Incorporating contextual information into KGWAS for interpretable GWAS discovery

Cheng Jiang; Brady Ryan; Megan Crow; Kipper Fletez-Brant; Kashish Doshi; Sandra Melo Carlos; Kexin Huang; Burkhard Hoeckendorf

cs.LGMar 26, 2026

Incorporating contextual information into KGWAS for interpretable GWAS discovery

Authors: Cheng Jiang, Brady Ryan, Megan Crow, Kipper Fletez-Brant, Kashish Doshi, Sandra Melo Carlos, Kexin Huang, Burkhard Hoeckendorf,

Organizations: University of Michigan · gRED, Genentech · Stanford University

Abstract

Genome-Wide Association Studies (GWAS) identify associations between genetic variants and disease; however, moving beyond associations to causal mechanisms is critical for therapeutic target prioritization. The recently proposed Knowledge Graph GWAS (KGWAS) framework addresses this challenge by linking genetic variants to downstream gene-gene interactions via a knowledge graph (KG), thereby improving detection power and providing mechanistic insights. However, the original KGWAS implementation relies on a large general-purpose KG, which can introduce spurious correlations. We hypothesize that cell-type specific KGs from disease-relevant cell types will better support disease mechanism discovery. Here, we show that the general-purpose KG in KGWAS can be substantially pruned with no loss of statistical power on downstream tasks, and that performance further improves by incorporating gene-gene relationships derived from perturb-seq data. Importantly, using a sparse, context-specific KG from direct perturb-seq evidence yields more consistent and biologically robust disease-critical networks.

Figures & tables

Figure 1: Knowledge graph construction. A. The original KG in KGWAS consisting of variant-gene, gene-gene and gene-program edges; B. Our extension to KGWAS: (left) removing gene-program edges, and sparsifying remaining connections; (right) replacing gene-gene edges with contextually relevant relationships derived from Perturb-seq.

Figure 2: Distribution of cosine similarities between pairs of target genes annotated in STRING (positive) and randomly selected pairs of target genes (negative).

Figure 3: Contributions of different nodes and edge types in the KGWAS knowledge graph using a sample size of 10,000. Reported metrics are the total number of recalled independent loci summed across three selected traits. Standard deviations are computed across three training runs with different random seeds. In each subplot, bold indicates the model selected as the baseline in the subsequent panel. (A) Ablation of different relation types; (B) Ablation of V2G and G2G edge types; (C) Collapsing V2G and G2G edge types; (D) Adding context-specific information.

		Small cohort sample size
	# Edges	1,000	2,500	5,000	7,500	10,000	50,000
GWAS	0	14	15	29	36	49	201
KGWAS	12.1M	27.00 (1.00)	24.67 (1.53)	37.67 (2.08)	55.00 (1.73)	62.33 (2.52)	221.00 (4.36)
+ Remove gene programs	12.1M	26.33 (1.53)	24.00 (2.65)	36.33 (1.53)	53.33 (2.31)	65.33 (2.08)	220.00 (4.36)
+ Prune edge types	3.4M	25.67 (3.06)	26.33 (2.52)	46.33 (2.52)	58.33 (2.52)	67.33 (3.51)	221.00 (2.00)
+ Collapse GG edges	2.3M	25.67 (3.06)	26.00 (2.00)	46.33 (0.58)	55.00 (2.65)	66.67 (1.53)	218.67 (2.52)

Table 1: Top 100 independent loci replicated in large cohort GWAS. All values are summed across three selected traits: MCH, IRF, RDW. The top and bottom groups show models without and with contextual information, respectively.

		Small cohort sample size
	# Edges	1,000	2,500	5,000	7,500	10,000	50,000
Context-aware KGWAS ( $\star$ )	625K	30.67 (2.31)	34.67 (4.04)	53.00 (2.00)	58.33 (2.08)	79.67 (2.08)	229.00 (3.61)
$\star$ with randomized G2G	863K	24.67 (2.31)	27.33 (1.53)	47.33 (2.08)	54.33 (1.15)	67.33 (4.73)	221.33 (3.06)
$\star$ with dropped G2G	747K	27.67 (2.52)	29.67 (0.58)	38.33 (2.31)	51.00 (1.73)	72.00 (2.00)	224.33 (4.73)
$\star$ with randomized V2G and G2G	863K	19.67 (1.15)	26.33 (4.93)	43.00 (12.74)	50.67 (1.52)	61.67 (5.69)	211.00 (3.00)

Table 2: Ablation stuides on edge contributions to context-aware KGWAS. Top 100 independent loci replicated in large cohort GWAS over MCH, IRF, RDW traits are reported.

Figure 4: Consistency of disease critical networks for the rs61759901 variant in KGWAS (left) and context-aware KGWAS (right). Each plot aggregate nodes and edges from three seeded models trained on the MCH trait. Genes that are not significant in the K562 cell line (n=6) are shown with a gray border.

Biomedical KGs

Biomedical knowledge graphs organize evidence about genes, diseases, drugs, pathways, and patients for computational discovery and clinical reasoning.

See all 71 papers

GIP-RAG: An Evidence-Grounded Retrieval-Augmented Framework for Interpretable Gene Interaction and Pathway Impact Analysis
Mar 19, 2026Fujian Jia, Jiwen Gu, Cheng Lu +5LLM Reasoning with Graphs Gene Regulatory Network Inference
Kanghua Juntai Biotech Co. Ltd., Room 1504, Building 7, Dongwang Jingyuan, No. 768 Jingwang Road, Kunshan, Suzhou, Jiangsu Province, China · Department of Radiation Oncology, Cancer Treatment Center, The Second Affiliated Hospital of Hainan Medical University, Haikou, China · Department of Nutrition, University of California, Davis, USA +1
Knowledge Graph Modulated Deep Learning for Limited-Sample Clinical Data Analysis
May 22, 2026Yuwei Xue, Sakib Mostafa, James Zou +5Graph Representation Learning Medical Diagnosis
Department of Radiation Oncology, Stanford University, Stanford, California, USA · Department of Electrical Engineering, Stanford University, Stanford, California, USA · Department of Biomedical Data Science, Stanford University School of Medicine, Stanford, California, USA +6
A phenotype-driven and evidence-governed framework for knowledge graph enrichment and hypotheses discovery in population data
Apr 18, 2026Adela Bâra, Simona-Vasilica OpreaScientific Hypothesis Generation Knowledge Graphs
Department of Economic Informatics and Cybernetics, Bucharest University of Economic Studies, Bucharest, Romania

Mar 19, 2026J/K move · Enter open · S save

GIP-RAG: An Evidence-Grounded Retrieval-Augmented Framework for Interpretable Gene Interaction and Pathway Impact Analysis

Fujian Jia, Jiwen Gu, Cheng Lu, Dezhi Zhao+4

Kanghua Juntai Biotech Co. Ltd., Room 1504, Building 7, Dongwang Jingyuan, No. 768 Jingwang Road, Kunshan, Suzhou, Jiangsu Province, China · Department of Radiation Oncology, Cancer Treatment Center, The Second Affiliated Hospital of Hainan Medical University, Haikou, China · Department of Nutrition, University of California, Davis, USA+1

LLM Reasoning with Graphs Gene Regulatory Network Inference Biomedical KGs

Open paper AskPDF arXiv

Understanding mechanistic relationships among genes and their impacts on biological pathways is essential for elucidating disease mechanisms and advancing precision medicine. Despite the availability of extensive molecular interaction and pathway data in public databases, integrating heterogeneous knowledge sources and enabling interpretable multi-step reasoning across biological networks remain challenging. We present GIP-RAG (Gene Interaction Prediction through Retrieval-Augmented Generation), a computational framework that combines biomedical knowledge graphs with large language models (LLMs) to infer and interpret gene interactions. The framework constructs a unified gene interaction knowledge graph by integrating curated data from KEGG, WikiPathways, SIGNOR, Pathway Commons, and PubChem. Given user-specified genes, a query-driven module retrieves relevant subgraphs, which are incorporated into structured prompts to guide LLM-based stepwise reasoning. This enables identification of direct and indirect regulatory relationships and generation of mechanistic explanations supported by biological evidence. Beyond pairwise interactions, GIP-RAG includes a pathway-level functional impact module that simulates propagation of gene perturbations through signaling networks and evaluates potential pathway state changes. Evaluation across diverse biological scenarios demonstrates that the framework generates consistent, interpretable, and evidence-supported insights into gene regulatory mechanisms. Overall, GIP-RAG provides a general and interpretable approach for integrating knowledge graphs with retrieval-augmented LLMs to support mechanistic reasoning in complex molecular systems.

Abstract

Biomedical KGs

Explore similar work

GIP-RAG: An Evidence-Grounded Retrieval-Augmented Framework for Interpretable Gene Interaction and Pathway Impact Analysis

Knowledge Graph Modulated Deep Learning for Limited-Sample Clinical Data Analysis

A phenotype-driven and evidence-governed framework for knowledge graph enrichment and hypotheses discovery in population data