PaperlayerView exact revision on arXiv

Source-linked AI summary

DrugReason: Dynamic Multi-View Reasoning over Knowledge Graph and Language Evidence for Drug Repurposing

Zijie Liu, Hongxuan Li, Zhen Tan, Jinhao Duan, Baixiang Huang, Zunpeng Liu, Kai Shu, Tianlong Chen

arXiv:2609.06779v1cs.LGcs.AI

TL;DR

Drug-repurposing relations can depend on complex biological mechanisms and diverse evidence. DrugReason combines KG-derived evidence with LLM-generated mechanistic inference, routing views to specialized experts; experiments show strong performance across three benchmarks.

  • Problem

    Drug-involving relations can depend on complex biological mechanisms and diverse biomedical evidence, making prediction challenging.

    Source: §1 · p. 1

  • Method

    DrugReason combines KG-derived evidence and LLM-generated mechanistic rationales, adaptively routing views to specialized experts and sharing signals through cross-expert distillation.

    Source: §1 · p. 2

  • Results

    Across PharmaDB, DDInter, and DrugBank, DrugReason shows strong improvements over single-view and static-fusion baselines.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Heterogeneous, query-dependent evidence interpretation is a promising direction for reliable and interpretable drug repurposing.

    Source: §5 · p. 9

    DrugReason prioritizes candidate relations computationally; it does not experimentally validate therapeutic efficacy or safety.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Drug repurposing aims to identify new therapeutic uses for existing compounds and, compared with de novo drug discovery, offers a faster and more cost-effective path to clinical translation. However, the space of candidate drug-disease pairs is enormous and their underlying relationships often depend on complex multi-hop biological mechanisms, making it difficult to reliably predict which pairs represent true therapeutic relationships. Existing approaches tackle this from two directions: knowledge graph-based methods organize curated biomedical evidence into structured relational networks for grounded multi-hop reasoning, while LLM-based methods leverage pretrained knowledge to generate flexible mechanistic rationales. Yet neither is sufficient alone - KGs are confined to observed graph structure while LLMs lack factual grounding and risk hallucination. To address this gap, we propose DrugReason, a multi-view reasoning framework that integrates grounded KG reasoning with LLM-generated mechanistic inference for drug repurposing. DrugReason adaptively routes diverse reasoning paths to specialized experts conditioned on the query context, while a cross-expert distillation objective enables knowledge sharing without sacrificing expert specialization. Experiments on PharmaDB, DDInter, and DrugBank show that DrugReason improves average performance over strong single-view reasoning baselines and achieves competitive or superior results compared with graph-based alternatives, while providing interpretable routing-based predictions.

1 Introduction

Drug repurposing prediction requires reasoning over diverse, often multi-hop biological evidence, but knowledge graphs and language models each have distinct limits. DrugReason combines their evidence views, routes them to specialized experts, and reports gains over single-view reasoning baselines across three benchmarks.

  • Drug–disease and drug–drug relationships can depend on complex biological mechanisms and multi-hop connections rather than a single direct association.Source: §1 · pp. 1–2
  • KG methods offer grounded, interpretable paths but are limited by observed graph structure, while LLM rationales can be flexible yet unsupported without factual grounding.Static fusion may not fully use heterogeneous path-specific signals, motivating adaptive routing.Source: §1 · pp. 1–2
  • DrugReason improves average performance over strong single-view reasoning baselines on PharmaDB, DDInter, and DrugBank, with competitive or superior results against graph-based alternatives.Ablations indicate that both adaptive routing and cross-expert distillation contribute to the gains.Source: §1 · p. 2
  • DrugReason adaptively routes KG-derived and LLM-generated reasoning views to specialized experts and uses cross-expert distillation to share signals while preserving specialization.Routing distributions also provide evidence-level transparency into how views support predictions.Source: §1 · p. 2

2 Related Work

Prior work uses curated knowledge graphs for structured, interpretable drug-repurposing evidence and language models for broader mechanistic reasoning. The section contrasts these approaches with DrugReason's query-specific routing of heterogeneous evidence views.

Knowledge-Grounded Reasoning for Drug Discovery

  • KG path-based methods provide interpretable drug–disease mechanisms, while GNN representations remain bounded by observed graph structure.KG-to-LLM systems such as K-Paths, DrugReAlign, and graph2prompt verbalize paths, but most use KG evidence as static context.Source: §2 · p. 2

LLM Biomedical Reasoning

  • Biomedical language-model systems use domain knowledge, tools, or agent collaboration for therapeutic prediction, but generated rationales can remain unreliable without structured grounding and evidence-specific control.Source: §2 · pp. 2–3

Diverse LLM Reasoning Strategies

  • Chain-of-Thought and Tree-of-Thought prompts elicit sequential reasoning and alternative hypotheses, but typically treat generated paths uniformly.DrugReason instead treats KG-derived paths and LLM-generated rationales as heterogeneous views routed to specialized experts.Source: §2 · p. 3

3 Method

DRUGREASON predicts relation labels for drug–disease and drug–drug pairs by combining KG-derived evidence with LLM-generated mechanistic views. It routes each view to a specialized LLM expert, aggregates their predictions, and uses mutual distillation during training to share predictive signals while preserving specialization.

3.1 Problem Formulation and Overview

  • For drug–disease and drug–drug pairs, DRUGREASON predicts a relation label represented by a canonical label verbalizer.Each reasoning view is a serialized textual evidence instance generated from one reasoning strategy.Source: §3.1 · p. 3

3.2 Reasoning View Construction

  • KG-derived definitions and paths provide structured grounding, while training-time LLM-generated views supply mechanistic inference under four prompting strategies.The four strategies are chain-of-thought, tree-of-thought, forward–backward validation, and mechanism-focused reasoning; the teacher LLM is not queried at inference.Source: §3.2 · pp. 3–4

3.3 Adaptive Routing to Generative LLM Experts

  • A router assigns each serialized reasoning view to one of two generative LLM experts, whose routed label predictions are aggregated for the pair-level prediction.The implementation uses Llama and Qwen experts, and the router makes hard assignments using straight-through Gumbel–Softmax.Source: Figure 2 · p. 4, §3.3 · pp. 4–5

3.4 Cross-Expert Mutual Distillation

  • Symmetric mutual distillation exchanges temperature-scaled soft label distributions between the experts while routed supervised learning preserves their specialization.Both experts are evaluated on each view for distillation, but the supervised prediction loss remains routed.Source: §3.4 · p. 5

3.5 Training Objective and Inference

  • Training proceeds in two stages, adding mutual distillation in Stage 2; at inference, routed view-level predictions are aggregated into the final label.The mutual distillation loss and routing entropy regularizer are used only during training.Source: §3.5 · pp. 5–6

4 Experiments

Across drug–drug and drug–disease benchmarks, DRUGREASON achieves the best average performance across datasets and backbones, and is competitive with or better than graph-based alternatives. Ablations and routing analyses examine the contributions of routing, distillation, and heterogeneous evidence assignment.

4 Experiments

  • DRUGREASON achieves the best average performance across datasets and backbones; on DDInter, it raises Llama3.2/Qwen2.5 from no-reasoning’s 0.20%/3.34% to 60.00%/72.00%, and on PharmaDB improves 43.25%/45.63% to 68.25%/65.48%.Source: §4 · p. 6, Table 1 · p. 7
  • On DDInter, the best single prompt scores 34.00/23.36, while the best static fusion scores 20.40/8.61; static fusion also trails the best single prompt on PharmaDB.The controls include concatenation, majority vote, mean vote, and graph embeddings.Source: Table 2 · p. 7, §4 · p. 7
  • On DDInter, Qwen2.5 reaches 72.00% versus GraphTrans (71.00%); on PharmaDB, Llama3.2 reaches 68.25% versus GAT (67.11%), but results are backbone-dependent.Llama3.2 trails GraphTrans on DDInter, while Qwen2.5 falls slightly below GAT on PharmaDB.Source: Table 3 · p. 7, §4 · p. 7
  • Removing routing drops DDInter/Qwen2.5 accuracy from 72.00% to 12.80%; mutual distillation generally improves average robustness, and the full model has the best average performance for both experts.Distillation is not beneficial in every cell: removing it slightly raises PharmaDB/Llama3.2 accuracy from 68.25% to 68.65%.Source: Table 4 · p. 7, §4 · p. 8
  • Routing regularization weight λ1 has a larger performance impact than distillation weight λ2; Qwen2.5 drops sharply at λ1 = 0.05 before partly recovering, while Llama3.2 is relatively stable.Performance generally degrades at larger weight values.Source: Figure 3 · p. 8, §4 · p. 8
  • The router assigns non-trivial routing mass to both KG-derived and LLM-generated views, with distributions varying by dataset and backbone.The analysis is a diagnostic of non-collapsed routing behavior; the ablation results provide direct evidence that routing-based assignment affects prediction.Source: Figure 4 · p. 8, §4 · p. 8

5 Conclusion

DRUGREASON integrates KG grounding with training-time LLM-generated rationales, routing heterogeneous views to specialized experts and using mutual distillation to share predictive signals. Across DDInter, DrugBank, and PharmaDB, it achieves strong overall performance against reasoning, graph-based, and transformer baselines.

5 Conclusion

  • DRUGREASON routes heterogeneous reasoning views to two specialized generative experts and uses cross-expert mutual distillation to share predictive signals while preserving specialization.Its views combine KG-derived factual grounding with training-time LLM-generated mechanistic rationales.Source: §5 · p. 8
  • DRUGREASON achieves strong overall performance against reasoning, graph-based, and transformer baselines, while single-view and static-fusion strategies are often unstable.The conclusion also identifies explicit reasoning evidence as critical.Source: §5 · pp. 8–9
  • The results point to heterogeneous, query-dependent evidence interpretation as a promising direction for reliable and interpretable drug repurposing.Source: §5 · p. 9

Limitations

DRUGREASON prioritizes candidate relations computationally; it does not experimentally validate therapeutic efficacy or safety. Its performance and cost are bounded by KG quality, possible unsupported LLM claims, and the use of two generative experts.

Limitations

  • DRUGREASON prioritizes candidate relations computationally rather than experimentally validating therapeutic efficacy or safety.Source: Limitations · p. 9
  • Missing or noisy biomedical KG relations can affect the reasoning views that DRUGREASON constructs.Source: Limitations · p. 9
  • LLM-generated views may contain unsupported mechanistic claims, although routing and mutual distillation help reduce reliance on any single view.Source: Limitations · p. 9
  • Using two generative LLM experts increases computational cost compared with single-model prompting or graph-only baselines.Source: Limitations · p. 9
  • Future work will explore more efficient routing, stronger automatic rationale verification, and broader prospective drug-repurposing evaluation.Source: Limitations · p. 9

A Prompt Templates and Input Filtering

The appendix documents prompt templates and deterministic filtering for reasoning views. Gold labels are excluded from model inputs, and teacher-generated rationales are used during training, not inference.

A Prompt Templates and Input Filtering

B Implementation Details

DRUGREASON training proceeds in two stages: Stage 1 optimizes routed supervised prediction with routing regularization, and Stage 2 adds cross-expert mutual distillation.

B Implementation Details

  • Stage 1 optimizes the routed supervised prediction objective together with routing regularization; Stage 2 retains that objective and adds cross-expert mutual distillation.Source: B Implementation Details · p. 12
  • λ1 controls routing-load regularization, while λ2 controls the mutual-distillation loss.Source: B Implementation Details · p. 12
  • Expert training uses lrs1 and lrs2 in Stages 1 and 2, respectively, while lrmisc is used for non-expert components such as the router.Source: B Implementation Details · p. 12

C Additional Control Results

The controls test whether DrugReason’s gains come from using more evidence or from KG structure alone. Static prompt controls compare single-view reasoning with input- and output-level fusion.

C Additional Control Results

  • The controls use the main experiments’ candidate pairs and evaluation protocol to test whether gains come merely from more evidence or KG structure alone.Source: C Additional Control Results · p. 12

C.1 Static Prompt-Fusion Controls

  • Single-prompt variants evaluate one reasoning view at a time, while concat-all combines all views in one prompt.Source: §C.1 · p. 12
  • Majority vote and mean vote fuse predictions at the output level across reasoning views.Source: §C.1 · p. 12

C.2 Knowledge Graph Embedding Controls

The section reports knowledge graph embedding baselines as controls for testing whether KG structure alone is sufficient without language-level mechanistic reasoning.

C.2 Knowledge Graph Embedding Controls

  • Table 9 reports knowledge graph embedding baselines that test KG structure alone without language-level mechanistic reasoning.Source: §C.2 · p. 12
  • The KGE results complement the compact KGE column in Table 2 of the main text.Source: §C.2 · p. 12

D LLM Usage

GPT-4o generated training-time mechanistic rationales, while LLMs also assisted with manuscript editing. The appendix describes prompt templates, training hyperparameters, and control results.

D LLM Usage

  • GPT-4o generated training-time mechanistic rationales, but the teacher model was not queried during inference.Source: D LLM Usage · p. 12
  • LLMs assisted with manuscript editing but were not used to generate experimental results or alter reported numerical findings.Source: D LLM Usage · p. 12
  • Prompt templates replace placeholders with candidate entities, biomedical definitions, or KG paths, and omit gold labels.Source: Table 5 · p. 13
  • The Llama3.2 and Qwen2.5 training hyperparameter tables identify λ1 as routing regularization and λ2 as mutual distillation.Source: Table 6 · p. 13, Table 7 · p. 13
  • Static prompt controls show that using more reasoning views without adaptive routing does not reliably improve performance.Source: Table 8 · p. 14
  • The KGE controls evaluate graph structure without language-level reasoning views or expert routing.Source: Table 9 · p. 14

Related papers

Loading 2609.06779v1…