Aligning Biomedical Texts and Knowledge Graphs: A Systematic Comparison of Lightweight Alignment Strategies

Artem Bisliouk; Elizaveta Nosova; Heiko Paulheim; Andreea Iana; Rita T. Sousa

cs.CLAug 24, 2026

Aligning Biomedical Texts and Knowledge Graphs: A Systematic Comparison of Lightweight Alignment Strategies

Authors: Artem Bisliouk, Elizaveta Nosova, Heiko Paulheim, Andreea Iana, Rita T. Sousa

Organizations: University of Mannheim, Germany · Data and Web Science Group, University of Mannheim, Germany

Abstract

Biomedical knowledge exists in two complementary but distinct forms: unstructured scientific literature and structured knowledge graphs (KGs). Aligning them is essential for knowledge grounding, evidence retrieval, and KG completion, yet existing methods do not explicitly align free-text evidence with KG triples. We present a unified framework for systematically studying design choices for aligning biomedical text and KGs. With a text encoder and a KG embedding model both frozen, we learn only a lightweight projection between their spaces via a contrastive objective. This enables a fair comparison across six design dimensions: text encoder, KG embedding model, projection head, triple composition, training direction, and hard-negatives sampling. We construct CTD-Align, a corpus of over 22K one-to-one tripledocument pairs linking chemical-gene interactions from the Comparative Toxicogenomics Database to supporting PubMed passages. We evaluate alignment on it in two retrieval settings: document-to-triple and triple-to-document. We find that the triple composition and the training direction (i.e., shared retrieval space) have the greatest impact, whereas the text encoder and hard-negatives sampling matter little. Overall, simple choices win: projecting text into the KG space with a linear head over concatenated subject, predicate, and object embeddings performs best. These findings establish lightweight contrastive alignment as an effective, practical foundation for bridging biomedical text and KGs.

Figures & tables

Figure 1: Overview of our unified framework for biomedical text-KG alignment. Documents and KG triples expressing the same fact are embedded by frozen encoders; the only trained component is a lightweight projection aligning the two spaces under a contrastive objective. We vary six components, shown here as separate modules: the text encoder, KG embedding model, triple-composition operator, projection head, training direction, and hard-negative strategy. We evaluate alignment by retrieval in two settings: document-to-triple ( D2T ) and triple-to-document ( T2D ).

	# Relations	# Edges	# Nodes
	# Relations	# Edges	Chemicals	Genes	Diseases	Pathways	Total
Core	11	22,293	4,179	3,431	-	-	7,610
Enriched	14	1,048,611	13,773	50,709	7,292	2,363	74,137

Table 1: Statistics of the core and enriched CTD KGs.

Panel of Figure 2 — Figure 2: Distribution of relation types in the (a) core and (b) enriched CTD KGs. ”Others” combines 4 relation types from the enriched KG with fewer than 2,600 edges (0.44%): increases response to substance , decreases transport , increases abundance , increases mutagenesis .

Hyperparameter	RotatE	TuckER	RDF2Vec
Entity dim.	$\{100,\dots,500\}$ / 500	$\{100,200,300\}$ / 200	$\{100,200,300\}$ / 200
Relation dim.	$=$ entity dim	$\{30,65,100\}$ / 65	$=$ entity dim
Learning rate	$[10^{-5},5\times 10^{-4}]$ / $2.8\times 10^{-4}$	$[5\times 10^{-4},5\times 10^{-3}]$ / $5.7\times 10^{-4}$	—
Margin	$\{3,6,9,12\}$ / 9	—	—
Adv. sampling temp.	$[0.5,1.0]$ / 0.91	—	—
Negative samples	$[50,500]$ / 467	—	$\{10,15,25\}$ / 10

Table 2: Hyperparameter search spaces and best values selected for each model, reported in the format search space / best value . We use the following abbreviations: dim. = dimension, self-adv. sampling temp. = the self-adversarial sampling temperature of RotatE 56 . A dash (—) indicates that the hyperparameter was either not tuned (in which case, we report only the value used) or not applicable to the given model.

		D2T (document $\to$ triple)				T2D (triple $\to$ document)
Model	Config	MRR	Hits@1	Hits@10	Rank	MRR	Hits@1	Hits@10	Rank
Verbalized	BioBERT / –	8.41 $\pm$ 0.3	5.47 $\pm$ 0.3	13.69 $\pm$ 0.4	927 $\pm$ 45	7.41 $\pm$ 0.1	4.48 $\pm$ 0.1	12.98 $\pm$ 0.2	845 $\pm$ 51
Verbalized	PubMedBERT / –	4.43 $\pm$ 0.3	2.81 $\pm$ 0.2	7.45 $\pm$ 0.4	4826 $\pm$ 96	2.83 $\pm$ 0.2	1.68 $\pm$ 0.2	4.87 $\pm$ 0.3	5066 $\pm$ 136
Aligned (text $\to$ KG)	PubMedBERT / RDF2Vec	15.99 $\pm$ 0.6	8.74 $\pm$ 0.6	30.17 $\pm$ 0.7	38 $\pm$ 2	12.13 $\pm$ 0.3	5.92 $\pm$ 0.2	24.30 $\pm$ 0.8	55 $\pm$ 2
Aligned (KG $\to$ text)	BioBERT / RotatE	3.75 $\pm$ 0.3	1.34 $\pm$ 0.3	7.84 $\pm$ 0.6	349 $\pm$ 18	5.89 $\pm$ 0.2	2.21 $\pm$ 0.3	12.41 $\pm$ 0.3	147 $\pm$ 7
Aligned (bidirectional)	PubMedBERT / RotatE	15.15 $\pm$ 0.6	8.51 $\pm$ 0.6	28.27 $\pm$ 0.6	48 $\pm$ 3	5.74 $\pm$ 0.3	2.29 $\pm$ 0.3	11.97 $\pm$ 0.5	159 $\pm$ 5

Table 3: Retrieval performance in both settings. We report averages and standard deviations over five folds. Rank denotes absolute median rank. For each training direction, we report its best-on-average configuration; the baseline uses no KG embeddings. The Config column reports the text encoder and KG embedding model. All aligned configurations use a linear projection head over concatenated subject, predicate, object embeddings. The best results per column are highlighted in bold, the second best are underlined.

Panel of Figure 3 — Figure 3: Impact of the training direction, projection head, and triple-composition operator on MRR. Bars show both retrieval settings ( D2T , T2D ), except (d), which shows D2T only.

Panel of Figure 4 — Figure 4: Impact of hard-negatives selection and the text encoder on MRR, in both retrieval settings.

Biomedical KGs

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

See all 71 papers

Query-Conditioned Knowledge Alignment for Reliable Cross-System Medical Reasoning
May 18, 2026Yan Jiao, Jingran Xu, Pin-Han Ho +1KG Entity Alignment Knowledge Graphs
Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen, China · Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada · School of Computer Science and Engineering, Kyungpook National University, Daegu, South Korea
Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning
Nov 26, 2025Zhenchao Tang, Fang Wang, Haohuai He +13LLM Alignment LLM Fine-Tuning
Tencent AI for Life Sciences Lab, Shenzhen, China. · Sun Yat-sen University, Shenzhen, China. · The Hong Kong Polytechnic University, Hong Kong SAR, China. +2
Surgical Alignment in Knowledge Graph Training for Clinical Diagnosis with Large Language Models
Aug 27, 2026Saksham Khatwani, He Cheng, Majid Afshar +2LLM Reasoning with Graphs LLM Alignment
University of Colorado Anschutz Medical Campus · University of Wisconsin Madison · Loyola University

May 18, 2026J/K move · Enter open · S save

Query-Conditioned Knowledge Alignment for Reliable Cross-System Medical Reasoning

Yan Jiao, Jingran Xu, Pin-Han Ho, Limei Peng

Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China, Shenzhen, China · Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada · School of Computer Science and Engineering, Kyungpook National University, Daegu, South Korea

KG Entity Alignment Knowledge Graphs

Open paper AskPDF arXiv

Cross-domain knowledge alignment is essential for integrating heterogeneous medical systems, yet existing approaches typically treat entity alignment as a static matching problem, ignoring query context and cross-system asymmetry. This limitation is particularly critical in integrative medical settings, where correspondence between concepts is inherently context-dependent, non-bijective, and direction-sensitive. In this paper, we propose Query-Conditioned Entity Alignment (QCEA), which reformulates entity alignment as a query-conditioned correspondence problem. Instead of learning a fixed mapping between entity representations, QCEA treats the textual description of a source entity as a query and ranks candidate entities in the target graph, enabling context-dependent alignment. The framework integrates semantic encoding, graph-based representation learning, and a direction-aware transformation module to capture asymmetric and many-to-many correspondence across heterogeneous knowledge systems. We evaluate QCEA on TCM--WM knowledge graphs derived from SymMap, covering both symptom alignment and herb--molecule alignment tasks. Experimental results show consistent improvements over representative baselines, particularly on rank-sensitive metrics such as Hit@K and MRR. Furthermore, downstream retrieval-augmented generation (RAG) experiments demonstrate that improved alignment leads to better evidence retrieval, stronger grounding, and higher answer accuracy. These findings highlight that alignment is not merely a data integration step, but a key factor that shapes knowledge accessibility and reliability in cross-system medical reasoning.

Abstract

Biomedical KGs

Explore similar work

Query-Conditioned Knowledge Alignment for Reliable Cross-System Medical Reasoning

Aligning LLMs with Biomedical Knowledge using Balanced Fine-Tuning

Surgical Alignment in Knowledge Graph Training for Clinical Diagnosis with Large Language Models