Research Progress of Artificial Intelligence Technology in Pharmacology of Traditional Chinese Medicine

Abstract

In recent years, artificial intelligence (AI) technologies, particularly machine learning (ML) and deep learning (DL), have demonstrated significant potential in pharmacological research on traditional Chinese medicine (TCM). Due to the complexity of TCM compositions, targets, and pathways, conventional experimental methods encounter limitations in analyzing dose-response relationships and synergistic mechanisms. AI, with its advanced learning and data-processing capabilities, enables the integration of complex information, thereby enhancing the systematic approach, efficiency, and accuracy of research, and creating new opportunities for scientific discovery. This article reviews the latest progress in applying AI to TCM pharmacology, focusing on core algorithms, typical application scenarios, and future technological challenges, with the aim of providing theoretical support and methodological references for the modernization of TCM.

Introduction

In the current era of rapid technological advancement, artificial intelligence (AI), recognized as one of the most innovative and influential fields of computer science, ¹ is transforming multiple industries at an unprecedented pace. ² The primary goal of AI is to develop computational systems that demonstrate human-like intelligence and can perform complex tasks traditionally dependent on human cognition. ³ Since its emergence as a discipline in the mid-twentieth century, AI has achieved major progress in both algorithmic theory and engineering applications, ⁴ resulting in the development of technical frameworks such as machine learning (ML), deep learning (DL), natural language processing (NLP), and computer vision. ⁵ With the broad adoption of AI technologies and the continuous accumulation of medical data, AI now plays an increasingly crucial role in drug discovery and healthcare.^6,7

In pharmacology, AI-driven approaches based on ML and DL are being integrated throughout the entire drug development process (Figure 1). ⁸ Computational intelligence has redefined the previously inefficient research and development paradigm of traditional pharmacology through data-driven multi-objective optimization, multi-omics integration, and predictive modeling. ⁹ Generative design accelerates the identification of lead compounds, network-based multi-omics analysis elucidates complex mechanisms, and predictive models reduce the risk of clinical failure. Collectively, these advances have markedly improved research efficiency, reduced resource consumption, and accelerated the introduction of new drugs, ¹⁰ while embedding data-driven and intelligent decision-making into the foundations of pharmacological research. ¹¹

Figure 1.

The application of AI in drug development. QSAR: Quantitative Structure–Activity Relationship; NLP: Natural Language Processing.

As a fundamental component of TCM modernization, TCM pharmacology presents distinctive features compared with conventional biomedical research. ¹² TCM formulations are typically composed of multiple ingredients that act on numerous biological targets through synergistic interactions, forming intricate regulatory networks across diverse pathways. ¹³ This multi-component, multi-target, and multi-pathway nature complicates the accurate analysis of dose-effect relationships using traditional experimental techniques. ¹⁴ The advent of AI provides a promising means of overcoming these research bottlenecks. ¹⁵ With its powerful data-processing capabilities, AI can integrate high-dimensional heterogeneous datasets, construct multi-level predictive models, and improve the accuracy of simulations and predictions of pharmacological effects in TCM. ¹⁶ Furthermore, the strategic use of AI has the potential to transform the conventional research paradigm of TCM, creating innovative pathways for pharmacological studies. ¹⁷ The adoption of computational intelligence in TCM pharmacology is expected to generate novel approaches, unlock biomedical value, and promote both the preservation and modernization of TCM within the framework of contemporary scientific research. ¹⁸

Core AI Algorithms

The principal AI algorithms applied in pharmacology include machine learning (ML), deep learning (DL), and natural language processing (NLP). ML is a data-driven methodology that allows computers to autonomously learn from data and improve performance through algorithmic training, without the need for explicit programming. ¹⁹ By processing large datasets, ML algorithms can recognize patterns and trends, subsequently generating predictive models that adapt to new data inputs. ²⁰ ML is generally classified into three primary types: supervised learning, unsupervised learning, and reinforcement learning. Representative algorithms include support vector machines (SVMs), K-nearest neighbor (KNN), decision trees (DT), random forests (RF), AdaBoost, K-means clustering, convolutional neural networks (CNNs), artificial neural network (ANN) and other algorithms. Deep learning (DL) is a branch of ANN algorithm in the field of machine learning. Both of them are specific methods to realize AI, and they are also the core of AI and data science. ²¹

In the pharmacology, diverse AI algorithms have been widely adopted. Clustering, SVMs, KNN, and related methods have been used to study theoretical aspects of TCM properties. ²² These approaches can elucidate correlations between TCM properties and clinical efficacy, support clinical decision-making by analyzing medical records and prescriptions, and establish models to classify the “cold” and “hot” properties of medicinal substances. In TCM ²³ theory, cold herbs are often applied to clear heat and treat heat syndromes, while hot herbs are used to dispel cold and treat cold syndromes, forming a critical foundation for TCM diagnosis and treatment. ²⁴ Additionally, hierarchical clustering and related algorithms have been employed in prescription compatibility studies to identify compatibility patterns, assess similarities among prescriptions, and infer medication rules for specific diseases. ²⁵ In toxicity research, ANNs, SVMs, and other algorithms have been used to construct predictive models for evaluating the safety of TCM, such as predicting drug-induced nephritis and contraindications during pregnancy.^26,27

AI in Pharmacodynamic Studies of TCM

AI in Drug–Target Interaction Research

Computer-aided drug screening enables efficient pre-screening of candidate compounds by constructing energy prediction models. The core principle of this approach is to quantify ligand-receptor binding free energy (ΔG), where more negative values indicate greater complex stability. ²⁸ Molecular docking serves as a key method, generating spatial interaction maps of drug binding sites through three-dimensional conformation sampling and energy optimization. ²⁹ Modern algorithms incorporate parameters such as hydrophobic interactions, hydrogen bond networks, and electrostatic complementarity to develop computationally intelligent screening systems. However, traditional methods rely on static simulations, neglect the dynamic nature of proteins, and depend heavily on empirical parameters, which introduces limitations. ³⁰ Machine learning reduces systematic bias by deconstructing structural features, while deep learning has advanced conformation sampling and activity prediction. ³¹ The AtomNet system developed by Wallach et al applies a three-dimensional convolutional neural network to process molecular structures through spatial meshing, generating feature vectors for analysis. ³² Compared with conventional methods, Ragoza's CNN-based scoring function significantly improves conformation and activity identification, although deep learning models still encounter challenges in accurately quantifying binding strength. ³³ To address this, the Pafnucy system developed by Stepniewska-Dziubinska et al utilizes a three-dimensional convolutional architecture to effectively assess drug-target binding energy. ³⁴

Machine Learning of Cell Phenotypic Data

The development of modern biomedical data-sharing systems, which use standardized interfaces to integrate large-scale biological data, has established a robust foundation for computational pharmacology. Analytical models based on such structured information enable comprehensive exploration of interactions between chemical compounds and biological phenotypes. ³⁵ The adversarial deep generative network developed by Kadurin et al, trained on the NCI-60 compound dataset, has been applied to construct predictive models capable of efficiently screening anticancer molecules in the PubChem library. ³⁶ In phenotypic detection, high-content imaging combined with automated image analysis quantifies dynamic morphological changes in drug-stimulated cells, enhancing prediction accuracy. ³⁷ For instance, O'Duibhir et al utilized high-content microscopy to capture images of three distinct leukemia cell colonies in related studies, subsequently integrating machine learning-based automatic analysis to extract 21 morphological and texture parameters of the colonies. They quantified the impact of various epigenetic drugs on the cell colonies, revealing that this approach not only accurately identified the inhibitory effects of drug concentrations on colony growth (eg, L4 reduced colony numbers at concentrations as low as 20-65 nM), but also uncovered dynamic morphological changes that traditional methods struggled to capture (eg, GSK-LSD1 induced colony differentiation into single cells), the method enhanced the accuracy of predicting drug effects, providing valuable insights for phenotype-based drug discovery in leukemia. ³⁸ The MitoReID system developed by Yu et al applies deep learning to track mitochondrial substructures, accurately identifying drug action pathways and shifting from empirical evaluation to intelligent recognition. ³⁹ Polypharmacology introduces an innovative paradigm for studying the multi-component nature of TCM by analyzing regulatory networks to identify new targets and optimize formulations. ⁴⁰ Gujral's ensemble learning model integrates elastic network regression with transcriptome data and kinase inhibition databases, enabling high-precision identification of cancer-related kinases. ⁴¹

Collaborative Multi-Omics Research

The continuous development of omics technologies has generated large datasets relevant to TCM. ⁴² Applying AI to explore data networks enables systematic analysis of the pathological mechanisms of complex diseases and the pharmacodynamic components of TCM. For example, Li et al constructed a network equilibrium model using ML algorithms, integrating coexpression patterns with network topology features to establish a molecular network for cold-heat syndrome, and subsequently identified biomarkers associated with this syndrome. ⁴³ Through dynamic topology modeling, computational intelligence-driven network pharmacology allows systematic analysis of disease phenotype networks and TCM efficacy, signifying the digital reconstruction of traditional medical theories. ⁴⁴ Advances in sequencing technology have made large-scale transcriptome generation more cost-effective, and transcriptomic data now clarify mechanisms of drug action, promoting the application of systems biology in mechanism studies. ⁴⁵ Clinical omics datasets such as those from TCGA provide critical support for advancing precision therapy. ⁴⁶ These approaches have reconstructed cell differentiation trajectories and clonal evolution pathways, offering new insights into the roles of TCM compounds in the cellular microenvironment. ⁴⁷

AI in the Study of Mechanisms of Action of TCM

AI-Integrated Network Pharmacology

With the advancement of TCM research, network pharmacology has emerged as a crucial approach for multi-target drug discovery. By integrating systems biology, multi-omics, high-throughput data analysis, computer simulations, and complex network analysis, network pharmacology has been effectively applied to TCM research. ⁴⁸ For instance, an integrative in silico study demonstrated its application in identifying active ingredients and elucidating the mechanism of a medicative diet (XQCSY) for the prevention of non-alcoholic fatty liver disease (NAFLD), highlighting its multi-target therapeutic potential. ⁴⁹ In TCM network pharmacology, AI algorithms support the prediction of potential disease targets, analysis of complex interactions, and identification of synergistic effects of combined drugs, thereby enhancing predictive accuracy. ⁴⁴ The integration of AI with network pharmacology has markedly improved the efficiency of pharmacological research in TCM. Given the multi-component and multi-target characteristics of TCM, conventional methods struggle to fully analyze complex effects. ⁵⁰ AI accelerates this process through advanced data processing and pattern recognition (Figure 2). Its large-scale analytical capacity enables rapid screening of potential ingredients and targets, improving both accuracy and efficiency. ⁵¹ Furthermore, AI contributes to toxicity prediction and safety assessment of TCM components, reducing risks during development. The integration of AI and network pharmacology not only enhances multi-target drug discovery but also introduces new methodologies and tools for the modernization of TCM and the advancement of precision medicine. ⁵² Interdisciplinary collaboration will further deepen the understanding of TCM pharmacology, promoting its modernization and global development.

Figure 2.

Comparison of Traditional Chinese Medicine Research and AI-Empowered Methods.

Currently, ML technology has demonstrated significant application value in systematic pharmacology research. As a representative supervised learning model, the support vector machine (SVM) exhibits strong performance in classification and regression tasks and is widely applied in screening drug candidates for complex diseases. ⁹ Based on the holistic perspective of traditional Chinese medicine, Dai and colleagues investigated the pathological mechanisms of Huntington's disease (HD) through protein interaction networks, screened high-affinity active ingredients, and validated their efficacy against core targets using an SVM classifier, among them, the SVM model for the CK2A2 target predicted an R² of 0.97, and the MLR model predicted an R² of 0.96, and all models’ external validation R² was within a reasonable range, reflecting the accuracy of the models’ predictions. ⁵³ This study integrated machine learning with systems pharmacology to identify novel TCM formulations and their effect-related proteins, providing a reference for the development of multi-target therapies. ⁵³ In ensemble learning, random forest constructs predictive models by integrating multiple decision trees, demonstrating clear advantages in multivariate data analysis. Using the FDA-approved drug target database, Wu et al developed a random forest-based predictive framework to identify multi-target active molecules and elucidate the synergistic mechanism of aconitum. This innovative approach offers a new strategy for discovering active ingredients in TCM and expands the potential for developing multidimensional treatment strategies for complex diseases. ⁵⁴

Studies on Deep Learning Extension Compounds and Targets

As a fundamental component of ML, deep learning (DL) has gained significant attention due to its hierarchical information processing capabilities. This approach enables efficient analysis and feature abstraction of high-dimensional data by constructing multi-layer neural network structures. ⁵⁵ Among various deep learning architectures, models such as multi-layer perceptron (MLP), convolutional neural network (CNN), deep neural network (DNN), and recurrent neural network (RNN) exhibit distinct characteristics, making them suitable for different applications. ³⁹ As a basic neural network model, MLP demonstrates strong performance in predicting compound-target interactions due to its fully connected topology. By simulating the computational processes of biological neurons, this network effectively evaluates drug-protein similarity and stability. ⁵⁶ DNNs, developed by increasing the number of hidden layers in MLPs, provide enhanced processing capabilities for complex tasks such as biological activity prediction and gene regulatory network modeling. ⁵⁷ Ji et al developed a compound-protein interaction (CPI) prediction model based on deep neural networks (DNN), which successfully expanded the target number for 46 compounds from Astragalus membranaceus and Hedyotis diffusa from 558 to 698. The model demonstrated impressive performance, achieving an accuracy of 93.31% on a test set containing 1,014,627 CPI data points, significantly outperforming KNN (91.27%), random forests (RF, 90.08%), and XGBoost (91.40%). Moreover, the combination of anti-breast cancer compounds, identified using the expanded target set in cellular experiments, exhibited an IC₅₀ as low as 19.41 μM and a CI value of 0.682, highlighting the model's robust target expansion capability, prediction performance, and practical application value in drug discovery. ⁵⁸ In spatial feature extraction, CNNs offer significant advantages in predicting small molecule-protein affinity through specialized convolutional operations. ³⁰ For sequence data processing, RNNs and their improved variants, such as LSTM networks, effectively capture temporal dependencies, playing a crucial role in virtual drug screening and multi-target drug development. ⁵⁹ For example, Li et al developed a customized compound library for RIPK1 using generative deep learning (GDL) model, which incorporated transfer learning, regularization enhancement, and sampling enhancement strategies, generating a virtual library containing 79,323 molecules. After screening, they identified a novel RIPK1 inhibitor, RI-962, which demonstrated potent inhibitory activity with an IC₅₀ value as low as 5.9 nM. When administered at a dose of 40 mg/kg, RI-962 significantly improved the survival rate of TNF-α-induced systemic inflammatory response syndrome (SIRS) mice, increasing it from 10% to 90%. Furthermore, the compound exhibited favorable pharmacokinetic properties and safety, underscoring the value of this model in discovering novel drugs and highlighting the potential of deep learning in drug development. ⁶⁰ These deep learning models not only extend the scope of ML applications but also provide robust technical support for drug discovery and development.

AI in Optimizing TCM Prescriptions and Mining Compatibility Rules

As a fundamental aspect of disease prevention and treatment in TCM, the concept of drug compatibility emphasizes synergistic effects and toxicity regulation, creating an important intersection with AI technology. ¹⁷ In terms of technical implementation, multi-modal algorithm evaluation has been developed: partial least squares discriminant analysis is employed to examine spectral-effect relationships of compound components, while SVM classification models are used to predict drug efficacy. ⁶¹ The DeepSynergy model was the first to apply deep learning to drug synergy prediction, achieving breakthrough accuracy in predicting anticancer drug synergistic effects by integrating chemical properties of drugs with cell line genomic features. ⁶² Within the framework of systems biology, innovative models continue to emerge. DrugCell achieved interpretable drug response prediction and collaborative design through the ANN–VNN fusion architecture. ⁶³ The TCMFP model quantitatively predicts TCM prescriptions from single-drug effects to combination effects through a three-level scoring system consisting of Hscore (network targeting), Pscore (empirical compatibility), and FmapScore (intelligent optimization). ⁶⁴ Despite notable progress, existing models face several challenges, including assumptions of structural similarity among compounds, limitations in modeling dimensions, and the lack of visualization tools for dynamic multi-component interactions. ³⁰ Addressing these issues requires the development of cross-scale modeling techniques that integrate single-cell sequencing data with dynamic modeling approaches.

Large Language Models in TCM Pharmacology

With the rapid advancement of artificial intelligence and natural language processing, large language models (LLMs) have gained prominence across various fields. ⁶⁵ Through deep learning and unsupervised pretraining on extensive corpora, LLMs effectively simulate human language generation processes, demonstrating strong abilities in both comprehension and text generation. ⁶⁶

In TCM pharmacology, LLMs exhibit unique advantages. By studying and analyzing TCM literature, these models construct comprehensive knowledge bases, providing accurate information on pharmacological effects, chemical components, and formulation methods. ⁶⁷ LLMs facilitate correlation analyses between herbal components and targets, uncover pharmacological mechanisms, and offer valuable references for new drug development. For instance, TCMBank, a TCM database developed at Sun Yat-sen University, integrates data on herbs, bioactive components, and gene targets, significantly advancing studies of TCM molecular mechanisms. ⁶⁸ In addition, the BATMAN-TCM 2.0 database contains nearly 500 herbs and their compound components, focusing on relationships between TCM ingredients and targets, and supporting target prediction and drug discovery. ⁶⁹ TCMNP, a network pharmacology data processing and visualization toolkit developed in R, supports component analysis, enrichment studies, and target research. ⁷⁰ Therefore, LLMs are driving the transformation of TCM pharmacology from experience-driven to data-driven research, although their application must remain grounded in rigorous pharmacological evidence. In the future, combining LLMs with experimental validation will accelerate the modernization and internationalization of TCM.

Conclusion

Through long-term practice, TCM has developed a distinct theoretical framework. However, its evolution is constrained by limited understanding of molecular mechanisms and micro-level principles, resulting in a lack of scientific explanations. These limitations have posed significant challenges to modernization and innovation. Integrating modern science and technology with TCM theory and investigating therapeutic mechanisms can bridge traditional and modern medicine. Future research should focus on applying natural language processing techniques to explore the complex mechanisms and empirical formulas described in ancient literature, while integrating deep learning to construct pharmacological association models for TCM prescriptions. ⁷¹ Furthermore, combining Organ-on-a-Chip (OoC) technology with AI should be emphasized. OoC can simulate the physiological microenvironment of human organs to generate dynamic pharmacodynamic data, ⁷² and AI can model dose-effect relationships of TCM components based on these data to improve the accuracy of in vitro screening.^73,74

Despite its potential, the application of AI in TCM research faces challenges and limitations. Inconsistencies and gaps in TCM data may affect model training accuracy. Moreover, the complexity and “black box” nature of AI models reduce interpretability. In clinical applications, transparent decision-making processes are essential to ensure model safety and effectiveness, limiting broad adoption in medical research. ¹⁶ Therefore, improving data quality and enhancing model interpretability have become key objectives for the future development of AI applications in TCM.