The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
Knowledge-Driven Neuro-Symbolic Reasoning for Personalized
Oncology Treatment Recommendation Based on Multi-Modal
Medical Knowledge Graph
Liuqing Yanga, Hong Wana, Jiangping Zhub, Pei Zhoub and Zhongjian Wanga,∗
aBCPM Data Limited, Building A, 1st Floor, 1480 North Tianfu Avenue, Chengdu High-tech Zone, Chengdu, 610041, Sichuan, China
bCollege of Computer Science, Sichuan University, (Wangjiang Campus) No. 24 South Section 1, Yihuan Road, Chengdu, 610065, Sichuan, China
A R T I C L E I N F O
Keywords:
Neuro-Symbolic AI
Knowledge Graph
Treatment Recommendation
Precision Oncology
Explainable AI
Multi-Modal Learning
A B S T R A C T
Personalized oncology treatment recommendation is a critical clinical task that requires in-
tegrating complex, multi-modal patient data with established medical knowledge to ensure
both accuracy and safety. While deep learning models excel at capturing latent patterns from
high-dimensional data, their opaque decision-making processes and inability to strictly enforce
clinical constraints hinder their adoption in high-stakes medical domains. Conversely, traditional
rule-based systems offer high interpretability but struggle to scale with complex, heterogeneous
data. To address these challenges, we propose the Knowledge-driven Neuro-Symbolic Network
(K-NeSyNet), a novel framework for personalized oncology treatment recommendation. K-
NeSyNet is grounded in a newly constructed Multi-Modal Oncology Knowledge Graph (MM-
OKG) that unifies genomic mutations, medical imaging features, clinical text, and structured
medical guidelines from publicly available sources including TCGA, DGIdb, KEGG, and NCCN
guidelines. The core innovation of K-NeSyNet is a three-channel differentiable symbolic rea-
soning mechanism that explicitly models guideline recommendations, mutation-target matching,
and contraindication penalties for each patient-drug pair. These symbolic signals are dynamically
fused with the outputs of a knowledge-aware graph attention neural reasoning module via an
adaptive gated fusion network. Crucially, the fusion gate learns to balance neural and symbolic
confidence in a patient-specific manner, while contraindication evidence enters both the symbolic
score and a dedicated safety objective to down-weight clinically risky drugs. Extensive experi-
ments on a real-world multi-modal oncology dataset comprising 4,781 patients across 10 cancer
types demonstrate that K-NeSyNet consistently outperforms eight state-of-the-art baselines.
Specifically, K-NeSyNet achieves the highest F1@10 of 0.9227, NDCG@10 of 0.9656, and
Jaccard similarity of 0.9366, while maintaining competitive Clinical Guideline Consistency.
Ablation studies confirm the indispensable role of each component, with the removal of the
fusion gate causing the most significant performance degradation. Furthermore, K-NeSyNet
provides transparent, score-decomposed explanations for its recommendations, offering a crucial
step toward trustworthy AI-assisted clinical decision support.
1. Introduction
Precision oncology has fundamentally transformed cancer treatment by tailoring therapeutic strategies to the
molecular and clinical characteristics of individual patients [1, 2]. Unlike the traditional “one-size-fits-all” paradigm,
precision oncology leverages genomic profiling, medical imaging, and clinical records to identify the most effective and
least harmful treatment for each patient. However, the exponential growth in multi-modal patient data, the complexity
of drug-gene-disease interactions, and the rapid evolution of clinical guidelines present significant challenges for
clinicians seeking to make optimal treatment decisions in a timely manner [3, 4].
Computational approaches, particularly deep learning, have shown remarkable promise in automating aspects of
clinical decision-making [5–7]. Knowledge graph (KG)-based methods further enhance recommendation quality by
incorporating structured biomedical knowledge, such as drug-drug interactions and drug-disease associations [8–10].
Despite these advances, existing methods suffer from three critical limitations when applied to the oncology treatment
recommendation task.
∗Corresponding author
yanglq@bcpmdata.com (L. Yang); med@bcpmdata.com (H. Wan); zjp16@scu.edu.cn (J. Zhu); zhoupei@scu.edu.cn (P. Zhou);
wangzj@bcpmdata.com (Z. Wang)
ORCID(s): 0009-0002-9623-5877 (L. Yang); 0009-0003-4523-7263 (H. Wan); 0000-0002-9756-5264 (J. Zhu);
0000-0003-1631-3637 (P. Zhou); 0000-0003-2247-3848 (Z. Wang)
Yang et al.: Preprint submitted to Elsevier Page 1 of 15
.CC-BY 4.0 International licenseIt is made available under a
display the preprint in perpetuity.
is the author/funder, who has granted medRxiv a license to(which was not certified by peer review)for this preprint The copyright holderthis version posted June 2, 2026.;https://doi.org/10.64898/2026.06.01.26354443doi:medRxiv preprint
Page 2
K-NeSyNet: Knowledge-Driven Neuro-Symbolic Reasoning for Oncology Treatment Recommendation
First, most current models are unimodal or bimodal, relying primarily on electronic health records (EHR) or
molecular data alone, while neglecting the rich complementary information available in medical imaging and clinical
narratives. In oncology, treatment decisions are inherently multi-modal: a pathologist examines histopathology slides,
a radiologist reviews CT scans, a molecular biologist interprets genomic mutations, and a clinician synthesizes all of
these with the patient’s clinical history [11]. Failing to integrate these diverse modalities leads to an incomplete patient
representation.
Second, existing deep learning approaches operate as “black boxes” that lack the ability to explicitly encode and
enforce clinical rules. In oncology, certain drugs are strictly contraindicated for patients with specific genetic mutations
or comorbidities (e.g., immunotherapy monotherapy is generally not recommended as first-line treatment for EGFR-
mutant non-small cell lung cancer). A purely neural model may learn statistical correlations that approximate these
rules but cannot guarantee strict adherence, posing serious safety risks [12, 13].
Third, current methods provide limited explainability. Clinicians require transparent reasoning paths to trust and
adopt AI-generated recommendations. A recommendation system that simply outputs a ranked list of drugs without
explaining why a particular drug is recommended (or not recommended) is unlikely to gain clinical acceptance [14].
To address these challenges, we propose the Knowledge-driven Neuro-Symbolic Network (K-NeSyNet), a novel
framework that synergistically combines the pattern recognition strengths of neural networks with the logical rigor
and transparency of symbolic reasoning. K-NeSyNet is built upon a newly constructed Multi-Modal Oncology
Knowledge Graph (MM-OKG) that integrates patient-level multi-modal features (imaging, text, genomics) with
structured biomedical knowledge from TCGA [15], DGIdb [16], KEGG [17], and NCCN clinical guidelines.
The core innovation of K-NeSyNet lies in its three-channel differentiable symbolic reasoning module, which
computes explicit symbolic scores for each patient-drug pair along three clinically meaningful dimensions: (1)
guideline support, which measures whether a drug is recommended by clinical guidelines for the patient’s cancer
type; (2) target matching, which evaluates whether the drug targets any of the patient’s mutated genes; and (3)
contraindication penalty, which identifies drugs that are contraindicated given the patient’s specific mutation profile.
These symbolic scores are then dynamically integrated with the neural predictions from a knowledge-aware graph
attention network (KGAT) through an adaptive gated fusion network. The fusion gate is conditioned on both the
patient’s multi-modal embedding and the relative magnitudes of the neural and symbolic scores, enabling patient-
specific balancing of data-driven and knowledge-driven signals. Crucially, contraindication evidence is injected into
both the symbolic score and a dedicated safety objective, so clinically risky drugs are consistently down-weighted
during learning and inference.
Our main contributions are summarized as follows:
• We propose K-NeSyNet, a neuro-symbolic framework that integrates multi-modal patient representations with
differentiable symbolic reasoning for personalized oncology treatment recommendation.
• We design a three-channel symbolic reasoning module (guideline, target-matching, contraindication) coupled
with an adaptive gated fusion network that dynamically balances neural and symbolic predictions in a patient-
specific manner.
• We construct MM-OKG, a comprehensive multi-modal oncology knowledge graph integrating data from TCGA,
DGIdb, KEGG, and NCCN guidelines, encompassing 4,781 patients, 15,615 entities, and 22,167 triples.
• We conduct extensive experiments demonstrating that K-NeSyNet outperforms eight state-of-the-art baselines in
both recommendation accuracy and clinical safety, while providing transparent, score-decomposed explanations.
2. Related Work
2.1. Knowledge Graph-Based Recommendation
Knowledge graphs have been widely adopted in recommendation systems to provide rich auxiliary information
beyond user-item interactions. KGNN [8] employs graph neural networks to aggregate neighborhood information in a
KG, enabling the model to capture high-order structural patterns. MKR [9] proposes a multi-task learning framework
that jointly trains KG embedding and recommendation, sharing latent features between the two tasks. KGAT [10]
introduces an attentive message-passing scheme over the KG, allowing the model to discriminate the importance of
different neighbors during aggregation. More recently, heterogeneous graph-based approaches [18] have been applied
Yang et al.: Preprint submitted to Elsevier Page 2 of 15
.CC-BY 4.0 International licenseIt is made available under a
display the preprint in perpetuity.
is the author/funder, who has granted medRxiv a license to(which was not certified by peer review)for this preprint The copyright holderthis version posted June 2, 2026.;https://doi.org/10.64898/2026.06.01.26354443doi:medRxiv preprint
Page 3
K-NeSyNet: Knowledge-Driven Neuro-Symbolic Reasoning for Oncology Treatment Recommendation
to electronic health records, demonstrating the potential of graph attention mechanisms in clinical settings. While
these methods effectively leverage KG structure for general recommendation, they do not incorporate domain-specific
clinical rules or multi-modal patient data, limiting their applicability in the medical domain.
2.2. Medication Recommendation
Medication recommendation has emerged as an important subfield of clinical decision support. GAMENet [5]
constructs a drug-drug interaction (DDI) graph and uses a graph-augmented memory module to generate safe drug
combinations. SafeDrug [6] further improves safety by incorporating a global DDI graph into the training objective,
explicitly penalizing unsafe drug combinations. COGNet [7] introduces a copy-or-generate mechanism that leverages
historical medication records to improve recommendation accuracy. Recent work has also explored drug combination
synergy prediction using deep learning [19] and knowledge graph-driven drug repurposing [20]. While these methods
represent significant progress in safe medication recommendation, they are primarily designed for general medication
scenarios and do not address the unique challenges of oncology, such as the need to integrate genomic mutation profiles,
medical imaging, and cancer-specific clinical guidelines.
2.3. Neuro-Symbolic AI in Healthcare
Neuro-symbolic AI seeks to combine the learning capabilities of neural networks with the reasoning capabilities
of symbolic systems [21, 22]. In healthcare, this paradigm has shown promise for tasks requiring both data-driven
pattern recognition and adherence to established medical knowledge. For instance, knowledge-infused learning has
been applied to integrate clinical guidelines into neural models for disease diagnosis [23]. However, most existing
neuro-symbolic approaches in healthcare either use symbolic knowledge only as a regularization signal during training
or employ a simple additive combination of neural and symbolic outputs. Neuro-symbolic systems have also been
applied to link prediction on biomedical knowledge graphs [24]. Our work advances this line of research by proposing
a dynamic, patient-specific fusion mechanism with explicit safety enforcement, specifically tailored for the oncology
treatment recommendation task.
2.4. Multi-Modal Learning in Medicine
Multi-modal learning has gained significant attention in medical AI, driven by the availability of diverse data
modalities including medical images, clinical text, and genomic data [25]. Recent work has demonstrated the benefits
of integrating imaging and genomic data for cancer prognosis [26] and treatment response prediction. Gated fusion
networks have been proposed to adaptively weight modality contributions in multi-modal settings [27], and attention
mechanisms have been explored for precision medicine applications [28]. However, most multi-modal medical
models focus on classification or prognosis tasks rather than treatment recommendation, and few integrate structured
knowledge graphs with multi-modal patient representations. K-NeSyNet addresses this gap by constructing a unified
multi-modal knowledge graph that serves as the foundation for both neural and symbolic reasoning.
3. Methodology
In this section, we present the K-NeSyNet framework in detail. As illustrated in Fig. 1, K-NeSyNet consists of four
core components: (1) Multi-Modal Oncology Knowledge Graph (MM-OKG) construction, (2) Multi-Modal Patient
Encoder, (3) Dual-Channel Reasoning Engine (neural and symbolic), and (4) Adaptive Gated Fusion Network.
Yang et al.: Preprint submitted to Elsevier Page 3 of 15
.CC-BY 4.0 International licenseIt is made available under a
display the preprint in perpetuity.
is the author/funder, who has granted medRxiv a license to(which was not certified by peer review)for this preprint The copyright holderthis version posted June 2, 2026.;https://doi.org/10.64898/2026.06.01.26354443doi:medRxiv preprint
Page 4
K-NeSyNet: Knowledge-Driven Neuro-Symbolic Reasoning for Oncology Treatment Recommendation
Figure 1: The overall architecture of the Knowledge-driven Neuro-Symbolic Network (K-NeSyNet). The framework consists
of three main components: (1) a Multi-Modal Patient Encoder that integrates medical imaging, clinical text, and genomic
mutations; (2) parallel reasoning paths including a Knowledge-Aware Graph Attention Network (KGAT) for neural pattern
recognition and a Differentiable Symbolic Reasoning Module with three clinically meaningful channels (guideline matching,
mutation-target matching, and contraindication penalty); and (3) an Adaptive Gated Fusion Network that dynamically
balances neural and symbolic signals while incorporating safety-aware contraindication penalties.
3.1. Problem Formulation
Let = {𝑝1, 𝑝2, … , 𝑝𝑁 } denote the set of 𝑁 patients and = {𝑑1, 𝑑2, … , 𝑑𝑀 } denote the set of 𝑀 candidate
drugs. Each patient 𝑝𝑖 is associated with multi-modal features: imaging features 𝐱𝑖𝑚𝑔
𝑖 ∈ ℝ512, clinical text embeddings
𝐱𝑡𝑥𝑡
𝑖 ∈ ℝ768, and genomic mutation profiles 𝐱𝑔𝑒𝑛
𝑖 ∈ ℝ256. The knowledge graph = (, , ) consists of entities ,
relation types , and triples ⊆ × × . The goal is to learn a scoring function 𝑓 (𝑝𝑖, 𝑑𝑗 ) ∈ [0, 1] that predicts the
appropriateness of drug 𝑑𝑗 for patient 𝑝𝑖, considering both data-driven patterns and clinical knowledge.
3.2. Multi-Modal Oncology Knowledge Graph (MM-OKG)
We construct the MM-OKG by integrating data from four publicly available sources:
TCGA Clinical and Genomic Data. We retrieve clinical records and somatic mutation profiles for 4,781 patients
across 10 cancer types from The Cancer Genome Atlas (TCGA) via the NCI Genomic Data Commons (GDC) API [15].
Each patient record includes demographics, cancer type, staging information, treatment history, and a list of somatic
mutations.
Drug-Gene Interactions. We obtain drug-gene interaction data from DGIdb v5 [16] , resulting in structured triples
of the form (drug, targets, gene).
Biological Pathways. We retrieve 20 cancer-related signaling pathways from the KEGG database [17] via its REST
API, establishing gene-pathway associations.
Clinical Guidelines. We encode treatment recommendations from NCCN clinical guidelines as structured rules
mapping cancer types and molecular subtypes to recommended drug regimens.
The resulting MM-OKG contains 15,615 entities, 15 relation types, and 22,167 triples. Entity types include patients,
drugs (80), diseases (40), genes (10,586), and guidelines (128). Relation types include diagnosed_with, has_mutation,
treated_with, targets, interacts_with, belongs_to_pathway, and guideline_recommends, among others.
Yang et al.: Preprint submitted to Elsevier Page 4 of 15
.CC-BY 4.0 International licenseIt is made available under a
display the preprint in perpetuity.
is the author/funder, who has granted medRxiv a license to(which was not certified by peer review)for this preprint The copyright holderthis version posted June 2, 2026.;https://doi.org/10.64898/2026.06.01.26354443doi:medRxiv preprint
Page 5
K-NeSyNet: Knowledge-Driven Neuro-Symbolic Reasoning for Oncology Treatment Recommendation
3.3. Multi-Modal Patient Encoder
The Multi-Modal Patient Encoder transforms heterogeneous patient features into a unified representation. For each
modality 𝑚 ∈ {𝑖𝑚𝑔, 𝑡𝑥𝑡, 𝑔𝑒𝑛}, we apply a modality-specific linear projection followed by layer normalization and
ReLU activation:
𝐡𝑚
𝑖 = ReLU(LayerNorm(𝐖𝑚𝐱𝑚
𝑖 + 𝐛𝑚)) (1)
where 𝐖𝑚 ∈ ℝ𝐷×𝑑𝑚 and 𝐛𝑚 ∈ ℝ𝐷 are learnable parameters, 𝑑𝑚 is the input dimension of modality 𝑚, and 𝐷 is the
unified embedding dimension.
To dynamically weight the contribution of each modality based on its informativeness for a given patient, we
employ a cross-modal attention mechanism:
𝛼𝑚
𝑖 = exp(𝐰⊤
𝑎𝑡𝑡𝐡𝑚
𝑖 )
∑𝑚′∈{𝑖𝑚𝑔,𝑡𝑥𝑡,𝑔𝑒𝑛} exp(𝐰⊤
𝑎𝑡𝑡𝐡𝑚′
𝑖 ) (2)
where 𝐰𝑎𝑡𝑡 ∈ ℝ𝐷 is a learnable attention vector. The final patient embedding is the weighted sum:
𝐡𝑖 = ∑
𝑚∈{𝑖𝑚𝑔,𝑡𝑥𝑡,𝑔𝑒𝑛}
𝛼𝑚
𝑖 𝐡𝑚
𝑖 (3)
This attention mechanism allows the model to adaptively emphasize the most relevant modality for each patient.
For instance, genomic features may be more informative for patients with actionable mutations, while imaging features
may be more critical for patients where tumor morphology drives treatment decisions.
3.4. Knowledge-Aware Neural Reasoning Module
The neural reasoning module leverages the KG structure to compute data-driven drug scores. We employ a multi-
head graph attention network (GAT) [29] that operates on the MM-OKG to learn entity embeddings that capture both
local neighborhood structure and global graph patterns.
For each entity 𝑒 ∈ , we initialize its embedding 𝐞(0) ∈ ℝ𝐷 and iteratively update it through 𝐿 layers of graph
attention:
𝐞(𝑙+1) = MultiHeadGAT(𝑙)(𝐞(𝑙), {𝐞(𝑙)
𝑛 ∶ 𝑛 ∈ (𝑒)}) (4)
where (𝑒) denotes the neighbors of entity 𝑒 in the KG. Each attention head computes:
𝛽(𝑙)
𝑒,𝑛 = exp(LeakyReLU(𝐚(𝑙)⊤[𝐖(𝑙)𝐞(𝑙)‖𝐖(𝑙)𝐞(𝑙)
𝑛 ]))
∑𝑛′∈ (𝑒) exp(LeakyReLU(𝐚(𝑙)⊤[𝐖(𝑙)𝐞(𝑙)‖𝐖(𝑙)𝐞(𝑙)
𝑛′ ])) (5)
After 𝐿 layers of message passing, we obtain the final drug embedding 𝐝(𝐿)
𝑗 for each drug 𝑑𝑗 . The neural score for
patient-drug pair (𝑝𝑖, 𝑑𝑗 ) is computed as:
𝑠𝑛𝑒𝑢𝑟𝑎𝑙
𝑖𝑗 = 𝜎(𝐡⊤
𝑖 𝐖𝑠𝑐𝑜𝑟𝑒𝐝(𝐿)
𝑗 ) (6)
where 𝐖𝑠𝑐𝑜𝑟𝑒 ∈ ℝ𝐷×𝐷 is a learnable bilinear scoring matrix and 𝜎(⋅) is the sigmoid function.
3.5. Differentiable Symbolic Reasoning Module
The symbolic reasoning module computes explicit, interpretable scores for each patient-drug pair along three
clinically meaningful channels.
Channel 1: Guideline Support Score. This channel evaluates whether drug 𝑑𝑗 is recommended by clinical
guidelines for patient 𝑝𝑖’s cancer type and stage:
𝑠𝑔𝑢𝑖𝑑𝑒
𝑖𝑗 = 𝟙[∃ guideline 𝑔 ∈ ∶ recommends(𝑔, 𝑑𝑗 , cancer_type(𝑝𝑖), stage(𝑝𝑖))] (7)
In practice, this is implemented as a differentiable lookup in the guideline knowledge base, returning 1.0 if the drug is
guideline-supported and 0.0 otherwise.
Yang et al.: Preprint submitted to Elsevier Page 5 of 15
.CC-BY 4.0 International licenseIt is made available under a
display the preprint in perpetuity.
is the author/funder, who has granted medRxiv a license to(which was not certified by peer review)for this preprint The copyright holderthis version posted June 2, 2026.;https://doi.org/10.64898/2026.06.01.26354443doi:medRxiv preprint