Decoding enzymatic landscapes: a knowledge graph–enhanced large language model framework for microbial enzyme production and catalysis systems

Abstract

Microbial enzyme production and catalysis systems are crucial aspect of biotechnological research. However, building them from trustworthy published experimental data presents a major obstacle for both manual and automated techniques. Here, we introduce MEPAM (Microbial Enzyme Production and Catalytic Activity based on LLM), a question-answering system designed to accurately address inquiries related to enzyme production and catalytic reactions. Specifically, by training three machine learning models with >0.98 accuracy, we identified 11,068 high-quality, relevant articles from the Web of Science. Leveraging DeepSeek-V3 with zero-shot learning, we developed an ontology-driven knowledge representation that extracted 12,434 entities and 35,918 relations with 0.78 extraction accuracy and constructed a structured knowledge graph. Compared to few-shot learning and other machine learning methods, our framework achieved significantly higher extraction accuracy. Using this framework, we developed MEPAM based on retrieval-augmented generation and prompt engineering. Finally, using MEPAM, we extracted a comprehensive network involving the expression profiles, precise culture conditions, and substrate preferences for cellulase, demonstrating the strong utility of this tool. Compared with traditional LLMs, particularly GPT-4o, MEPAM exhibited superior performance, achieving significantly higher answer accuracy (0.86 vs. 0.52) and nearly eliminating hallucinations. MEPAM is available at http://180.76.108.212. This framework provides context-rich, verifiable insights, thus bridging predictive modeling with experimental validation to facilitate the exploration of microbial enzymatic systems.

Graphical abstract

Highlights

• •MEPAM deciphers Enzymatic Landscapes using host profiles, culture conditions & substrates.
• •MEPAM shows higher answer accuracy (0.85 vs. 0.52) and nearly hallucinations-free.
• •DeepSeek-V3 hit 0.78 accuracy in triple extraction, surpassing fine-tuned SciBERT.
• •A knowledge graph was constructed from 11,068 relevant, high-quality articles.
• •A proof-of-principle case study is conducted for a cellulase network.

1. Introduction

Enzymes serve as highly efficient biocatalysts [1], with broad applications spanning the industrial, agricultural, and environmental sectors [2]. These applications generally require the enzymes to exhibit high soluble expression [3] and high catalytic activity [4]. Despite extensive research focused on identifying hosts for high-efficiency enzyme production, optimizing fermentation parameters, and precisely adjusting enzymatic reaction conditions, a comprehensive, systems-level understanding of the network among these factors controlling microbial enzyme production and catalytic activity is still lacking.

As expertise varies widely among individuals, a knowledge network integrating published, high-quality research findings could serve as a powerful research tool for both scientific and practical applications. Currently, the query “[microorganism] AND [enzyme]” yields over a million results in the Web of Science database alone, suggesting an overwhelming abundance of research articles related to microorganisms, enzymes, and substrates. However, identifying high-quality, relevant articles specifically focused on microbial enzyme production and catalytic activity can pose a serious challenge, even with continuous refinements to the search queries. Exacerbating this issue, life science research language employs complex and sometimes redundant terminology, a wealth of topic-specific abbreviations, and abstract or vague concepts. For instance, the same microorganism can have different names in different databases: the organism called Bacillus subtilis 168 in NCBI is listed as DSM402 in the German Collection of Microorganisms and Cell Cultures (DSMZ). Furthermore, terms such as “xylose” and “cellulose” can describe both growth medium components and enzymatic reaction substrates, increasing the potential for confusion.

Historically, aggregating sufficient knowledge to construct a microbial enzyme production and catalysis network has required extensive, manual screening and annotation of published literature to identify relevant factors, such as microbial hosts, validated enzymes, appropriate substrates, fermentation conditions, and reaction parameters. As a result, statistical analyses or modeling across studies using tools such as knowledge graph-based approaches [5,6] commonly suffer from small dataset coverage, deficiencies in information extraction, and labor-intensive data collection from the literature [7,8], rendering large-scale comparative analyses impractical. Recent advances in generative large language models (LLMs), such as DeepSeek [9,10], Qwen [11,12], LLaMA [13], and GPT [14,15], provide a potential alternative. However, although these models are capable of providing a grammatically correct answer to virtually any question, including those related to scientific research, the responses may originate from unreliable sources or hallucinatory outputs [16,17], which raises concerns about the potential for generating inaccurate or even harmful information [[18], [19], [20], [21]]. Despite these limitations, models such as ChatExtract, developed by Morgan and colleagues [22], and biomedical natural language processing (BioNLP) techniques [23] have demonstrated remarkable capability for extracting complex data and appear to have strong contextual comprehension skills. To fully leverage the strengths of LLMs while mitigating their limitations, knowledge graph–enhanced LLMs (KG-LLMs) have emerged as a viable strategy for synergizing the strengths of both approaches while overcoming the above limitations of knowledge graphs and LLMs [24]. The use of well-engineered prompts might further improve the accuracy of general-domain LLMs [25].

Here, we built MEPAM (Microbial Enzyme Production and Catalytic Activity based on LLM), a question-answering system designed to accurately address inquiries about microbial enzyme production and catalytic activity based on high-quality, published data. Compared with other current generative LLMs, MEPAM is literature-traceable, hallucination-free, and capable of providing accurate answers to user queries. This novel tool is available at http://180.76.108.212. Our versatile framework (KG-LLMs) could also be used to build domain-specific intelligent systems across other scientific fields.

2. Results

2.1. Development and application pipeline of the MEPAM question-answering system

To facilitate investigations of microbial enzyme production conditions and catalytic activity networks that require thorough analysis of multi-source data, we constructed MEPAM, a question-answering system based on a knowledge graph–enhanced LLM. To construct the model, we initially formulated a relatively comprehensive search query, which retrieved 77,087 potentially relevant articles from the Web of Science database since 2000. From these hits, three models, BioBERT [26], SciBERT [27], and XGBoost [28] were constructed to obtain a refined corpus of 11,068 publications, which were then converted into machine-readable text using GROBID (GeneRation Of BIbliographic Data) and CERMINE (Content ExtRactor and MINEr) (Fig. 1A).

Fig. 1.

Workflow for knowledge graph-augmented LLM construction. A Article retrieval and cleaning, resulting in a curated collection of 11,068 high-quality, highly relevant articles. B Screening, evaluation, and triplet extraction using large language models. C Construction of the knowledge graph database and vector dataset. D Development of the question-answering system.

Next, to investigate microbial enzyme production conditions and catalytic activity networks, we established three core nodes (microorganism, enzyme, and substrate) connected by two key relationships (microbial enzyme production and catalytic conditions for substrate degradation). This framework forms two fundamental triplets: (Microorganism)-[Fermentation Conditions (attributes: pH, temperature …)]→(Enzyme) and (Enzyme)-[Enzymatic Parameters (attributes: temperature, ions …)]→(Substrate). We then employed DeepSeek-V3 with zero-shot learning to extract the structured triplets from the Abstract and Results sections of all 11,068 articles, which resulted in a set of 35,918 triplets stored in the knowledge graph (Fig. 1B and C). At the same time, we used BGE-M3:567m to generate embeddings from all 11,068 articles, which were subsequently stored in a vector database. Finally, we developed MEPAM, a question-answering system based on Retrieval-Augmented Generation (RAG) that utilizes prompt engineering and a chatbot interface (Fig. 1D).

2.2. Collection and screening of research articles

To ensure data traceability and accuracy, we first designed a systematic strategy for data acquisition and screening. Based on the Web of Science database, we formulated the following search query template: “[Fermentation] AND [soluble expression] AND [enzyme] AND [Bacillus subtilis OR Escherichia coli OR Saccharomyces OR Trichoderma reesei OR Aspergillus]” and restricted the search to the “Topic”, “Title”, and “Abstract” of original English research articles published between 2000 and 2025. This initial search retrieved 77,087 candidate articles (Fig. 2A). Their journal impact factor (IF) distribution is summarized in Fig. S1; 2804 articles were published in journals with IF > 10. These articles, forming the basis for knowledge graph development, spanned 4356 journals, the most frequent of which were Applied and Environmental Microbiology (1,905), Journal of Biological Chemistry (1,891), and Journal of Bacteriology (1,871) (Table S1).

Fig. 2.

Article collection and filtering for knowledge graph construction. A Preliminary screening of the Web of Science database using search queries. B Performance evaluation of the three classification models on the test set. Error bars indicate standard deviations (SD). C Filtering the article dataset using the three classification models to determine the final article collection. D The distribution of journal impact factors for the 11,068 articles.

To identify articles from this initial collection that were highly relevant to our study, we trained three classification models (BioBERT, SciBERT, and XGBoost) using a balanced training set, which consists of abstracts of 300 positive articles and 900 negative articles. All three models demonstrated strong performance in the test set (ACC, AUC, and F1 score >0.98) (Fig. 2B). When applied to the pool of 77,087 candidate publications, the models predicted 51,211 (BioBERT), 40,816 (SciBERT), and 31,216 (XGBoost) positive articles, respectively (Fig. 2C). To ensure the reliability of these studies, we conservatively retained only 11,068 articles that were unanimously identified by all three models for our final dataset (Fig. 2C). The selected articles covered a broad range of journal IFs (Fig. 2D), including 263 articles published in journals with IF > 10. These articles spanned 88 journals, among which papers from Applied and Environmental Microbiology (332), Bioresource Technology (279), and Applied Microbiology and Biotechnology (198) occurred with the highest frequency (Table S2). This rigorous multi-model consensus approach guaranteed that inputs for knowledge-graph construction were both accurate and representative.

2.3. Screening, evaluation, and information extraction of large language models

To evaluate the capacity of current state-of-the-art general LLMs (i.e., DeepSeek-V3, Qwen2.5-72B, Llama3.3-70B, and GPT-4o) for zero-shot triplet extraction, we manually curated 100 articles from the pool of 11,068 publications. For each article, we manually extracted textual information from the title, Abstract, and Results sections individually and annotated entities (e.g., microorganisms, enzymes, substrates) and relationships (e.g., fermentation conditions, enzymatic reaction parameters) using BRAT (v1.3), which yielded 422 triplets (Fig. 3A). We then performed structured data extraction from the 100 curated articles with the OntoGPT framework using the four LLMs mentioned above, generating outputs in YAML format. These models yielded 412, 402, 437, and 482 triplets, respectively (Table S3). To evaluate the consistency of the results, we compared the model-generated triplets against the manually annotated ground truth triplet set. This analysis, conducted via Qwen2.5-14B, demonstrated that DeepSeek-V3 (linkML) achieved the highest average scores across all metrics: Precision (0.78), Recall (0.78), and F1 score (0.76). The other three models showed relatively poor performance, with all three metrics ranging between 0.5 and 0.6. Notably, all linkML-based models outperformed their non-linkML counterparts (Fig. 3B and Fig. S2–S4). Additionally, compared to the fine-tuned machine learning-based NLP model SciBERT, DeepSeek-V3 (linkML) also demonstrated significant advantages across all three evaluation metrics, achieving 41.8%, 59.2%, and 55.1% improvements in Precision, Recall, and F1-score, respectively (Fig. 3B).

Fig. 3.

Extraction of triplets from articles using DeepSeek-V3. A Triplet schema. B Evaluation of extraction accuracy by the four LLMs and SciBERT. C Example of triplet extraction from an abstract using DeepSeek-V3.

A comparison of the computational efficiency and cost structure of these four LLMs revealed that DeepSeek-V3 offered the most economical pricing at ¥0.5 per million tokens (64K context length), followed by Llama3-3:72B (¥0.8; 128K), Qwen2.5-72B (¥4; 128K), and GPT-4o (¥18.1; 128K). All pricing information reflects publicly available model usage costs as of April 2025. DeepSeek-V3 (0.42 h) also demonstrated the fastest processing time in extracting triplets from the 100 research articles, followed by Llama3-3:72B (0.75 h), GPT-4o (0.83 h), and Qwen2.5-72B (1.92 h) (Table 1). Considering its accuracy, throughput, and cost, we selected DeepSeek-V3 for large-scale triplet extraction from all 11,068 articles following document preprocessing through the GROBID and GERMINE pipelines, which converted PDFs to structured XML files with 0.94 (Abstract) and 0.89 (Results) accuracy (Fig. S5), ensuring reliable text mining inputs.

Table 1.

Comparison of the computational efficiency and cost of LLMs.

LLM	Price (¥per million tokens)	Context length	Processing time (hours per 100 articles)
DeepSeek-V3	0.5	64K	0.42
Llama3-3:70B	0.8	128K	0.75
Qwen2.5-72B	4.0	128K	1.92
GPT-4o	18.1	128K	0.83

For instance, the following statements were obtained from the text of an abstract (Fig. 3C).

"1. A thermostable l-asparaginase was produced from Bacillus licheniformis UDS-5.

2. It secreted l-asparaginase optimally in the medium, pH 7, containing 0.5% (w/v) peptone, 1% (w/v) sodium chloride, 0.15% (w/v) beef extract, 0.15% (w/v) yeast extract, 3% (w/v) l-asparagine at 50 °C for 96 h.

3. It displayed optimal catalysis at 70 °C in 20 mM Tris-Cl buffer, pH 8."

DeepSeek-V3 successfully extracted the following.

• •Microorganism: Bacillus licheniformis UDS-5
• •Effect: optimal, temperature: 50 °C, pH: 7, incubation period: 96 h, medium: 0.5% (w/v) peptone, 1% (w/v) sodium chloride, 0.15% (w/v) beef extract, 0.15% (w/v) yeast extract, 3% (w/v) l-asparagine
• •Enzyme: l-asparaginase
• •Effect: optimal, ions: Ca²⁺, temperature: 70 °C, pH: 8
• •Substrate: l-asparagine

In total, the pipeline extracted 12,434 entities, including 1842 microorganisms, 7310 enzymes, and 3282 substrates, with 35,918 relations, comprising 19,386 fermentation conditions and 16,532 enzymatic catalytic conditions.

2.4. Analysis of microorganism and enzyme entities

To conduct a systematic, in-depth analysis of information extracted from published articles by DeepSeek-V3, we first collected all extracted microorganism and enzyme entities. The microorganism entities encompassed 170 genera and 1842 species (Table S4–S5). Analysis of triplet data revealed that the most prevalent microorganisms were Saccharomyces spp. (32.0%), followed by Escherichia spp. (28.0%) and Trichoderma spp. (12.1%). Other genera, including Bacillus, Aspergillus, Lactobacillus, Pichia, Streptomyces, and Clostridium, were present at relatively low frequencies. At the species level, Saccharomyces cerevisiae dominated the Saccharomyces spp. group (95.7%), while Escherichia coli accounted for nearly all occurrences within Escherichia spp. Similarly, Trichoderma reesei represented the vast majority (96.0%) of Trichoderma spp. (Fig. 4A). These results suggest that S. cerevisiae, E. coli, and T. reesei have significant potential as fermentation host strains for the production of enzymes.

Fig. 4.

Systematic analysis of microorganism and enzyme entities extracted from published articles. A Distribution of extracted microorganism entities. B Distribution of extracted enzyme entities. C Proportions of the most common enzymes studied in the three most common hosts. D Proportion in which each host from (C) appears in studies of the three most commonly produced enzymes.

We then classified the 7310 enzyme entities identified in the triplet data (Table S6) into six functional categories for quantitative analysis, including oxidoreductases, transferases, hydrolases, lyases, isomerases, and ligases, with hydrolases (40.9%) and oxidoreductases (26.0%) constituting the major categories (Fig. 4B). Among hydrolases, cellulose-degrading enzymes were particularly abundant, with cellulases representing 36.4% of hydrolases, followed by β-glucosidases (19.3%), while the oxidoreductase category was dominated by alcohol dehydrogenase (25.2%) and xylose reductase (14.7%) (Fig. 4B).

Sorting enzymes by host revealed a preponderance of cellulose-degrading enzymes among fungal hosts. S. cerevisiae primarily produced cellulase (4.1%) and β-glucosidase (3.6%), in addition to the xylose metabolism enzyme xylitol dehydrogenase (3.5%). T. reesei predominantly produced cellulase (55.5%), β-glucosidase (8.4%), and cellobiohydrolase (9.3%). By contrast, E. coli produced a more diverse range of enzymes (0.05%–3.67% each), likely due to its well-studied genetics, high genetic tractability, fast growth, and low cultivation costs (Fig. 4C). We sorted hosts according to enzyme production, showing that 66.8% of cellulases were produced by T. reesei, while 10.2% were produced by S. cerevisiae. T. reesei (28.5%) and S. cerevisiae (27.5%) were employed with roughly equal frequency to produce β-glucosidase. By contrast, 46.4% of alcohol dehydrogenases and 63.4% of xylose reductases were produced by S. cerevisiae, followed by 30.4% and 7.1%, respectively, produced by E. coli (Fig. 4D). These results demonstrate that the production of cellulose-degrading enzymes, particularly cellulase and β-glucosidase, in T. reesei or S. cerevisiae, and alcohol dehydrogenase production by S. cerevisiae, were dominant research topics across the 11,068 collected articles [[29], [30], [31], [32], [33]]. The primary application involved the conversion of cellulosic agricultural waste (e.g., corn stover, sugarcane bagasse) into fermentable sugars for subsequent bioethanol production.

2.5. Analysis of fermentation conditions and enzymatic reaction parameters

To explore information about microbial enzyme production and catalytic activity networks obtained by our triplet extraction model, we analyzed the fermentation conditions and enzymatic reaction parameters for each of the above entities. Specifically, DeepSeek-V3 with zero-shot learning extracted 19,386 non-redundant fermentation conditions and 16,532 non-redundant enzymatic reaction parameters (Table S7–S8). As an example, to illustrate how networks regulate microbial enzyme production and catalytic activity, we examined cellulase fermentation by T. reesei and cellulase hydrolysis of different substrates.

Fermentation conditions for cellulase production by T. reesei typically involved moderate temperatures of ∼30 °C and weakly acidic pH (4.5–6, commonly 5). However, temperatures of 28 °C and 50 °C were also frequently reported (Fig. 5A). Cellulose, lactose, glucose, and Avicel served as the primary carbon sources, while ammonium sulfate, urea, and ammonium sulfate were the most common nitrogen sources (Fig. 5B). Accordingly, fermentation media were frequently enriched with these components (Fig. 5C). Additionally, fermentation durations were mostly concentrated at 48, 72, 120, 144 and 168 h, with a moisture content of 70%. Other parameters varied widely, including agitation speed (120–500 rpm), aeration rate (0.3–4 vvm), and fermentation volume (100 mL–5 L). The maximum cellulase yield reported in these studies reached 100 g/L, and the highest recorded enzyme activity was 296 U/(mL·min) (Fig. 5D).

Fig. 5.

Microbial enzyme production and catalytic activity network analyses. A Fermentation temperature and pH. B Carbon and nitrogen sources. C Media and growth supplements. D Fermentation parameters. E-H Hydrolysis conditions for lignocellulose, cellulose, CMC, or Avicel substrates with optimal temperature, pH, substrate, and/or activity.

We also evaluated cellulase catalytic activity towards four substrates: lignocellulose, cellulose, carboxymethyl cellulose (CMC), and Avicel. Overall, cellulose was the preferred substrate and had the highest number of experimentally tested catalytic reaction parameters. The optimal conditions for cellulase hydrolysis of cellulose were 30–50 °C and pH 5, demonstrating characteristics of mesophilic, weakly acidic cellulases (Fig. 5F). Similar trends were observed for lignocellulose and CMC (Fig. 5E and G). By contrast, Avicel hydrolysis was optimal at 30, 50, 60, and 80 °C under weakly acidic conditions, indicating thermotolerant enzyme activity (Fig. 5H). However, relatively few studies documented substrate concentration and absolute enzyme activity.

2.6. Construction and comparison of the question-answering system

To enable efficient knowledge retrieval and management of complex relationships, we developed a retrieval-augmented generation (RAG)-based question-answering system using the 11,068 research articles mentioned above. For reliable performance, the extracted triplets were stored in the NebulaGraph graph database, while article metadata (titles, DOIs), Abstracts, and Results sections were embedded and indexed in a vector database. The system processes queries through a structured pipeline: First, Mistral 8x7B extracts key terms from the input question and retrieves semantically relevant document snippets from the vector database. These terms are then converted into Cypher queries via a predefined text2cypher module in order to fetch associated entities and relationships from NebulaGraph. Finally, the resulting graph-structured data and vector search results, along with their sources, are fed as context into Mistral 8x7B to generate accurate, traceable answers (Fig. 1D). This integrated approach ensures precise responses and provides provenance links to the original article. MEPAM is available at http://180.76.108.212.

To validate the accuracy and effectiveness of the MEPAM question-answering model, we compared it with several established large language models (LLMs) such as DeepSeek-V3, GPT-4o, Qwen2.5-72B, and Mixtral 8x7B. Additionally, we conducted an ablation setup using only the Mixtral 8x7B model (denoted as Mixtral 8x7B (No KG)) to evaluate its memorization and reasoning capabilities without external knowledge retrieval. MEPAM outperformed all other models in terms of Exact Match, Precision, Recall, F1-score, and Faithfulness. MEPAM achieved scores of 0.73, 0.84, 0.85, 0.84, and 0.93 respectively, surpassing the best-performing baseline LLM, GPT-4o, which scored 0.37, 0.55, 0.55, 0.54, and 0.67 for the same metrics (Fig. 6A–E, Table S9–S14). MEPAM had almost no hallucinations, while GPT-4o exhibited a hallucination rate of 0.33. The other models, DeepSeek-V3, Qwen2.5-72B, and Mixtral 8x7B, showed higher hallucination rates (>0.50; Fig. 6F, Table S9–S14). Due to the retrieval-augmented mechanism, MEPAM has a slightly longer response time than other large language models (approximately 5 s for MEPAM compared to approximately 1 s for the other LLMs), which ensures the accuracy and effectiveness of its answers. Above all, these results indicate that MEPAM is a reliable tool, with minimal hallucinations, for accurately answering questions related to microbial enzyme production and catalytic activity networks.

Fig. 6.

Comparison of MEPAM with other traditional large language models. A–F show comparative analysis between MEPAM and traditional large language models in terms of Exact match (A), Precision (B), Recall (C), F1 score (D), Faithfulness (E), and Hallucination rate (F). Asterisks indicate P-values, as determined by t-test: ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001.

For instance, to address the input query, “What is the optimal fermentation pH for producing cellulase using Aspergillus niger or Trichoderma reesei?”, MEPAM first extracts the key search terms (“optimal”, “pH”, “producing”, “cellulase”, “Aspergillus niger”, and “Trichoderma reesei”) using text2cypher. These terms are then converted into a NebulaGraph Cypher query:

MATCH (m:microbiome)-[p:produce]->(e:enzyme)

WHERE e.enzyme.label == ‘cellulase’ AND (m. microbiome.id == ‘Aspergillus niger’ OR m. microbiome. id == ‘Trichoderma reesei')

RETURN p.ph.

The graph database returns: {'p.ph’: ['4.0 and 5.0']}. This result, combined with semantically relevant vector search results, is then fed as context into Mixtral 8x7B to generate the final response: “The optimal pH range for producing cellulase using Aspergillus niger or Trichoderma reesei is 4.0 and 5.0.” By comparison, the direct Mistral 8x7B output (“The optimal pH range … using Aspergillus niger … between 4.5 and 6.0, the optimal pH for Trichoderma reesei … within a range of 4.5 to 6.0.”) shows that the MEPAM system provides more precise, experimentally supported answers (Fig. S6A).

In another example, when querying MEPAM with question such as, “Which strains are suitable for the production of mannanase?”, it provides a more comprehensive response than the Mistral 8x7B base model's answer. In addition to listing suitable host strains, MEPAM offers comparative analysis between different hosts and includes relevant production yield data (Fig. S6B). When a query falls outside the retrieval scope of MEPAM or the retrieved content is insufficient to provide a complete answer, the system transparently presents Mistral 8x7B's response after outputting the retrieved results, along with the disclaimer: “This information may be incomplete. Additional context from LLM”, ensuring that users can distinguish between knowledge graph–enhanced and LLM-generated answers.

3. Discussion

In this study, we developed MEPAM, a question-answering system that addresses questions related to microbial enzyme production and catalytic activity networks. This tool mitigates hallucinations commonly reported in LLM-only based systems to provide accurate answers, while also tracing responses back to the original article, offering a reliable foundation for experimental design and further research.

In the present study, we integrated fermentation conditions for enzyme production into the “produce” relationship category and integrated catalytic reaction parameters into the “catalysis” relationship. This holistic storage approach ensures data completeness while enhancing retrievability and usability, allowing researchers to quickly access relatively comprehensive details about fermentation and catalytic conditions. However, due to language complexity in biological sciences, as mentioned in the Introduction, our dataset contains 37 distinct terms for Escherichia coli and 26 terms for Bacillus subtilis (Table S15), including abbreviations, subspecies names, and strain identifiers from different databases. Such inconsistencies create discrepancies between manually curated and LLM-extracted data, which may contribute to the observed reduction in precision metrics. For entity disambiguation, in the absence of directly applicable reference databases, we implemented an extended-context approach utilizing a 64K-token window that incorporated complete Abstract and Results sections from each article. This methodology enables the LLM to effectively utilize comprehensive semantic context for enhanced entity resolution accuracy.

While many studies have demonstrated that fine-tuned or few-shot LLMs typically achieve higher accuracy than zero-shot approaches [23,25], we argue that zero-shot learning may be more suitable for analyzing microbial enzyme production and catalytic activity networks because complex terminology, ambiguous abbreviations, and inherently vague concepts prevalent in biological research often result in substantial variation between descriptions of similar studies. Moreover, zero-shot learning provides broader generalizability, transferability, and adaptability compared to few-shot approaches [[34], [35], [36]], enabling rapid knowledge graph construction through schema adjustments. Additionally, zero-shot learning helps mitigate hallucinations [37]. During preliminary tests of few-shot learning in our study, we observed that example inputs stating an optimal growth temperature of 37 °C resulted in some models generating false responses that repeated this parameter, even when unsupported by source text.

Our initial goal was to consolidate the most commonly reported fermentation conditions for producing a given target enzyme in a specific host, as well as the most frequently used substrate types and reaction conditions to maximize enzyme yield and activity during enzymatic catalysis. For example, we analyzed the optimal conditions for cellulase production by T. reesei and determined that moderate temperatures (∼30 °C) and weakly acidic (pH 5) conditions were the most frequently employed parameters. Similarly, the optimal conditions for cellulose hydrolysis by cellulase were 30–50 °C and pH 5. In addition, we summarized the general expression conditions and enzymatic reaction parameters for most enzymes based on the eukaryotic and prokaryotic production hosts, as presented in Table S16 and Table S17. However, we encountered limited absolute data on enzyme yields or activities in the articles. Most reports describe only relative improvements (e.g., “increased yield” or “enhanced activity”) based on their respective baselines (e.g., wild-type enzymes), making it difficult to assess current industrial production levels. To systematically extract meaningful fermentation and reaction conditions that could help improve yield or activity, we introduced an “effect” parameter in both relationship categories, including increase/decrease/optimum categories for fermentation conditions and enhance/inhibit/stable/maximum categories for enzymatic reaction parameters. Only data entries containing these specified effect terms were retained for further analysis.

In conclusion, MEPAM provides a high-accuracy and traceable question-answering tool for exploring microbial enzyme production and catalytic activity networks. Moreover, the system demonstrates the potential of integrating complex biological information through knowledge graphs and question-answering frameworks, offering broad prospects for future research and applications in the field.

4. Materials and methods

4.1. Optimized model construction and article screening using BioBERT, SciBERT, and XGBoost

The training framework for Transformer-based BioBERT and SciBERT maintained consistent hyperparameters by employing the AdamW optimizer [38] for fine-tuning over 10 epochs with a learning rate of 5e⁻⁵, adjusting batch sizes to 8–16 based on GPU memory constraints, and setting the maximum sequence length to 512 tokens to preserve critical contextual information. For XGBoost, TF-IDF vectors (1–3-gram, 3000 features) were generated from preprocessed text after removing HTML tags and filtering with NLTK's English stopword list, using model architecture specifically configured with max_depth = 6 and n_estimators = 1800, which was optimized through a grid search to complement the Transformer models' characteristics. All three models were trained on the same dataset, which consists of abstracts of 1200 articles (300 positive and 900 negative articles), with performance evaluated using 10-fold cross-validation of metrics including accuracy, AUC-ROC, recall, and F1 score.

4.2. Retrieval of full-text PDFs and impact factors

The full-text PDFs of the 11,068 articles were supplied by Chongqing VIP Information Co., Ltd. PDF files were converted into structured XML format using GROBID (https://github.com/kermitt2/grobid) and CERMINE (https://github.com/CeON/CERMINE). Python scripts were then employed to parse the XML files, extracting the contents of the “Abstract” and “Results” sections through identification and matching of section headers. The extracted text was subsequently saved in plain-text format.

Concurrently, journal impact factor data were obtained from the Web of Science export files. A customized Shiny application (https://leqdaydayup.shinyapps.io/leqdaydayup_wos/) was utilized to merge and deduplicate batch-downloaded records while standardizing journal names.

4.3. Schema construction for entities and relationships

Three core entities (microorganisms, enzymes, and substrates) and two key relationships (fermentation conditions and enzymatic reaction parameters) were defined, establishing fermentation conditions as attributes linking microorganisms to enzymes and enzymatic reaction parameters as attributes connecting enzymes to substrates for the construction of a comprehensive entity–relationship schema. Using the DSPY (Data Schema Processing for YAML) framework (https://github.com/stanfordnlp/dspy), schema descriptions were incorporated to generate a YAML configuration file interpretable by large language models (LLMs). This YAML file provides detailed definitions and constraints for all entities and relationships, including strict requirements to preserve the original text content verbatim during information extraction, prohibiting any modifications, interpretations, or generative outputs to prevent hallucination. The schema implementation enforces precise content extraction boundaries while maintaining contextual fidelity through explicit entity–relationship mapping, attribute value constraints, and source-text preservation protocols.

4.4. Manual annotation of entities and relationships

One hundred articles were randomly selected 11,068 publications for expert annotation using BRAT (https://github.com/nlplab/brat). The Abstract and Results sections of these articles were processed, with entity–relationship triplets manually labeled. The annotation schema represented microorganisms, enzymes, substrates, fermentation conditions, and enzymatic reaction parameters as distinct entities (visually differentiated by color coding). All annotated entities were exported from BRAT as.ann files and converted into YAML files with a standardized structure through manual labeling: each file first enumerated all head and tail entities, followed by complete triplet representations (e.g., (Microorganism) - [Fermentation Conditions (attributes: pH, temperature …)] → (Enzyme); (Enzyme) - [Enzymatic Parameters (attributes: temperature, ions …)] → (Substrate)). Files were systematically named using DOIs to enable subsequent model performance evaluation through DOI-based matching.

4.5. Entity and relationship extraction using large language models

To systematically extract structured triplet information, four state-of-the-art LLMs were implemented through API integration, thereby circumventing local hardware limitations: GPT-4o (via OpenAI API), Llama3-70B (via Lepton platform), Qwen2.5-72B (via Alibaba Cloud's Bailian), and DeepSeek-V3 (via DeepSeek API). The DSPY-generated YAML schemas were incorporated into the OntoGPT framework to standardize the extraction pipeline. Each model processed the input texts to produce structured YAML outputs containing triplets (microorganisms, enzymes, substrates, fermentation parameters, and enzymatic reaction parameters). All extracted knowledge units were persistently stored as YAML files indexed by DOI, maintaining traceability for downstream validation.

4.6. Comparative evaluation of large language models in triplet extraction

For systematic comparison, an automated evaluation pipeline was developed that parsed YAML files using Python scripts to decompose triplets into head entities (microorganisms and enzymes), tail entities (enzymes and substrates), and relational attributes (fermentation conditions and enzymatic reaction parameters); structurally align these components into formalized (head entity, relationship, tail entity) representations; and perform DOI-indexed matching between manually annotated and model-generated triplets. The Qwen2.5-14B model served as the evaluation arbiter, which received specially designed prompts.

The OntoGPT framework was evaluated through comparison of schema-guided extraction (LinkML YAML) with direct prompt-based extraction (natural language descriptions + example YAML) using identical LLM backends (GPT-4o/Llama3-70B/Qwen2.5-72B/DeepSeek-V3). Both groups were subjected to identical evaluation protocols measuring structural compliance (valid YAML schema adherence), content fidelity (attribute-level matching with LLM arbitration), and overall extraction performance (precision recall metrics across all entity–relation types).

4.7. Comparative evaluation of the fine-tuned SciBERT

The manually annotated data from 100 articles were converted into BIO format using Python scripts, which performed tokenization and generated corresponding label files. The data were then divided into 5 folds for cross-validation, with each fold containing 20 randomly selected articles as the test set, while the remaining articles were split into training and validation sets at a 1:4 ratio, ensuring that each article appeared in the test set exactly once. SciBERT [27], which was downloaded from the HuggingFace Transformers library, was fine-tuned using the AdamW optimizer [38] with a learning rate of 5e⁻⁵, a batch size of 16, and 10 training epochs. To address long text truncation, a sliding window approach was implemented. Model performance was evaluated using the seqeval (https://github.com/chakki-works/seqeval) toolkit to compute F1 score, precision, and recall.

4.8. Knowledge graph construction

The knowledge graph was systematically constructed by processing all 11,068 published records through DeepSeek-V3 for triplet extraction. Python scripts parsed the YAML output files to decompose triplets into head entities, tail entities, and their relationships, assigning sequential unique identifiers to each entity. These structured data were stored across two CSV files: one containing all entities with their corresponding IDs and attributes, and the other maintaining relationship mapping between entity pairs. These tabular representations were then programmatically converted into Cypher query language statements using custom Python scripts. The resulting Cypher commands were executed in NebulaGraph (https://github.com/vesoft-inc/nebula) to establish the graph database infrastructure, where nodes (entities) were first created with their properties, followed by the formation of directed edges (relations) between corresponding head and tail entities based on their pre-assigned IDs, ultimately leading to a complete knowledge graph with all validated triplets.

4.9. QA system implementation

A hybrid retrieval-augmented generation system was developed by extracting titles, DOIs, Abstracts, and Results sections from all 11,068 publications and processing these texts through Ollama's BGE-M3-567m embedding model (https://github.com/ollama/ollama) to generate vector representations. The textual data and their relationships were stored in NebulaGraph using NebulaPropertyGraphStore, while PropertyGraphIndex was used to establish a unified indexing system with local vector embedding for efficient retrieval. The system employs dual retrieval mechanisms, in which a semantic similarity search identifies the top 5 most relevant entries based on vector proximity, while keyword extraction from user queries feeds into text2cypher (https://github.com/vesoft-inc/nebula) to generate executable Cypher queries for structured knowledge graph traversal. Retrieved evidence chains are synthesized into comprehensive prompts for Mistral 8x7B, which performs final answer generation with built-in reliability checks. The model first evaluates prompt sufficiency and then either delivers a complete response or explicitly labels LLM-generated supplemental information where gaps exist, ensuring transparent and accurate knowledge delivery (Fig. 1D).

4.10. Comparison of MEPAM with other traditional large language models

To compare MEPAM with other LLMs, a high-quality benchmark dataset was systematically constructed for evaluation. Thirty specific questions involving various enzymes (including β-glucuronidase, aspartate transcarbamylase, laccase, and cellulase) were collected from published scientific literature. Each question was traced back to its original source publication, and the DOI was recorded for traceability. Key data were then manually extracted from the relevant text passages and standardized into concise and unambiguous reference answers. The questions cover multiple critical enzyme properties, such as kinetic parameters (e.g., K_m, V_max), optimal reaction conditions (pH, temperature), substrate specificity, and inhibitors or activators.

To eliminate evaluation bias caused by differences in response styles across models, a uniform post-processing procedure was applied to all raw model outputs: each model-generated answer was manually reviewed to precisely extract values, units, or entries corresponding directly to the dimension referenced in the reference answer. The extracted content was then formatted into a string consistent with the structure of the reference answer (e.g., temperature units unified to °C, ions written as Zn²⁺). This step ensured fairness and consistency in the evaluation.

A multi-dimensional set of metrics was employed for comprehensive assessment, which was defined and calculated as follows.

• •Exact match: A value of 1 is assigned if the model answer exactly matches the reference answer string; otherwise, 0 is assigned. This metric measures absolute accuracy.
• •Precision: Calculated as the number of matching tokens divided by the total number of tokens in the model answer. This measures the accuracy of the information provided.
• •Recall: Calculated as the number of matching tokens divided by the total number of tokens in the reference answer. This measures the completeness of the information provided.
• •F1score: The harmonic mean of precision and recall (F1 = 2 × (Precision × Recall)/(Precision + Recall)), serving as a core comprehensive metric for overall performance.
• •Faithfulness: Evaluates whether the model's output remains faithful to the provided factual source. A score of 1 is assigned for fully correct or non-contradictory answers and 0 for answers containing contradictory information; for intermediate cases, the score is determined proportionally based on the fraction of the model's extracted content that overlaps with the reference answer.
• •Hallucination rate: Measures the proportion of responses containing specific erroneous information unsupported by the knowledge source. A score of 1 is assigned if the entire output consists of hallucinated content and 0 if no hallucinated content is present. For intermediate cases, the score is calculated proportionally based on the fraction of the model's extracted content that introduces unsupported or erroneous information (vague statements or missing information are not considered hallucinations but may reduce other scores).

4.11. Data processing and meta-analysis

All structured data extracted from the knowledge graph underwent systematic processing and statistical analysis using Python 3.10.9. The Pandas library (v1.5.3) was employed for comprehensive data wrangling, including cleaning, transformation, and aggregation of Excel-formatted datasets, with a particular focus on microbial host–enzyme pairing frequencies, fermentation parameter distributions, and production yield correlations. Statistical validation was performed using SciPy (v1.9.3) for outlier detection (Tukey's method), normality testing (Shapiro–Wilk), and distribution analysis, while metadata cross-verification ensured consistency with annotations in the source literature. The complete data processing pipeline, including all analytical workflows and quality control measures, has been made publicly available in a GitHub repository (https://github.com/tongqichang/MEPAM) to ensure reproducibility and transparency.

4.12. Implementation of the visualization pipeline

The visualization workflow was implemented through a multi-platform approach combining custom Python scripts and specialized tools. Matplotlib and Seaborn libraries were used to generate segregated pie charts and clustered bar plots with publication-quality formatting, while ChiPlot (https://www.chiplot.online/) produced circular packing diagrams to visualize hierarchical data relationships. To analyze knowledge graph topology, Gephi (v0.9.7) was employed with ForceAtlas2 layout algorithms and modularity-based community detection, enabling both the identification of macro-level network patterns and examination of micro-level entity relationships. All visualizations maintained a consistent color-coding scheme corresponding to the entity types and relationship categories established in the knowledge graph.

CRediT authorship contribution statement

Qichang Tong: Writing – original draft, Validation, Software, Methodology, Investigation, Data curation. Lincong Zhou: Writing – original draft, Validation, Software, Methodology, Investigation, Data curation. Xu Liu: Formal analysis. Xiaoqing Liu: Formal analysis. Ningfeng Wu: Writing – review & editing. Yuan Wang: Formal analysis. Huiying Luo: Supervision. Bin Yao: Supervision. Jian Tian: Writing – review & editing, Conceptualization. Dongfei Han: Writing – review & editing, Conceptualization. Xianghua Yan: Writing – review & editing, Conceptualization. Feifei Guan: Writing – review & editing, Writing – original draft, Validation, Software, Methodology, Investigation, Formal analysis, Data curation, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Data availability

All datasets, computational models, and source code utilized in this study are publicly available in the Supplementary Materials and through our GitHub repository: https://github.com/tongqichang/MEPAM. The question-answering system, MEPAM, is available at http://180.76.108.212.