K-Paths: Reasoning over Graph Paths for Drug Repurposing and Drug Interaction Prediction

TL;DR

K-Paths introduces a training-free, model-agnostic framework to extract diverse, biologically meaningful multi-hop paths from large biomedical knowledge graphs. By augmenting the KG, retrieving $K=10$ diverse shortest paths with a diversity-aware Yen’s algorithm, and converting paths for LLM or GNN reasoning, it enables inductive, zero-shot and supervised reasoning about unobserved drug–drug and drug–disease interactions. The approach yields substantial zero-shot gains for LLMs, reduces KG size by up to 90% for GNNs while preserving performance, and provides interpretable reasoning chains that improve explainability. The work offers a scalable bridge between KGs and LLMs for biomedical discovery and releases code and retrieved paths as a benchmark for inductive reasoning.

Abstract

Biomedical knowledge graphs (KGs) encode rich, structured information critical for drug discovery tasks, but extracting meaningful insights from large-scale KGs remains challenging due to their complex structure. Existing biomedical subgraph retrieval methods are tailored for graph neural networks (GNNs), limiting compatibility with other paradigms, including large language models (LLMs). We introduce K-Paths, a model-agnostic retrieval framework that extracts structured, diverse, and biologically meaningful multi-hop paths from dense biomedical KGs. These paths enable the prediction of unobserved drug-drug and drug-disease interactions, including those involving entities not seen during training, thus supporting inductive reasoning. K-Paths is training-free and employs a diversity-aware adaptation of Yen's algorithm to extract the K shortest loopless paths between entities in a query, prioritizing biologically relevant and relationally diverse connections. These paths serve as concise, interpretable reasoning chains that can be directly integrated with LLMs or GNNs to improve generalization, accuracy, and enable explainable inference. Experiments on benchmark datasets show that K-Paths improves zero-shot reasoning across state-of-the-art LLMs. For instance, Tx-Gemma 27B improves by 19.8 and 4.0 F1 points on interaction severity prediction and drug repurposing tasks, respectively. Llama 70B achieves gains of 8.5 and 6.2 points on the same tasks. K-Paths also boosts the training efficiency of EmerGNN, a state-of-the-art GNN, by reducing the KG size by 90% while maintaining predictive performance. Beyond efficiency, K-Paths bridges the gap between KGs and LLMs, enabling scalable and explainable LLM-augmented scientific discovery. We release our code and the retrieved paths as a benchmark for inductive reasoning.

Figures (3)

• Figure 1: Schematic representation of the predictive framework for unobserved interactions. (a) Problem formulation: Given a network of observed interactions among known entities (drugs or diseases) and a broader biomedical network containing additional relationships between various entities (e.g., drugs, diseases, genes, etc.). The task is to predict unobserved interactions between a known entity and an emerging entity or two emerging entities. (b) Augmented KG: The observed interactions and biomedical network are integrated to create a richer representation for the task.
• Figure 2: K-Paths Overview. (1) Given a query about the effect of an entity ($u$) on another entity ($v$), (2) K-Paths extracts reasoning paths from an augmented KG connecting ($u$) and ($v$). (3) These paths are filtered for diversity and (4a) transformed into natural language descriptions for LLM inference. (4b) The retrieved paths can also be used to construct a subgraph, enabling GNNs to leverage more manageable information for training and prediction.
• Figure 3: Influence of path selection strategies on Llama 3.1 8B. Diverse paths are essential for performance improvement.