Untargeted metabolomic profiling of CJ and CJ-ELNs was performed by LC-MS. A total of 6 samples (including 3 CJ samples and 3 CJ-ELNs samples) were analyzed using a ultra-high-performance liquid chromatography (UHPLC) system coupled to a Q Exactive HF-X mass spectrometer (Thermo Scientific). Chromatographic separation was performed on an HSS T3 column (maintained at 40°C) with a 12-minute linear gradient from 2% to 98% mobile phase B at a flow rate of 0.3 mL/min. Mass spectrometry (MS) data were acquired in both positive and negative electrospray ionization (ESI) mode (± ESI) using a data-dependent acquisition strategy (top 10 most intense ions). Raw data were first converted to the mzML format using ProteoWizard, followed by processing, using Compound Discoverer 3.3 (Thermo Fisher Scientific) for peak alignment (with maximum retention time shift of 0.5 min and mass tolerance of 10 ppm) and normalization (using the median of maximum peak areas). Compound identification was achieved by matching MS/MS spectra against the following databases: mzCloud, LipidMaps, KEGG, HMDB, and MassBank. The matching criteria were set to a mass tolerance of 10 ppm and a minimum match factor threshold of 10. A four-step filtering workflow was designed to quantitatively identify target compounds as follows (Figure 1).

Common and unique compounds derived from CJ and CJ-ELNs were visualized in a Venn diagram, and a word cloud analysis was conducted via Python. Functions and tools, and databases of key terms used in this study are listed in Table 1.

A multimodal framework of LC-MS combined with the DeepSeek-DF model was created, consisting of two major components of the input and output layers. The input layer integrated structural features of compounds (Morgan fingerprints), quantitative features (z score normalization), and literature-derived features (PubMedBERT embeddings). The core architecture was listed in Figure 2. Additionally, the output layer used multitask learning to simultaneously predict wound-healing activity via Sigmoid output and mechanism category via Softmax output.

The Automated Semantic Recognition (ASR) module and prompt engineering techniques of DeepSeek-R1 32B, as well as web searching were used to interpret the potential biological functions of the screened candidate compound with an annotation of functional labels. A plausibility assessment was then performed based on predefined criteria, including antioxidation, anti-inflammation, anticellular damage, antiapoptosis, wound healing and tissue regeneration, and cell proliferation. Each compound was evaluated and categorized using the following scoring scheme: √ (confirmed), × (not supported), and ? (uncertain). Taking the metabolite (-)-Epicatechin 3-O-gallate as an example, its function, category and possibility in the involvement of wound healing, tissue regeneration, antioxidant, and anticellular damage were predicted via the multimodal framework (Table 2). Following this preliminary filtering, manual curation was conducted to eliminate compounds of nonplant origin and those with low abundances. Ultimately, a refined set of characteristic natural products from CJ-ELNs with potential wound-healing properties was selected.

After conversion and normalization of the raw LC-MS data, a total of 829 and 2212 compounds were identified from CJ and CJ-ELNs. A Venn diagram visualized 1881 specific compounds in CJ-ELNs (Figure 3). “Acid,” as the most frequent term across all entries of metabolite names, was detected by a word cloud analysis (Multimedia Appendix 1). It suggested that acidic compounds were highly enriched in CJ-ELNs.

A total of 1881 candidate compounds enriched in CJ-ELNs were functionally annotated and classified using the self-designed multimodal framework of LC-MS combined with DeepSeek models. They were categorized into 20 distinct classes, including organic compounds, alkaloids, amino acids, biomolecules, organic acids, antioxidants, anti-inflammatory agents, energy metabolism-related molecules, phenolics, cytoprotective agents, alcohols, and others. Organic compounds were the leading category of compounds enriched in CJ-ELNs (Figure 4, Multimedia Appendix 2). Functionally, 43.33% (n=39) of compounds enriched in CJ-ELNs possessed the antioxidant property. With the assistance of DeepSeek, we specifically screened compounds enriched in CJ-ELNs with targeted functions of antioxidation, anti-inflammation, anticellular damage, antiapoptosis, wound healing and tissue regeneration, and cell proliferation.

We estimated the overall expression levels of compounds across the six target functions derived from the DeepSeek model within this multimodal framework, visualizing the results in radar chart format after log2-transformation. (Figure 5). Notably, compounds with the antioxidant function possessed the highest expression levels, proving the antioxidant mechanism of CJ-ELNs in wound repair. Finally, a secondary filtering of compounds with targeted functions was conducted. We manually excluded nonplant–derived compounds, including those of animal origin, synthetic chemicals, and other nonbotanical sources. In addition, compounds with low expression levels in CJ-ELNs were also removed. As a result, a total of 46 naturally active compounds derived from CJ-ELNs with targeted functions were identified (Figure 6 and Multimedia Appendix 3). Citric acid was the most abundant compound with the targeted functions, which was consistent with the finding from the word cloud analysis.

This study innovatively integrated DeepSeek models with LC-MS to successfully predict the major natural products of CJ-ELNs responsible for wound healing. DeepSeek’s ASR semantic recognition and prompt engineering worked together to generate initial classification labels. Moreover, an automatic assessment effectively, rapidly, and accurately achieved the goal of data mining of specific compounds for targeted functions.

AI techniques, particularly LLMs, have become an unstoppable force for reshaping medical research [19,20]. Traditionally, LC-MS is a powerful analytical technique to identify and quantify active ingredients in traditional Chinese medicine (TCM) herbs. However, a rapid and accurate recognition of compounds with targeted functions, and a quantitative analysis of trace concentrations in complicated samples can be challenging [21]. We expected that an integration of LC-MS and LLMs would benefit TCM research, including the acceleration of active ingredient screen, precise targeting of interested compounds for certain diseases, and anchoring the promising candidates for developing new drugs. DeepSeek is an intelligent system based on a large-scale pre-trained language model, exhibiting strong capabilities in text understanding, knowledge reasoning, and cross-modal collaborative analysis, particularly excelling in processing information within Chinese-language contexts [22,23]. It enables rapid processing and analyzing massive volumes of both unstructured and structured data, thus digging biological insights out of complex omics datasets [24,25].

In the present study, we first created a four-step filtering workflow and quantitatively identified target compounds from CJ-ELNs by LC-MS. The cloud word analysis emphasized the term of acid among screened compounds enriched in CJ-ELNs. Acidic compounds derived from traditional Chinese herbals are established for the role of clearing heat and detoxifying [26]. Numerous studies have reported that acidic compounds in plants exert antioxidant, antibacterial, and anti-inflammatory effects through mechanisms such as scavenging free radicals, alleviating oxidative stress, modulating inflammatory factors, stimulating fibroblast proliferation, promoting collagen deposition, enhancing epithelialization, and inducing angiogenesis [27,28]. To achieve a precise data mining of compounds with relevant functions, DeepSeek models lent a hand that specifically screened compounds in CJ-ELNs with targeted functions of antioxidation, anti-inflammation, anticellular damage, antiapoptosis, wound healing and tissue regeneration, and cell proliferation. Finally, naturally active compounds in CJ-ELNs were resurfaced for their promising potentials in wound repair. For example, studies have shown that baicalin accelerates the wound healing process by downregulating the expression of pro-inflammatory cytokines (IL-6 and IL-1β) while upregulating the anti-inflammatory factor IL-10, and by promoting the secretion of various growth factors (VEGF, FGF-2, PDGF-β, and CTGF) [29]. The combination of LC-MS with DeepSeek paves the way to further analyses of therapeutic targets from traditional Chinese herbs for wound healing and tissue regeneration [30,31].

Limitations in this study should be noted. Firstly, bioactive compounds derived from CJ-ELNs were mined via LC-MS and a single LLM, namely, DeepSeek-R1. Other cutting-edge LLMs such as Claude, GPT-4 and Liama [32] can be further analyzed for the assistance of LC-MS in identifying interested compounds. Secondly, the 46 naturally active compounds derived from CJ-ELNs with targeted functions should be validated in in vivo and in vitro experiments. Lastly, the workflow we have established requires further validation on independent datasets. We shall address the aforementioned issues in subsequent work, including evaluating the efficacy of compounds through cell migration and transdermal tissue compatibility assays, verifying their efficacy via macroscopic imaging and H&E staining following animal wound modelling interventions, and validating potential pathways involved through Western blot and immunohistochemical analysis.

We innovatively designed a multimodal framework of LC-MS combined with DeepSeek models that rapidly and accurately identify naturally active compounds from CJ-ELNs. This AI-powered system innovatively integrates the traditional analytical technique with modern large language models, showing a huge potential in modern medicine and TCM research.