Trends and Hotspots in the Convergence of Artificial Intelligence and Cell and Gene Therapy: A Patent Analysis.

INTRODUCTION: The convergence of Artificial Intelligence (AI) and Cell and Gene Therapy (CGT) represents a frontier in next-generation therapeutic innovation, yet the structural patterns of technology convergence within this interdisciplinary field remain poorly characterized. This study aims to develop a computational framework for predicting which technology linkages are most likely to strengthen within this convergence landscape, thereby providing a data-driven basis for R&D resource allocation.

METHOD: Global patent data in the AI-CGT field were collected to construct a time-series heterogeneous network in which International Patent Classification (IPC) codes serve as technology nodes and individual patents as knowledge nodes, capturing co-occurrence, affiliation, and citation relationships among technologies. At the macro forecasting level, a Temporal Heterogeneous Graph Neural Network (THGNN) was employed for link prediction. The model aggregated multi-hop neighbor information through meta-paths and leveraged Long Short-Term Memory (LSTM) networks to capture technological evolution dynamics, with the aim of identifying IPC pairs with the highest future association weights. At the micro-analysis level, Latent Dirichlet Allocation (LDA) topic modeling was applied to patents associated with the core IPC pairs. The optimal number of topics was determined by combining perplexity curves with inter-topic distance visualization to reveal the key technological content of these IPC associations.

RESULT: The temporal heterogeneous graph neural network model achieved an Area Under the Curve (AUC) of 0.7539, Average Precision (AP) of 0.7862, and F1 score of 0.7568 for link prediction. Based on this model, multiple high-confidence convergence pathways were identified, among which the predicted potential for association strengthening between biochemical detection (C12Q1) and neural networks (G06N3) and between genetic engineering (C12N15) and neural networks (G06N3) received the highest prediction scores (0.89 and 0.80, respectively), suggesting that these pairs are likely to constitute the primary axes of technology convergence in this field.

DISCUSSION: The two major convergence directions identified above correspond to the core bottlenecks in CGT clinical translation-namely, the precision and safety of gene editing, and the quality control and efficacy monitoring of therapeutic products. These findings reveal a patent-based convergence landscape in which AI-related patent classifications increasingly co-occur with both detection and therapeutic methods, indicating a trend of AI integration across the diagnostic-totherapeutic continuum at the level of patent disclosures. The identified convergence directions are computational predictions based on patent co-occurrence patterns and carry inherent uncertainty; they represent technology convergence signals rather than confirmed implementation. The "macroprediction- micro-analysis" bi-level framework proposed in this study provides a quantifiable, interpretable, and transferable analytical paradigm for technology opportunity identification in interdisciplinary pharmaceutical research.

CONCLUSION: This study introduces THGNN into technology opportunity prediction in the AI-CGT convergence field and establishes a complete analytical chain from macro-level association identification to micro-level technology interpretation through LDA topic modeling. The findings offer a data-driven decision-making reference for pharmaceutical companies and research institutions seeking to strategically position their R&D pipelines within complex technology landscapes.