual knowledge (LLM-generated pharmacological descriptions). A chemistry-oriented expert receives primarily structural evidence; a clinical expert receives relational and semantic cues. Allocation updates across iterations guided by conclusion-agent feedback.

  3. Analysis and decision: Each expert retrieves and filters evidence from its assigned sources, then produces a DDI prediction and structured rationale. A conclusion agent synthesizes these into a final prediction or, if uncertainty or conflict persists, generates iteration feedback specifying which new experts and knowledge sources should be prioritized next.

A traditional classifier pre-filters DDI candidates to reduce context overflow before the agent pipeline begins.

Key Contributions

  • Formulates DDI prediction as a mechanism-conditioned knowledge orchestration problem rather than single-pattern classification.
  • Introduces a dynamic context flow mechanism that routes per-expert, per-instance, per-iteration knowledge subsets (contrasted with static 1-to-1, 1-to-N, and broadcast N-to-N schemes).
  • Adaptive expert instantiation: agent roles are not fixed—they are revised based on conclusion-agent feedback about reasoning gaps.
  • Iterative feedback loop closing the planner → experts → conclusion → planner cycle.
  • Demonstrated interpretable agent-level mechanistic rationales alongside quantitative SOTA results.

Results

All results from Table 2 (mean ± std, 5 runs), S1 (known–new) and S2 (new–new) splits:

DrugBank (Accuracy / F1): - S1: DDIAgents 55.75 / 45.52 vs. best prior agent MDAgents 54.69 / 39.52; best non-agent TextDDI/DDI-GPT ~50.68 / 34.71 - S2: DDIAgents 38.21 / 18.17 vs. best prior agent AgentVerse 36.48 / 14.01; paper reports +13.77% F1 improvement in S2 over next-best agent (MDAgents 15.97)

TWOSIDES (Hit@5 / NDCG@5): - S1: DDIAgents 53.94 / 21.21 vs. MDAgents 51.42 / 20.07 - S2: DDIAgents 42.26 / 14.08 vs. AgentVerse 40.38 / 13.14; paper reports +7.15% NDCG@5 in S2 over next-best agent - Single LLM performs near worst (S1 DrugBank Acc 8.71), ruling out knowledge leakage as driver

Average inference time: 12.83 s/query (DrugBank), 15.66 s/query (TWOSIDES).

Limitations

  • Maximum iteration count (\(T = 3\)) and agent count (\(m = 3\)) are fixed hyperparameters; the framework's cost and ceiling under larger budgets are not evaluated.
  • S2 (new–new) performance remains substantially weaker than S1 across all methods, suggesting fundamental generalization difficulty for unseen drugs.
  • Backbone is Qwen2.5 series; robustness of the dynamic routing signal across different backbone LLMs is not established.
  • Knowledge source retrieval and filtering details are relegated to an appendix; sensitivity to retrieval quality is not ablated in the main text.
  • The paper is truncated; ablation studies on individual components (context flow alone vs. instantiation alone) may not be fully reported here.

Relevance to Harnesses / Meta-Harnesses

DDIAgents is a domain-specific agentic harness with an explicit meta-harness layer: the planner agent functions as an orchestrator that dynamically configures the sub-agent team and their information contexts at each iteration, while the conclusion agent functions as a feedback router that rewires the harness topology between rounds. The N(t)-to-N(t) dynamic context flow taxonomy—contrasted empirically with fixed 1-to-1 (Virtual Lab), 1-to-N (SciAgents), and N-to-N broadcast (Coscientist, MDAgents) patterns—provides a principled vocabulary for classifying harness knowledge-routing architectures. For researchers building general meta-harnesses, the core insight is that per-instance adaptive routing of context to sub-agents, driven by an intermediate synthesis signal, consistently outperforms any static wiring scheme regardless of how capable the individual agents are.