AI-powered TargetMap: Enabling system-level target discovery through full-path reasoning on a unified knowledge graph
Xizhi Jin, Sijie Wang, Jiahe Chen, Shuhao Shen, Fangjie Yan, Jian Wu, Qiaojun He, Hongxia Xu, Ruijia Wu, Ji Cao, Bo Yang
TL;DR
TargetMap is an AI platform that uses a unified knowledge graph and full-path reasoning to discover drug targets by analyzing biological networks at a system level.
Contribution
Introduces TargetMap, an AI-powered platform that enables system-level drug target discovery through full-path reasoning on a unified knowledge graph.
Findings
TargetMap uses a unified knowledge graph to facilitate system-level therapeutic target discovery.
Full-path graph reasoning enables global mechanistic analysis by leveraging LLM capabilities.
Interactive visualization allows exploration of signaling networks and pathway linearization into temporal narratives.
Abstract
Modern drug discovery faces a critical challenge: high attrition rates often stem from an incomplete understanding of how individual targets operate within complex, system-wide biological networks. While computational tools such as graph neural network (GNN) excel at extracting local topological patterns, they often fall short in capturing the global semantic context and long-range dependencies within mechanistic pathways. To bridge this gap, we introduce TargetMap, an AI-driven knowledge graph platform powered by an LLM-based full-path graph reasoning algorithm. Our approach leverages the contextual reasoning capabilities of LLMs by representing entire biological pathways—as captured by structured knowledge graphs—as coherent narratives for holistic analysis. This enables the generation of testable hypotheses for distinct therapeutic targets based on a system-level understanding of…
Click any figure to enlarge with its caption.
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9Peer Reviews
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Graph Neural Networks · Graph Theory and Algorithms · AI-based Problem Solving and Planning
Introduction
Target-based drug discovery (TBDD) has been a foundational paradigm for modern therapeutic development, successfully delivering numerous treatments against well-characterized molecular entities.1^,^2 However, the overall productivity of pharmaceutical R&D has declined, with low success rates particularly in complex diseases such as oncology and central nervous system disorders.3^,^4 This suggests that a sole reliance on a single-target approach may be insufficient for understanding polygenic and system-wide diseases. As Sadri noted, the inherent reductionism of classical TBDD, while powerful for certain targets, may overlook the critical dynamics of pathological networks and their compensatory mechanisms.5 Consequently, there is a growing consensus on the need for a more integrated perspective that complements target-centric views with a network-level understanding of disease mechanisms.5^,^6^,^7^,^8 An ideal framework for this purpose requires two synergistic components: (1) a comprehensive mapping of disease pathogenesis and (2) advanced analytical algorithms capable of interpreting these maps from a global perspective.
The foundation for such a framework relies on high-quality, mechanistic data. Existing biological databases can be broadly categorized by their primary focus. Resources such as BioGRID,9 STRING,10 and SIGNOR 3.011 offer extensive catalogs of biomolecular interactions, providing invaluable breadth and a global network view. In contrast, expert-curated pathway databases such as KEGG,12 Reactome,13 and WikiPathways14 deliver depth by detailing specific molecular mechanisms with high precision. A central challenge, therefore, is not a lack of data, but rather the integration gap between these two paradigms: how to unify discrete, deep mechanistic maps into a coherent, analyzable network without losing their rich biological context.
On the algorithmic front, graph neural networks (GNNs) combined with feature engineering have become the standard for modeling biological networks.15^,^16^,^17 Despite their success in learning from local node neighborhoods, GNNs face inherent limitations in global path-based reasoning. Issues such as over-smoothing can dilute information across long paths, hindering the model’s ability to reason about end-to-end causal chains—a capability essential for understanding disease etiology.18 Meanwhile, large language models (LLMs) such as DeepSeek-R119 have demonstrated remarkable proficiency in understanding complex contexts and performing multi-step reasoning. The emerging field of leveraging LLMs for scientific knowledge graphs presents a promising avenue to overcome the semantic limitations of structure-only models.20^,^21
To address these interconnected challenges, we developed TargetMap (Figure 1), a knowledge graph platform designed to shift the paradigm from localized network analysis to holistic mechanistic reasoning. Our core contribution is the LLM-based Full-Path Graph Reasoning algorithm, a methodology engineered to overcome the semantic and topological limitations of GNNs for pathway-level analysis. By structuring and linearizing established biological pathways from knowledge bases into coherent narratives, we enable the LLM to apply its powerful reasoning capabilities to interpret the entire pathway context. Where GNNs analyze local connectivity, our algorithm reasons about higher-order mechanisms and dependencies across the pathway. This core reasoning engine is supported by a unified knowledge base that integrates mechanistic pathways, multi-omics data, and clinical evidence, and is operationalized through Graph RAG and interactive visualization. By bridging deep mechanistic data with AI-driven holistic pathway interpretation, TargetMap provides a powerful tool for generating biologically plausible and testable hypotheses for therapeutic targets. The platform is freely accessible at https://www.targetmap.zjuaim.top/.Figure 1. System architecture and functional framework of TargetMapTargetMap integrates diverse biomedical datasets to construct a comprehensive reference atlas of human signaling networks through three interconnected layers:(A) Data source layer aggregates standardized biological information from established repositories, including mechanistic pathway maps (Reactome, WikiPathways, KEGG), protein knowledgebases (UniProt), target information (TTD), disease annotations (DOID), and compound data (DrugBank).(B) Knowledge graph layer synthesizes a unified human biological signaling network through systematic graph data curation and integration. It delivers core graph functionalities, including multi-entity querying, interactive network visualization with comprehensive metadata display, and configurable graph representation options.(C) AI analysis layer provides advanced graph analytics capabilities, including a graph-based question-answering system, graph retrieval-augmented generation framework, GNN-based link prediction algorithms, and full-path graph reasoning algorithm.
Results
Core interface architecture
TargetMap’s core interface is structured into three panels, as shown in Figure 2, comprising the graphic panel, the function panel, and the detail/setting panel. The Graphic Panel serves as the primary visualization workspace, establishing a unified framework of cellular processes. Node categories are visually differentiated using distinct background colors, while protein-specific nodes are further classified employing family-specific geometric icons. Additionally, attributes such as the presence of mutations, relocations, post-translational modifications, and the unfolded relationship count are simultaneously integrated into each node’s representation to facilitate the rapid visual identification of functional roles. Unlike conventional static pathway maps, TargetMap adopts an interactive, user-driven exploration paradigm. Nodes and edges are equipped with context-aware interactive features: hover actions trigger radial highlighting of associated connections and activate context-sensitive panels displaying functional annotations and evidence-curated references, while selective clicks reveal deeper mechanistic insights. Furthermore, users can iteratively expand the network topology by tracing upstream/downstream or related neighboring nodes, such as disease nodes, from any selected node. The graph supports undo and redo operations, enabling the systematic reconstruction of interaction hierarchies. To mitigate visual complexity, the panel supports a subgraph isolation mode, allowing focused investigation of specific network modules.Figure 2. Interactive web interface of TargetMapThe interface architecture comprises three functionally integrated panels.Left function panel: A modular organization of core tools including multi-entity search, dynamic graph manipulation controls, analytical utilities, and contextual help resources.Central graphic panel: Interactive visualization workspace for exploring protein-target-drug-disease networks.Right detail/setting panel: Dynamic display of entity-specific metadata and customizable visualization settings.
The function panel is integrated into the left-hand sidebar interface and organized into five modules. The Search module supports both single-entity and multi-entity path queries, allowing users to retrieve and directly integrate results into the current graph visualization. Adjacent to this, the Utils module offers various graph manipulation utilities for enhanced workflow efficiency. For deeper insights, the analysis module features advanced algorithms for sophisticated data analysis. Furthermore, the download module enables export functionality for extracted data and visualizations, while the help module offers valuable instructional resources such as operational diagrams and video tutorials to assist users. The right-side detail/setting panel dynamically adapts its display based on user interactions. Node or edge selection activates the detail panel, which presents entity-specific multidimensional data. Without selection, the setting panel becomes accessible, offering granular control over visualization parameters, allowing users to customize the display and presentation style of the graph to suit their specific requirements. A detailed demonstration of these operational features is provided in Video S1.
Video S1. An introductory tutorial to the TargetMap web platform and its knowledge graph functionalities, related to Figure 2
Multi-source information visualization
When a user selects an entity node, the detail panel overlays the Setting Panel on the right interface, as shown in Figure 3, displaying three hierarchically structured modules: properties, relations, and neighbors. The Properties module typically comprises four subsections, with content dynamically determined by node category. Taking protein nodes as an example, the information subsection presents the metadata curated from UniProt, including accession ID, functional annotations, and cross-referenced database identifiers. The structure subsection features an interactive three-dimensional molecular visualization interface,22 while the mutation subsection aggregates somatic mutation profiles from TCGA and ICGC. The clinical subsection integrates mRNA expression data from TCGA and GTEx, with differential expression analyzed using the RSEM algorithm for quantification. Statistical significance was evaluated using Kruskal-Wallis tests for multi-tissue distribution comparisons, Wilcoxon rank-sum tests for tumor-normal differential expression analyses, and Wilcoxon signed-rank tests for matched paired-sample comparisons, with FDR correction applied to account for multiple testing (adjusted p < 0.05). Expression patterns are visualized through three complementary approaches: (i) tissue-specific boxplots depicting anatomical distribution, (ii) tumor-normal comparative boxplots highlighting disease-associated variations, and (iii) paired-sample plots with connecting lines emphasizing intra-patient expression dynamics across matched specimens. The relations module systematically catalogs signaling pathways associated with the selected entity, allowing filtration by evidence source and relation type. The neighbors module enumerates directly connected nodes, such as modified protein nodes, disease-related nodes, and drug nodes, enabling users to comprehensively understand the contextual relationships and biological significance of the target entity within the network.Figure 3. Integrated visualization platform for multi-omics biological entity informationThe platform comprises three interconnected modules.(A) The properties module displays curated metadata from reference databases, including an interactive 3D molecular structure viewer, mutation profile visualizations, and tissue expression data integrated from TCGA and GTEx databases.(B) The relations module facilitates the filtering and exploration of signaling pathway associations.(C) The neighbors module categorizes and displays directly connected biological entities.
Advanced search
TargetMap’s search interface features two distinct modes: Quick Search and Advanced Search, which are shown in Figure 4. The Quick Search module facilitates rapid retrieval of individual entities, including proteins, targets, drugs, and diseases. As an example, users can input partial gene names, protein synonyms, or protein UniProt IDs. As users input queries, a context-aware auto-completion system dynamically generates suggestions, enabling efficient search execution by selecting predefined options. Family and group queries will include their members in the results.Figure 4. TargetMap search interface overviewThe interface provides complementary search modalities for network-based discovery.(A) Quick search enables entity-specific queries via a simplified interface with auto-complete functionality.(B) Advanced search allows relationship-based exploration through dual input panels that cross-reference entries.
The Advanced Search mode supports complex queries involving multi-entity relationships. Its interface is structured into two input panels: List A and List B. Both panels incorporate integrated auto-completion functionality supporting all four entity types. List A retrieves interactions associated with the input entity, while List B constrains outputs to interactions exclusively between List A and List B entries. TargetMap’s default configuration emphasizes biological mechanism pathway information by applying weighted scoring that prioritizes protein mechanism pathway interactions. The system identifies the most relevant connections by integrating information from the top-k shortest paths and extracting the most concise pathway representation for visualization. Users can tailor both the weighting parameters and pathway display preferences to meet their specific analytical needs.
Knowledge graph-enhanced AI question answering
In contrast to traditional chunk-based text organization approaches, knowledge graphs offer a structured and contextually enriched framework for representing validated biological knowledge.23 This characteristic confers substantial advantages in enhancing LLMs' capacity to reason over complex biological systems. Within TargetMap, the knowledge graph operates as a foundational organizational scaffold, enabling the systematic management of multi-scale biological data. The platform’s graph-driven question-answering (QA) system converts dynamically retrieved entity-attribute data and relational patterns from subgraphs derived from user interactions into machine-interpretable text. This transformation allows LLMs to better contextualize task semantics during reasoning processes, thereby supporting end-users in deriving biologically meaningful insights from interconnected graph content.
The retrieval-augmented generation (RAG) framework represents an AI architecture that dynamically integrates external knowledge retrieval with generative capabilities, substantially enhancing answer accuracy and source traceability.24^,^25^,^26 TargetMap’s implementation extends this paradigm through a multi-stage workflow: biological entity recognition, graph node alignment, and intent-aware subgraph construction. Contextual features—including pathway topology, node annotations, and query semantics—are methodically reintegrated into the LLM to generate responses with improved accuracy and verifiable sourcing (Figure 5). To accommodate diverse query dimensions, the module adopts a bifurcated retrieval strategy. For single-entity queries, we integrated the reasoning paradigm of the knowledge graph-based thought (KGT) framework.27 Transcending standard retrieval methods that rely solely on local neighborhood expansion, this approach implements a “Graph Schema-Based Inference” mechanism. Specifically, the system parses the query to identify entity types, conducts reasoning over the high-level graph schema to deduce optimal relational pathways, and synthesizes a structured query to extract an intent-aligned subgraph. In contrast, for multi-entity scenarios, the system prioritizes topological connectivity by retrieving the top-20 shortest paths between entities, thereby elucidating explicit structural associations. Ultimately, contextual features derived from both strategies—spanning pathway topology and node annotations—are methodically synthesized and fed into the LLM to facilitate high-accuracy generation.Figure 5. Integrated knowledge graph architecture for biological pathway analysis and target predictionThe system processes human signaling pathway information from the Reactome Database (N = 13526 relations) through a multi-stage pipeline that first discriminates between essential regulatory nodes and redundant intermediates using topological analysis. After text data transformation consolidates multi-step processes into unified mechanistic representations, the pipeline combines pathway data with entity information. This integrated knowledge framework branches into three parallel analytical approaches: a graph-based QA system for extracting entity-attribute relationships and transforming graphs into semantic contexts; a Graph RAG system for intent-driven knowledge extraction and subgraph synthesis; and a full-path graph reasoning algorithm that contextualizes pathway analysis and evaluates druggability.
Unlike conventional RAG systems, TargetMap incorporates interactive subgraph visualizations that allow users to spatially explore results within their network context. This functionality transforms passive answer retrieval into an active “Observe-Reason-Expand-Verify” discovery loop, as demonstrated in our breast cancer case study (Figure S2). In this workflow, the system successfully guided the exploration of the oncogenic protein AKT1, utilizing chain-of-thought (CoT) reasoning to formulate hypotheses regarding PTK6-mediated phosphorylation as a signaling “switch” and its synergistic pathogenicity with BRCA1/2-mediated DNA damage responses. By dynamically expanding the network to include these interactions, TargetMap not only supported the theoretical framework but also delineated specific, testable pathways, confirming the platform’s utility in generating high-value scientific hypotheses.
Graph neural network-based link prediction in biological networks
Link prediction represents a fundamental downstream task in knowledge graph reasoning, holding particular significance for analyzing complex biological networks. Current computational methodologies primarily fall into two paradigms: knowledge graph embeddings (KGEs), which model entities and relations in low-dimensional vector spaces, and graph neural networks (GNNs), which utilize message-passing mechanisms to capture topological dependencies.15^,^16^,^17 To establish a rigorous evaluation framework, we leveraged the TargetMap database, converting biological interactions into standardized triple representations. We employed a temporal slicing strategy28^,^29 to prevent information leakage: models were trained on historical data preceding a designated timestamp and validated on subsequent temporal sequences to assess their capability in predicting future link formation. Our comparative study implemented a diverse set of baselines. For KGE approaches, we applied translational and semantic matching models, including TransE,30 RESCAL,31 DistMult,32 ComplEx,33 and RotatE.34 Furthermore, to evaluate deep structural learning capabilities, we incorporated classic GNN architectures: GCN,35 GraphSAGE,36 GAT,37 and the relation-aware RGCN38 (For details, see Method S1 and Table S1).
Despite systematic hyperparameter optimization—exploring diverse strategies for message passing layers, dropout rates, and aggregation functions—standard GNN architectures yielded suboptimal results on this dataset. We attribute this performance bottleneck to inherent structural mismatches between standard GNN architectures and the nature of biological signaling. GNNs are primarily designed to refine node representations through local neighborhood aggregation. However, biological signaling is defined by precise, long-range directed cascades (e.g., activation chains) rather than simple homophily or local proximity. In the absence of rich initial node attributes (such as protein sequences or textual descriptions), the message-passing mechanism tends to over-smooth the representations, failing to capture the distinct regulatory roles of proteins within complex pathways.
Consequently, the top-performing structure-based baselines were the translational KGE models (specifically TransE), which better model direct relational plausibility. However, even the optimized TransE model achieved recall rates plateauing at approximately 25% in target ranking tasks (see Tables S2 and S3 for further analysis). This limitation highlights the critical challenge of preserving global semantic integrity during computational modeling. While KGEs capture local structural dependencies, they struggle to reason over the cross-scale characteristics—such as pathway interdependencies and regulatory feedback loops—necessitating more advanced architectures capable of deciphering multi-layered interactions.
A high-performance framework for target prediction
The recent paradigm shift in LLMs presents transformative potential. With context window capacities expanding exponentially from 4K to 8K tokens to the current 128K-1M ranges, LLMs now demonstrate unprecedented capability in processing complex biological pathway data.39^,^40 Leveraging this advancement, we developed an architecture that integrates full-pathway contextual data with entity descriptions through LLM-based target prediction pipelines.
The structure of the full-path graph reasoning algorithm is illustrated in Figure 6. The algorithm begins with a deep semantic analysis of the user query, prioritizing entity recognition and intent decomposition. Beyond simply extracting core entities (e.g., disease names), the system analyzes the query context to auto-generate a series of granular sub-questions. This step is pivotal for uncovering implicit research intent; for instance, a query regarding therapeutic targets often necessitates an understanding of underlying pathogenic or drug-resistance mechanisms. Based on these sub-questions, the system performs entity expansion to recommend supplementary nodes, ensuring a reasoning scope that is both precise and sufficiently broad. Subsequently, the model retrieves comprehensive knowledge graph data corresponding to these entity lists and transforms this structured information into natural language “event texts,” thereby enabling the LLM to effectively interpret complex biological relationships.Figure 6. Architecture of the LLM-based full-path graph reasoning algorithmThe framework integrates full-pathway biological context with LLM reasoning capabilities, structured into three phases.(1) Question understanding and task classification: The system parses user queries to extract core entities and decomposes complex intents into granular sub-questions, while recommending supplementary entities for comprehensive graph retrieval.(2) Intelligent task routing and execution: Tasks are routed to specific pipelines (mechanism analysis, target prediction, or scientific question). The core target prediction pipeline (red section) features an iterative “hypothesis-validation” closed loop. The model generates candidates based on knowledge graph data and validates them against disease pathways; failed validations trigger automated re-prediction (up to 3 iterations) to ensure high confidence.(3) Unified output generation: The final module synthesizes the reasoning context, traceable evidence chains, and Mermaid-based pathway visualizations into a structured report.
Following data ingestion, the system classifies the task into one of the three distinct modules: mechanism analysis, target prediction, or scientific question exploration. In the Scientific Question Exploration module, the system demonstrates the capacity to synthesize complex, multi-disease hypotheses, such as differentiating oncogenic drivers and metabolic contexts across hepatocellular carcinoma, colorectal cancer, and pancreatic cancer (Figure S1). Within the core target prediction module, the architecture enforces a rigorous “prediction-validation” iterative closed loop. Initially, the model proposes potential targets based on available graph information. These candidates are then cross-referenced against the entity list to retrieve detailed attributes, neighbor relationships, and disease graph overlaps. This leads to a critical validation phase where the model evaluates the mechanistic alignment between target attributes and disease pathways. This rigorous validation was exemplified in a breast cancer case study, where the system successfully deduced the synthetic lethality between POLQ inhibition and BRCA-deficiency by tracing compensatory DNA repair mechanisms (Figure S3). If a candidate fails validation—due to insufficient pathway overlap or druggability concerns—the system excludes it and automatically triggers a re-prediction. This iterative process cycles up to three times to ensure the identification of high-confidence targets or to exhaust potential options, ultimately synthesizing the full reasoning process and evidence chain into a unified, structured analysis report.
Performance on target prediction
To validate the efficacy of the chain-of-thought (CoT) logic embedded within our Full-Path Graph Reasoning algorithm, we first evaluated the intrinsic reasoning capabilities of three long-context architectures: GLM4-long,41 GPT-4o-mini-high, and Gemini 2.0 Flash (detailed results are shown in Table S6). Given that therapeutic target discovery necessitates structured inference rather than simple pattern matching, we specifically assessed whether this CoT scaffolding could mitigate generation artifacts. We observed that under standard direct prompting—such as broad directives to “predict potential therapeutic targets for breast cancer”—the models frequently failed. Lacking a logical guide, they were prone to hallucinations, often fabricating plausible-sounding but biologically non-existent proteins. However, guiding the models through our structured CoT process resulted in a significant performance shift. This approach successfully transitioned the architectures from generating speculative artifacts to identifying clinically validated targets, confirming that the CoT protocol is essential for unlocking the models' latent knowledge for robust, evidence-based prediction.
We adopted the superior-performing Gemini 2.0 Flash as our base model. Building on this validated reasoning base, we quantified the efficacy of our 'Full Pathway Formalism' against representative structural baselines. As shown in Figure 7A, our approach yielded an Overall Recall@20 of 72.9%. This is substantially higher than the baseline LLM without pathway context (42.9%), confirming a key insight: while CoT serves as the reasoning scaffold, the integration of structured mechanistic knowledge is essential for accurate prediction.Figure 7. Performance evaluation of the full-path graph reasoning algorithm for target prediction(A) Comparative analysis of Recall@20 scores for our full-path reasoning method against baseline models and ablation variants, stratified by evidence levels.(B) Generalizability assessment of the full-path reasoning framework versus a baseline model across four major cancer types, n = 6. Data are represented as mean ± SD.(C) Robustness analysis under network perturbation scenarios, illustrating performance changes in response to low-degree node pruning and high-degree node attacks. Statistical significance was determined using one-way ANOVA and is denoted by asterisks: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001.
Furthermore, our method’s performance is more than double that of the structure-based baseline (the optimized TransE model, 25.7%). This significant margin indicates that our framework successfully deciphers complex biological semantics that purely topology-based methods fail to capture. Ablation studies further revealed that each component of our data processing pipeline—such as the rule-based handling of intermediate complexes and the inclusion of regulatory information—is critical for achieving this peak performance.
To ensure our framework’s utility is not confined to a single disease, we tested its generalizability across four major malignancies: liver, breast, colorectal, and stomach cancer (see Tables S7–S10 for further analysis). In all four cases, as shown in Figure 7B, our Full-Path Graph Reasoning algorithm consistently and significantly outperformed the baseline, demonstrating its broad applicability as a versatile tool for oncology research.
Finally, we conducted network perturbation “stress tests” to assess the model’s robustness and the biological validity of its reasoning process, as illustrated in Figure 7C. When peripheral, low-degree nodes were removed (“low-degree pruning”), the model’s performance remained stable and superior to the baseline, showcasing its resilience to incomplete data. Conversely, when central, high-degree hub nodes were targeted (“High-Degree Attack”), performance dropped sharply. This result powerfully validates our model, as it confirms that its predictions are critically dependent on key, biologically central nodes within the network, reflecting a genuine understanding of the biological system’s architecture.
Discussion
TargetMap is an AI-driven knowledge graph database that systematically integrates manually curated mechanistic maps with diverse multidisciplinary datasets (Figure 8) to address the critical need for a holistic, system-level approach to understanding complex biological networks. The platform’s key innovation transcends reductionist paradigms by offering researchers an environment to explore intricate disease mechanisms from a global perspective. Through its dynamic network visualization, TargetMap empowers users to intuitively navigate complex biological signaling pathways, identify potential targets, and formulate hypotheses. Furthermore, the integration of advanced AI methodologies, particularly LLMs for knowledge-graph question answering, retrieval-augmented generation, and all-path graph inference, distinguishes TargetMap from existing resources. This synergistic combination of comprehensive data, user-friendly design, and cutting-edge AI tools offers a unique and valuable platform for the scientific community.Figure 8. Statistical analysis of data integration and network topology in TargetMap(A) Proportional distribution of biological entities across primary reference databases.(B) Quantitative distribution of major biological entity types within the TargetMap knowledge graph.(C) Venn diagram illustrates the overlap and unique contributions of protein-protein relationship data from mechanistic pathway databases (Reactome, KEGG, and WikiPathways) and drug-protein interactions from DrugBank.(D) Schematic representation of the interconnected entity relationships in TargetMap, linking diseases, pathogenic mechanisms, proteins, and drugs.(E) Frequency analysis of the top 10 most common protein-protein and drug-protein interaction types, alongside the frequency of mutation-related disease-protein relationships.(F) Power-law fit of the network degree distribution, indicating a scale-free topology where most nodes have few connections and a small number of hub nodes are highly connected.(G) Example of a detailed data structure for a relationship, showing the start and end nodes, relation type, pathway context, and a descriptive summary with references.
Notably, the full-path reasoning paradigm pioneered by TargetMap resolves a critical modality mismatch prevalent in emerging graph-based QA frameworks, such as KGT,27 KG-Rank,42 and Hybrid LLM-KG.43 While these frontier approaches effectively leverage discrete subgraphs for factual retrieval, the non-linear and fragmented nature of such inputs often lacks explicit directional guidance. This structural discontinuity can lead to attention fragmentation and illogical hallucinations during the deep, multi-hop inference required for target discovery. By explicitly linearizing complex biological networks into coherent, sequence-based causal narratives, TargetMap achieves a precise semantic alignment with the inherent sequential processing mechanism of LLMs. This alignment significantly reduces reasoning overhead and facilitates the rigorous modeling of transitive causality.
For the continued evolution of TargetMap, several strategic directions are paramount. First, implementing adaptive learning frameworks will enable continuous integration of newly generated experimental data, facilitating dynamic updates to the existing knowledge graph—particularly in supplementing and refining information about entities and their interrelationships. Second, incorporating patient-derived multi-omics datasets and clinical outcome correlations will enhance the system’s predictive accuracy in evolving biological contexts. Finally, expanding the knowledge graph’s temporal dimension will enable more dynamic modeling of disease progression trajectories and drug resistance mechanisms.
The convergence of knowledge graphs and LLMs is rapidly transforming diverse fields. Within this evolving landscape, the synergistic integration of these technologies to enhance all-path graph inference algorithms represents a particularly promising avenue for advancing research. Recent advancements, exemplified by DeepSeek-R1, highlight the rapid evolution of LLMs, particularly in expanding their context length capabilities. Although current models, such as DeepSeek-R1, still exhibit limitations in processing truly extended contextual information, further gains in the context length manageable by LLMs are anticipated to render TargetMap optimally poised to leverage these advancements. This will enable TargetMap to achieve even greater sophistication in knowledge discovery and complex reasoning, ultimately facilitating deeper insights into intricate biological networks and underlying disease mechanisms.
In summary, TargetMap will play an increasingly important role in drug development and biological systems research through continuous data expansion and functional optimization. We remain committed to providing a comprehensive, dynamic, and efficient research tool to support scientists in achieving breakthrough advances in hypothesis generation, target discovery, and drug development.
Limitations of the study
While TargetMap demonstrates significant potential in mechanism exploration, we acknowledge limitations in the current experimental scope. Notably, standard GNN baselines exhibited suboptimal performance compared to KGE models, a discrepancy likely attributable to the inherent sparsity of initial node features in biological networks. Future work will therefore investigate pre-training strategies to enhance node initialization for GNNs. Additionally, the evaluation of our interactive modules remains qualitative due to the absence of established ground-truth datasets for biological question-answering. To address this, we plan to collaborate with domain experts to construct high-quality, annotated benchmarks for rigorous quantitative assessment, while simultaneously conducting formal user studies to empirically validate the platform’s improvements in time efficiency and hypothesis generation over traditional database workflows.
Resource availability
Lead contact
Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Xizhi Jin ([email protected]).
Materials availability
All materials generated in this study will be available upon request.
Data and code availability
We have open-sourced the complete TargetMap dataset, the core implementation of the Full-Pathway Graph Reasoning framework, the reference baseline model code, and the complete test data. The accession link (Zenodo: https://doi.org/10.5281/zenodo.18756427) is listed in the key resources table. Any additional information required to reanalyze the data reported in this article is available from the lead contact upon request.
Acknowledgments
This work was supported by the “Pioneer” and “Leading Goose” R&D Program of Zhejiang (grant number 2024C03003 to Bo Yang) and the Zhejiang Provincial Natural Science Foundation of China (grant number LRG26H310001 to Ji Cao).
Author contributions
J.C. and X.J. conceived and designed the study. X.J. and S.W. analyzed and drafted the article. J.C. and S.S. were involved in the data collection and web development processes. H.X., F.Y., R.W., J.W., and B.Y. critically revised the article. All of the authors have read and approved the final article.
Declaration of interests
The authors declare no competing interests.
STAR★Methods
Key resources table
REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataData used in this studyZenodohttps://doi.org/10.5281/zenodo.18756427Reactomehttps://reactome.orgVersion 93KEGGhttps://www.genome.jp/keggRelease 115.0WikiPathwayshttps://www.wikipathways.orgRelease 20250610Disease Ontology (DOID)http://disease-ontology.orgRelease 2025-06-27UniProthttps://www.uniprot.orgRelease 2025-03Therapeutic Target Database (TTD)https://db.idrblab.net/ttd/Release 2024-01-10DrugBankhttps://go.drugbank.comVersion 5.1.13AlphaFold 3https://alphafold.google.comServer 3.0TCGA & GTExhttps://portal.gdc.cancer.govRelease V8ClinicalTrials.govhttps://clinicaltrials.govAccessed March 2025Software and algorithmsDeep Graph Library-Knowledge Embedding (DGL-KE)https://github.com/awslabs/dgl-keVersion 0.1.1Deep Graph Library (DGL)https://www.dgl.ai/Version 2.5
Method details
Systematic collection and processing of pharmaceutical data resources
The construction of high-quality knowledge graphs fundamentally depends on meticulously curated datasets to ensure accurate representation of biological mechanisms. For pathway-level data integration, TargetMap employs Reactome13 as its primary knowledge repository, supplemented by KEGG12 and WikiPathways.14 These resources are distinguished by their expert-curated, manually annotated datasets, which provide detailed, mechanism-oriented insights into biological pathways.
At the entity level, biological entities are systematically classified using standardized ontologies derived from established biomedical repositories (Figures 8A–8C). This includes Disease Ontology (DOID) for diseases,44 UniProt for proteins,45 Therapeutic Target Database (TTD) for therapeutic targets,46 and DrugBank for pharmaceutical compounds.47 Essential attributes and functional annotations were methodically extracted from these repositories to establish a foundational biological framework.
To further enrich the knowledge graph, we integrated multidimensional pharmacological datasets. Three-dimensional protein structural data were procured from AlphaFold 3.48 Comprehensive somatic mutation profiles were systematically aggregated from The Cancer Genome Atlas (TCGA)49 and the International Cancer Genome Consortium (ICGC)50 data portals (data downloaded March 2025). Raw mutation data was filtered to retain only non-synonymous variants to focus on functionally relevant genetic alterations. Disease-specific expression signatures were derived through integration of transcriptomic data from TCGA and Genotype-Tissue Expression (GTEx) project.51 Gene expression levels, quantified using the RSEM52 algorithm, underwent Trimmed Mean of M-values (TMM) normalization. Differential expression analysis between normal and malignant tissues across 33 cancer types was conducted using Wilcoxon rank-sum tests for tumor-normal comparisons, Kruskal-Wallis tests for multi-tissue comparisons across GTEx primary sites, and Wilcoxon signed-rank tests for paired tumor-normal sample analyses, with an FDR-adjusted p-value < 0.05 considered statistically significant. Clinical context was incorporated through structured metadata extraction from ClinicalTrials.gov,53 focusing on key fields including trial phase, therapeutic modality, and recruitment status. The integration of these heterogeneous datasets enables comprehensive characterization of potential targets within multiple biological contexts, thereby facilitating evidence-based target prioritization.
Data redundancy processing and relationship extraction
During the visualization of biological pathway data, we observed substantial redundancy in protein complex node representations, particularly within the Reactome database, which may impair interpretability. To address this, we developed a rule-based simplification approach. We defined “transitional complexes” as those meeting the following topological criteria: (1) the node’s degree is exactly 2; (2) its two neighboring nodes are not directly connected; and (3) the complex itself is not a primary reactant or product in other interactions. The functional annotations of these identified transitional complexes were then programmatically merged into their immediate downstream effector complexes. This method streamlined pathway representations by reducing the total count of complex nodes by 45% (from 10,399 to 5,699) while preserving the core mechanistic flow. For clarity, proteins were distinguished by major subcellular localizations (extracellular, membrane, cytoplasmic, and nuclear), and a unified annotation system was applied to proteins with multiple post-translational modification states.
To enhance the visualization fidelity of biological knowledge graphs, we developed an LLM-based relationship extraction system. We first retrieved ontological definitions of common biological relationships (e.g., phosphorylation, ubiquitination) from the Gene Ontology (GO) database.54 These definitions were then structured into a detailed prompt provided to the DeepSeek-R1 model to perform relationship classification and generate human-readable explanations. All LLM-generated outputs underwent a rigorous two-stage expert verification process by two independent biochemists. Any disagreements were resolved by a third senior expert. This process confirmed a 95% accuracy rate for the initial LLM outputs, with the remaining 5% being manually corrected to ensure reliability. The verified relationships were further organized through semantic similarity analysis. Specifically, we used a pre-trained Sentence-BERT model (all-MiniLM-L6-v2) to generate vector embeddings for each relationship’s description, followed by k-means clustering to group them into 9 functionally coherent, higher-level parent relationship types, significantly reducing representational complexity. (A detailed list of these 9 parent types and their constituent relationships is provided in Tables S4 and S5). Each curated relationship is annotated with detailed contextual information, including its type, pathway context, descriptive summaries, and supporting references (Figure 8G).
Finally, to decipher disease mechanisms, we integrated pathogenic gene-induced disease pathway information from Reactome with normal physiological pathways, using mutant proteins as the connecting elements. This, combined with drug-protein interaction data from DrugBank, allowed for the construction of comprehensive disease-protein-target-drug relationship chains (Figures 8D and 8E). The degree distribution of the resulting network, like many biological interaction networks, follows a power-law distribution, with a few nodes having many connections while most nodes have only one or a few (Figure 8F).
Text data transformation for LLM compatibility
To bridge the gap between our biological graph data and LLMs, we developed a linearization pipeline that transforms multi-step biochemical processes into cohesive, sequential narratives. This framework consolidates reaction events, regulatory mechanisms, and mutational data into a unified formal structure. Mathematically, the generated pathway narrative is defined as a sequence of state transitions originating from a substrate S and culminating in a final product E:
Where:
S represents the initial substrate nodes.
E represents the final product nodes.
ℒk denotes the intermediate complexes formed at step k.
Each transition operator θk encapsulates the specific biological context of the reaction step, defined as a tuple of elements:
Here, Rk denotes the core reaction event, Mutk indicates associated mutational alterations (denoted in the string output by !), and Regk represents regulatory elements (denoted by ˆ). Moreover, “|” means operator separating reaction branches and “_” designates intermediate complexes.
For example, consider the simplified pathway: “Mutated RAS binds to RAF, and this complex then phosphorylates MEK, a process positively regulated by KSR1”. Our pipeline transforms this into the following structured text string:
This formalism provides a structured methodology that encapsulates comprehensive biological mechanism information—including temporal sequencing, regulatory dynamics, and mutational data into a format amenable to analysis by LLMs while preserving mechanistic accuracy.
Web architecture
The technical architecture of TargetMap adopts a modern technology stack to ensure efficient system performance and a robust user experience. The system implements Neo4j as its graph database engine to manage network-based representations of multimodal biological relationships. The interactive graph visualization is custom-built from the ground up using the D3.js library. The backend API is developed on the Django framework, with the frontend implemented in Vue3 for a responsive user interface. Elasticsearch is integrated to enhance query performance through optimized full-text search capabilities.
Quantification and statistical analysis
Performance evaluation for target prediction
To quantitatively evaluate our framework’s performance, we defined target prediction as a ranking task. For any given disease, the model ranks candidate proteins according to their likelihood of being a valid therapeutic target. We established a ground truth dataset of known disease-target associations curated from the TTD.46 This dataset represents the “true positives” and is stratified into three evidence-based categories: “Clinically Validated” (Rclinical), “Existence of Inhibitors” (Rinhibitor), and “Academically Validated” (Racademic). Our primary evaluation metric is Recall@k, a standard measure for ranking performance, which we define as:
In our main experiments, we set k = 20, thus calculating Recall@20. To create a single, comprehensive performance score that reflects the translational value of different evidence levels, we calculated a simple arithmetic mean of the Unweighted Recall@20 scores and a Weighted Overall Recall@20. This metric prioritizes targets with greater clinical relevance by assigning weights to each category: “Clinically Validated” (weight: 4), “Existence of Inhibitors” (weight: 2), and “Academically Validated” (weight: 1). The final score was calculated using the following formula:
This weighting scheme is designed to prioritize the model’s ability to identify targets with higher therapeutic potential and established clinical relevance.
Statistical analysis
Data are presented as mean ± SD; graphs and statistical analyses were generated using GraphPad Prism 8.3.0; Multi-group comparisons were analyzed using a one-way ANOVA followed by Tukey’s post hoc test; significance levels: p < 0.05 (∗), p < 0.01 (∗∗), p < 0.001 (∗∗∗), p < 0.0001 (∗∗∗∗), ns (not significant, p ≥ 0.05).
The reference list from the paper itself. Each links out to its DOI / PubMed record.
- 1Santos R.Ursu O.Gaulton A.Bento A.P.Donadi R.S.Bologa C.G.Karlsson A.Al-Lazikani B.Hersey A.Oprea T.I.Overington J.P.A comprehensive map of molecular drug targets Nat. Rev. Drug Discov.162017193410.1038/nrd.2016.23027910877 PMC 6314433 · doi ↗ · pubmed ↗
- 2Emmerich C.H.Gamboa L.M.Hofmann M.C.J.Bonin-Andresen M.Arbach O.Schendel P.Gerlach B.Hempel K.Bespalov A.Dirnagl U.Parnham M.J.Improving target assessment in biomedical research: the GOT-IT recommendations Nat. Rev. Drug Discov.20202164813319988010.1038/s 41573-020-0087-3PMC 7667479 · doi ↗ · pubmed ↗
- 3Wong C.H.Siah K.W.Lo A.W.Estimation of clinical trial success rates and related parameters Biostatistics 20201927328610.1093/biostatistics/kxx 06929394327 PMC 6409418 · doi ↗ · pubmed ↗
- 4Cummings J.L.Morstorf T.Zhong K.Alzheimer's disease drug-development pipeline: few candidates, frequent failures Alzheimers Res. Ther.620143710.1186/alzrt 26925024750 PMC 4095696 · doi ↗ · pubmed ↗
- 5Sadri A.Is Target-Based Drug Discovery Efficient? Discovery and “Off-Target” Mechanisms of All Drugs J. Med. Chem.662023126511267710.1021/acs.jmedchem.2c 0173737672650 · doi ↗ · pubmed ↗
- 6Fernald K.D.S.Förster P.C.Claassen E.van de Burgwal L.H.M.The pharmaceutical productivity gap - Incremental decline in R&D efficiency despite transient improvements Drug Discov. Today 29202410416010.1016/j.drudis.2024.10416039241979 · doi ↗ · pubmed ↗
- 7Davidov E.Holland J.Marple E.Naylor S.Advancing drug discovery through systems biology Drug Discov. Today 8200317518310.1016/S 1359-6446(03)02600-X 12581712 · doi ↗ · pubmed ↗
- 8Nguyen D.-T.Mathias S.Bologa C.Brunak S.Fernandez N.Gaulton A.Hersey A.Holmes J.Jensen L.J.Karlsson A.Pharos: Collating protein information to shed light on the druggable genome Nucleic Acids Res.452017 D 995D 10022790389010.1093/nar/gkw 1072 PMC 5210555 · doi ↗ · pubmed ↗