Patent protection of biological genetic resources based on deep learning and artificial intelligence

Abstract

With the rapid development of artificial intelligence (AI), deep learning has provided new ideas for the patent protection of biological genetic resources in the field of intellectual property. This paper aims to explore the application of deep learning technology driven by artificial intelligence in the patent protection of biological genetic resources. By analyzing the advantages of deep learning models in extracting patent texts and technical features, the paper proposes a technical approach to optimize the patent protection of biological genetic resources. This paper constructs and optimizes a deep learning-based Recurrent Convolutional Neural Network (RCNN) model. It combines natural language processing techniques and image recognition algorithms to efficiently extract and analyze patent texts and technical features. The model’s performance is further enhanced by introducing Top-K max pooling strategy and pre-trained word vectors (such as GloVe). Experimental results show that the optimized RCNN model performs excellently in classifying patents related to biological genetic resources. It achieves an overall accuracy of 90.20% and an F1 score of 89.00%. In subcategories such as agriculture, medicine, and biotechnology, the model’s accuracy and F1 score both exceed 90%. Additionally, the use of Top-K max pooling strategy and pre-trained word vectors (GloVe) significantly improves the model’s feature extraction ability and classification accuracy. The proposed optimized RCNN model demonstrates strong adaptability and efficiency in the field of patent protection for biological genetic resources. Compared with existing technologies, the optimized model not only improves the accuracy and efficiency of patent text classification but also reduces the time and cost of manual review through automated processing. The research results provide new technical support for the intellectual property protection of biological genetic resources and offer practical experience for the application of deep learning in the field of intellectual property.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-25051-y.

Introduction

Research background and motivations

Since artificial intelligence (AI) has advanced so quickly, deep learning technology has found extensive application in a variety of domains, including image recognition, natural language processing (NLP), and creative scientific research. The international society has been very concerned recently about the use and preservation of biological genetic resources¹. Biological genetic resources are not only of great value in agriculture, medicine, ecology and other fields, but also involve the global issues of biodiversity protection and sustainable development. How to strengthen the intellectual property protection of these resources, especially patent protection, through technical means, has become an urgent problem to be solved².

The application of deep learning technology in intellectual property protection provides a new perspective for patent mining and management of biological genetic resources³. Biological genetic resources have become an important focus of international intellectual property competition in recent years because of their uniqueness and scarcity. However, traditional patent protection methods face many challenges. In this context, as an important branch of AI, deep learning provides technical support to solve these challenges⁴. Through big data analysis and pattern recognition, deep learning can automatically extract and analyze information related to biological genetic resources from massive patent databases, thus improving the accuracy and efficiency of patent protection.

Research objectives

The paper aims to investigate the use of AI-powered deep learning technologies for biological genetic resource patent protection. The technical route to maximize patent protection of biological genetic resources is presented by examining the benefits of deep learning models in patent text and technical feature extraction. The specific objectives are as follows: (1) Establish a deep learning model for patent protection of biological genetic resources: Mine key technical information in patent texts and maps through NLP and image recognition technology. (2) Reduce the amount of manual labor and use algorithm optimization to increase the speed and precision of patent analysis. In addition to offering the theoretical foundations for the intellectual property protection of biological genetic resources, this paper offers practical experience to further investigate the use of AI in the domain of intellectual property.

Literature review

The significance of patent text mining for technical innovation, intellectual property protection, and technology transfer has made it a research hotspot in recent years. Song et al. (2024) put forward a digital patent analysis system integrating NLP and Machine Learning (ML) technologies, aiming at improving the efficiency of patent recommendation, transferability evaluation and research team detection. This system showed great potential in improving technology transfer efficiency and enhancing enterprise competitiveness⁵. Miric et al. (2023) discussed the application of supervised ML in large-scale management research, especially in identifying AI patents. They compared the traditional keyword-based method with ML and found that the latter performed better in text classification⁶. By combining the features of lexical networks with deep learning models, Li et al. (2023) created a deep learning patent text classification technique that significantly increased classification accuracy. This approach successfully addressed the issue of the classic deep learning model’s inadequate feature extraction capabilities while working with lengthy texts (such as patent papers), particularly in the categorization of Chinese patent materials⁷. Wang et al. (2024) presented a patent text categorization technique based on Automatic Search Optimization (ASO) and Bidirectional-Long Short-Term Memory (Bi-LSTM) technology. The model was shown to have a high accuracy (above 88%) when handling the classification tasks of technical fields, technical impacts, and technical means labels. By using optimization procedures, the model’s accuracy and efficiency were further enhanced⁸. Whalen et al. (2020) provided a patent similarity dataset based on a vector space model, aiming at providing scholars and policy makers with insight into innovation and intellectual property rights through similarity measurement between patents⁹.

Amid the continuous development and widespread application of deep learning, NLP, and AI technologies, research in related fields is also expanding and deepening. Menaouer et al. (2025) proposed a deep learning-based sentiment classification model called Knowledge Graph Convolutional Network. It was used to analyze the sentiment of tweets related to the Russia-Ukraine war. They collected and processed a large amount of tweet data. Experiments showed the model’s accuracy and effectiveness in sentiment classification tasks. This demonstrated the potential of deep learning in social media text analysis¹⁰. In addition, Menaouer et al. (2023) studied an Android malware detection method based on stacked autoencoders and convolutional neural network (CNN). The study used deep learning techniques to detect and classify mobile malware. It achieved high accuracy and performance metrics. This further proved the value of deep learning in the field of information security¹¹. In the medical field, Abdeldjouad et al. (2024) conducted a systematic review and meta-analysis to assess the performance of AI in predicting adverse drug reactions in cancer patients. They found that AI models had high specificity and sensitivity in predicting adverse drug reactions. This showed great potential for AI in healthcare. However, it also pointed out the need for standardized and multicenter studies to improve the quality of evidence¹².

These studies not only show the wide use and great success of deep learning, NLP, and AI in different fields, but also prove their strong ability to handle complex text data and find valuable information. The development and use of these technologies offer useful lessons for the patent protection of biological genetic resources. They also provide a stronger theoretical basis and technical support for the deep learning-based method of protecting biological genetic resource patents proposed in this paper.

Research methodology

Patents of biological genetic resources

Biogenetic resource patents provide legal protection for the development of biogenetic materials and technology including bacteria, plants, animals, and genes^13–15. The use of biological genetic resources is expanding across a wide range of industries, from agriculture to medicine, given the advancements in science and technology, particularly the quick growth of genomics and biotechnology^16,17. Patent categories are shown in Fig. 1.

Fig. 1.

Patent category.

Figure 1 shows that patents are mainly divided into three major categories: agricultural patents, medical patents, and microbial patents. Agricultural patents cover things like crop improvement and genetic modification, such as patents for genetically modified corn. Medical patents involve drug formulas and gene therapies, like patents for gene-editing treatments for diabetes. Microbial patents focus on innovations and applications of microbial strains, with a patent for an antibacterial strain being a typical example. This figure helps people quickly understand the main types of patents related to biological genetic resources and the specific technical fields they cover. It also provides a basis for classifying patents when studying the application of patent text analysis models on different types of patents.

The patent protection of biological genetic resources mainly involves applying for and protecting patents on animal and plant genes, microbial strains, and biomolecules. This is to prevent misuse or improper exploitation. Deep learning technology can improve the intelligence level of patent application, examination, and management by automating data analysis. Traditional gene patent searches rely on keyword matching in gene databases. However, deep learning models can automatically learn the features of gene sequences and improve matching accuracy. For example, a variant of Bidirectional Encoder Representations from Transformers (BERT) can automatically identify text related to biological genetic resources in patent documents and compare it with existing patents. A biotech company developed a new rice gene disease-resistance technology. During the patent authorization process, deep learning can automatically analyze related patents applied worldwide to see if there’s any potential infringement risk. For example, using a Graph Neural Network (GNN) to analyze the citation relationships in biotech patents can help determine the ownership of the technology.

Given the large number of patents related to biological genetic resources, deep learning’s natural language processing (NLP) technology can automatically classify patent texts and generate brief summaries. For example, the Generative Pre-trained Transformer (GPT) model can be used to automatically extract the key innovations from patents, which improves the efficiency of patent examiners. The application areas of deep learning in the patent protection of biological genetic resources are shown in Table 1.

Table 1.

Application field of deep learning for patent protection of biological genetic resources.

Application area	Traditional method	Deep learning optimization scheme	superiority
Gene sequence alignment	Keyword retrieval	Similarity analysis of deep learning	Improve the matching accuracy
Infringement detection	Manual comparison	GNN analysis patent network	Discover hidden patent associations
Patent classification	Rule matching	NLP text classification	Improve classification automation
Abstract generation	Hand-written	Generative AI	Save time and improve efficiency

Patent text analysis model

Convolutional neural network

As a typical forward propagation neural network, Convolutional Neural Network (CNN) structure is shown in Fig. 2.

Fig. 2.

CNN structure.

A standard CNN has at least one convolutional layer. This layer is the core of the CNN and is responsible for extracting features from the input data. In the convolutional layer, data is processed through a weight-sharing mechanism. This mechanism reduces the number of parameters that need to be learned during training, which speeds up model convergence. Also, CNN can perform parallel processing because neurons in the same feature map share the same weights. This is a significant advantage over fully connected networks. The structure in the diagram shows the direction of data flow and processing in the network. It starts with the input data, which is processed through the convolutional layer to extract features. This prepares the data for classification or other tasks and gives a clear understanding of the basic architecture and working principle of CNN. This lays the foundation for understanding the optimization of CNN models and their application in patent text analysis.

CNN optimization model

In this paper, the structure and hyperparameter of CNN are optimized in detail. In the initialization stage of the word embedding layer of the model, this paper explores two different methods. The first method is to use pre-trained word vectors, while the second method is to initialize by random generation. For pre-training word vectors, this paper selects two neural network model word vectors based on different large-scale corpus training: Word2Vec (based on Google News corpus) and GloVe (based on Common Crawl and Wikipedia corpus). In the aspect of random initialization, this paper investigates two strategies: Uniform distribution and Gaussian distribution¹⁸. In the practical application of model training, this paper implements two learning strategies: Transfer Learning and Static Learning. Transfer learning involves taking the trained model parameters as the starting point and fine-tuning and optimizing the parameters for the new dataset¹⁹. Static learning is to keep the pre-trained word vector unchanged during the training process, and the word vector will not be modified with the model training and iteration²⁰. The structure of CNN optimization model is displayed in Fig. 3.

Fig. 3.

CNN optimization model structure.

Figure 3 shows the optimized model structure of a CNN. At the input stage of the model, the number 64 indicates the number of data samples processed in each iteration or training batch. The word embedding dimension is set to 300, which corresponds to the scale of the pre-trained word vector model used. In the third layer, the convolution operation uses three different kernel sizes (3, 4, and 5), with 100 kernels of each size. The dimensions of the feature maps obtained after convolution are closely related to the input sentence length and the kernel sizes. Two optimizations are made for the fully connected layer: Dropout parameter adjustment: Since this paper aims to extract key feature vectors from patent texts, each dimension is important. Therefore, the Dropout parameter is set to 0 to ensure all neurons are active. Tests on the H04 dataset show that removing the Dropout layer has minimal impact on classification accuracy. A dual-layer classifier design: The fully connected layer is responsible for outputting classification probability estimates. The first layer converts a 300-dimensional input vector into a 50-dimensional output vector, and the second layer further reduces it to 3 dimensions. This dual layer structure achieves nonlinear mapping, reduces the dimensionality of feature vectors, and improves the performance of the classifier²¹.

Construction of recurrent CNN model

Text representation occupies a fundamental position in the field of NLP. However, traditional models often ignore the context and word order information of the text, which limits their ability to capture sentence structure. Although these models can quickly convert text into vector form, they cannot handle key information such as word order²². As deep learning technology has advanced, the neural network-based approach has emerged as the primary way to extract sequential information from text. Although CNN is effective at extracting features from patent papers, it has certain drawbacks, such as trouble identifying long-term dependencies and choosing the right convolution kernel size. Important information could be overlooked if the convolution kernel size is too tiny. However, the calculation cost will rise if the convolution kernel size is excessively large^23,24. In order to overcome these challenges, this paper puts forward an improved model of Recurrent Convolutional Neural Network (RCNN), which uses its advantages in processing sequence data to automatically identify the order dependence among text sentences, and carries out optimization training according to the topic clustering results of patent texts. CNN model relies on the size of the convolution kernel to artificially determine the relationship among sentences. This method lacks flexibility²⁵, while patent texts often contain longer sentences. Although RCNN model is chosen at the expense of certain time complexity, it can obtain the relationship structure at the sentence level of the text, and its time complexity remains at O(n), which is lower in time and space complexity than Recurrent Neural Network (RNN) and other models²⁶.

1. Word embedding layer.

   In essence, the word embedding layer in RCNN model adopts a design of bidirectional RNN, which captures the context information of words (including the left and right texts) by using reverse and forward recurrent processes respectively. The specific mathematical expression is as follows:

   1

   2
   In RCNN model, represents the current word. Specifically, the vector representing the current word is determined by the context representation of the left text and the right text. represents the left text of the current word. represents the right text of the current word, and is the word vector of the current word. and represent weight parameters, and is a nonlinear function. The equation shows that the above representation of the current word is determined by the context representation of the text on the left side of the current word and the word vector on the left side of the current word. The text representation on the right is also calculated in the same way²⁷. Therefore, based on the contextual representation of the current word, people can infer the representation of the current word in the text:

   3
   The representation of words in a text is understood as the vector form of words. Then, a single-layer neural network is used to extract the potential semantic features of words, which is similar to the convolution layer in CNN. The specific operation details of convolution layer are as follows:

   4

   represents the weight parameter and represents the offset parameter. The symbol represents the activation function, and the hyperbolic tangent function is selected.

2. Pooling layer

   In this paper, the maximum pooling layer is adopted. A crucial aspect of text analysis is the close relationship between features and positions, for example, the positioning information of key sentence components in the text. In order to capture the essential information in the text, this paper uses the maximum pooling strategy to extract the most important features from the semantic vectors of words, even though the latent semantic vectors created in the previous step are unable to clearly highlight the significance of particular features²⁸. In addition, the pooling layer also effectively solves the problem of different text lengths, that is, texts (sentences) with different lengths are uniformly converted into vectors with fixed dimensions²⁹. The mathematical expression of the maximum pooling operation is as follows:

   5

3. Fully connected layer

   In the fully connected stage of the model, the feature information that was collected from the text input in the previous stage is first summarized by this layer using a single-layer neural network structure. The mathematical expression that best describes its detailed function is as follows:

   6
   Then, the model uses the Softmax function to estimate the possibility that the marked patent text belongs to each category, which can be expressed by the following equation:

   7

RCNN optimization model

In CNN and RCNN models, the pooling operation usually adopts the maximum pooling strategy, that is, the maximum value in the feature map is reserved and the remaining features are discarded. This method can effectively maintain the position information and rotation invariance of features in computer vision^30,31. However, it is not entirely reasonable to ignore the position information in patent text processing. Patent text contains an orderly combination of different sentence components (such as subject, predicate, object, etc.), and the position information of these components is very important for the sentence structure and order³². To solve the problem of ignoring positional information and order in traditional maximum pooling, this paper proposes to use a Top-K maximum pooling strategy. This strategy retains the Top-K largest values in the original sequence for a given value of K and keeps the relative position information of these values. Top-K maximum pooling can better express the multiple occurrences of features and their intensities while preserving the positional information. Overall, in patent text processing, Top-K maximum pooling has several advantages:

1. Retain the Benefits of Maximum Pooling and Reducing Model Complexity.

   Top-K maximum pooling inherits the advantages of maximum pooling, which can reduce the number of model parameters, prevent over-fitting, and solve the problem of indefinite patent text length³³. After pooling, the dimension of each feature map is reduced to l×k, which significantly reduces the computational complexity, thus improving the model efficiency and processing scale.

2. Preserve Location Information and Alleviate Bias in RNN Models.

   Top-K maximum pooling extracts important features from the text and retains position information, which alleviates the bias problem of RNN model. RNN usually thinks that the later words have greater influence, but in fact, important words may appear anywhere in the text^34–37. Top-K maximum pooling avoids over-reliance on sequence order by retaining high-score features and their position information, and captures key information of text more comprehensively.

The optimized RCNN combines the feature extraction ability of CNNs and the context modeling ability of RNNs. It also incorporates the Top-K max pooling strategy and pre-trained word vectors (such as GloVe). The specific features of these optimization strategies are as follows:

(1) Combining the strengths of CNNs and RNNs: CNNs are used for extracting local features and can quickly capture key information in the text. RNNs are used for modeling context information and can handle long-range dependencies in longer texts. (2) Top-K max pooling strategy: Unlike traditional max pooling or average pooling, Top-K max pooling retains the top K most important features and their position information, avoiding information loss. By selecting the top K features, the model’s computational load is reduced and processing efficiency is improved. (3) Pre-trained word vectors (GloVe): GloVe word vectors are trained on large-scale corpora and can better capture semantic relationships between words. Initializing the model with pre-trained word vectors significantly improves classification accuracy and feature extraction ability.

Experimental design and performance evaluation

Experimental design

This paper uses datasets related to biological genetic resource patents, which are mainly sourced from the following publicly available patent databases: WIPO (World Intellectual Property Organization): https://patentscope.wipo.int/; USPTO (United States Patent and Trademark Office): https://www.uspto.gov/patents/search. The dataset is a publicly available resource with no special access restrictions, but users must comply with the terms of use of each database. The experiment adopts a self-defined three-category patent labelling scheme, which classifies patent texts into three categories according to their technical fields: agriculture, medicine, and biotechnology. The classification is mainly based on the definitions of the major categories A01 (agriculture), A61 (medicine), and C12 (biotechnology) in Cooperative Patent Classification (CPC) and International Patent Classification (IPC), and combined with manual review to make the final judgment on specific cross-field patents. In terms of handling multi-label problems, to avoid excessively high modeling complexity, this paper adopts a single-label classification strategy. The training set, validation set, and test set adopt a time-based partitioning strategy (data before 2018 is used for training, and data after 2018 is used for validation and testing), and the results are compared to maintain the consistency. All experiments were repeated three times under different random seeds (seed = 42, 123, 2025). In the data preprocessing stage, this paper only uses the Title, Abstract, and Claims of the patent, without using the instruction manual to avoid text redundancy and noise. The maximum sequence length is uniformly 512, and the excess part is truncated on the right side.

Among them, agricultural patents mainly include technologies like crop improvement and genetic modification, such as patents related to genetically modified corn, with a sample size of 3,217, accounting for 33.4% of the total data. Medical patents focus on directions such as drug formulations and gene therapy, such as patents on gene editing for diabetes treatment, with a data volume of 3,982, accounting for 41.4%, which is the highest proportion among the three categories. Microbiological patents involve the innovation and application of microbial strains, such as invention patents for a certain antibacterial strain, with a total of 2,424 collected, accounting for 25.2%. Overall, the distribution of the three types of patent data is relatively balanced, which helps the model maintain stable classification capabilities in different types of texts.

This paper conducts model training and testing on a high-performance computing platform. The hardware environment used in the experiment includes an Intel Xeon Gold 6226R processor, an NVIDIA RTX 3090 graphics card (24GB memory), and 128GB Double Data Rate 4 (DDR4) memory. This configuration provides strong support for the training of large-scale deep learning models, which can effectively improve training efficiency and model convergence speed.

In terms of software environment, the systems and tools used in this article include: Ubuntu 20.04 (Ubuntu Linux operating system, Version 20.04, https://ubuntu.com/); Python 3.8 (Python programming language, Version 3.8, https://www.python.org/); TensorFlow 2.10 (TensorFlow deep learning framework, Version 2.10, https://www.tensorflow.org/); PyTorch 1.12 (PyTorch deep learning framework, Version 1.12, https://pytorch.org/). Auxiliary development and data processing use mainstream scientific computing and machine learning toolkits such as NumPy, Pandas, and Scikit-learn, which provide a solid foundation for the stability and reproducibility of the experiment.

In the experiment, the parameter settings of CNN and RCNN are shown in Table 2.

Table 2.

Parameter setting.

Parameter name	Value	Explanation
Batch Size	64	The amount of data per training iteration
Learning Rate	0.001	Use Adam optimizer
Embedding Dimension	300	Pre-training word vector dimension
Filter Sizes	[3, 4, 5]	Three convolution kernel sizes, 100 each.
Dropout rate	CNN: 0, RCNN: 0.2	Control overfitting
Top-K maximum pooling parameter	K = 5	Extracting K important features
Number of training rounds (Epochs)	50	Early stop strategy: stop when the loss is not improved for 5 rounds

The baseline model sources and parameter settings used for comparison in this experiment are shown in Table 3.

Table 3.

Baseline model parameter settings.

Models	Parameter settings	Source/implementation library
LR (Logistic Regression)	L2 regularization, learning rate = 0.01, max_iter = 1000	Scikit-learn
SVM (Support Vector Machine)	Kernel function = RBF, C = 1.0	Scikit-learn
CNN	Convolutional kernel size=[3,4,5], 100 for each type, Dropout = 0.5, learning rate = 0.001	TensorFlow 2.10
LSTM (Long Short-Term Memory)	Hidden layer = 128, Dropout = 0.2, learning rate = 0.001	PyTorch 1.12
Bi-LSTM	Hidden layer = 128 (64 in both directions), Dropout = 0.2, learning rate = 0.001	PyTorch 1.12
BERT	The maximum sequence length = 512, batch size = 32, Learning rate = 2e-5, training epochs = 10, tokenizer = WordPiece, truncation strategy = right truncation, optimizer = AdamW	HuggingFace Transformers
PatentBERT	The maximum sequence length = 512, batch size = 32, Learning rate = 2e-5, training epochs = 10, tokenizer = SentencePiece, Truncation strategy = right truncation, optimizer = AdamW	HuggingFace Transformers

Performance evaluation

This paper thoroughly examines the model’s performance in feature extraction and classification of biological genetic resource patent texts in particular fields using indicators such as accuracy, precision, recall, and F1 score to assess the suggested model’s performance.

Baseline model comparison

To comprehensively evaluate the performance of the proposed optimized RCNN model, this paper compares it with various baseline methods, including traditional machine learning models (LR and SVM), classical deep learning models (CNN, LSTM, and BiLSTM), and pre-trained language models (BERT and PatentBERT). The baseline model comparison results are shown in Fig. 4.

Fig. 4.

Baseline model comparison results.

In traditional methods, the overall performance of LR and SVM is weak, with accuracies of 72.5% and 74.3%, respectively, and F1 scores of only 69.5% and 71.6%, which is difficult to effectively capture complex semantic information in patent texts. With the introduction of deep learning, classification performance has been significantly improved: the accuracy of the CNN model has increased to 85.7%, and the F1 score is 83.8%. The LSTM and BiLSTM models are further improved, with accuracies of 87.3% and 88.5%, respectively, and F1 scores increase to 86.1% and 87.5%, indicating that the ability to model sequence dependencies helps improve classification performance. In terms of pre-training language models, the BERT model performs better than traditional deep learning methods, with an accuracy of 89.6% and an F1 score of 88.4%. This model demonstrates the advantages of large-scale general corpus pre-training in patent text classification. After further introducing the PatentBERT model trained on patent domain corpora, the accuracy improves to 90.1% and the F1 score reaches 88.8%, slightly higher than BERT. This indicates that domain specific corpora can better adapt to patent text features. Overall comparison shows that the optimized RCNN model performs the best among all models, with an accuracy of 90.2%, a precision of 88.7%, a recall of 89.4%, and an F1 score of 89.0%. It outperforms traditional models and general/domain pre-trained language models in all indicators. This indicates that the proposed optimization strategy can effectively improve the performance of the model in the task of patent classification of biological genetic resources, and enhance its ability to handle complex terms and long text structures while ensuring generalization ability. In addition, when comparing the performance of RCNN with BERT/PatentBERT, this paper further conducts a paired significance test, and the results show that the improvement of RCNN is statistically significant (p < 0.05).

In summary, the optimized RCNN model outperforms other benchmark models in all performance indicators, not only comprehensively leading in traditional accuracy indicators, but also demonstrating stronger text understanding and category differentiation abilities. It is especially suitable for processing complex terms and long sentence structures in patent texts, which verifies its wide applicability and superior performance in the context of biological genetic resource patent protection.

Influence of different pooling strategies

In this experiment, the effects of maximum pooling, average pooling and Top-K maximum pooling on the model performance in optimizing RCNN model are discussed. The results of different pooling strategies are shown in Fig. 5.

Fig. 5.

Effects of different pooling strategies.

The accuracy of the maximum pooling strategy is 88.60%, precision is 87.00%, recall rate is 86.40%, and F1 score is 86.70%. In contrast, the performance of the average pooling strategy is weak, with accuracy of 85.30%, precision of 84.10%, recall of 83.70% and F1 score of 83.90%. The Top-K maximum pooling strategy performs best, with accuracy of 91.00%, precision of 89.50%, recall of 90.20% and F1 score of 89.80%. This result shows that the Top-K maximum pooling strategy can better retain the key information of the text and improve the overall performance of the model.

Influence of word vector initialization method

This experiment analyzes the influence of different word vector initialization methods on the model performance, including Word2Vec, GloVe and Random Initialization. The influence result of the word vector initialization method is shown in Fig. 6.

Fig. 6.

The influence result of word vector initialization method.

In the experiment of word vector initialization method, different word vector initialization methods have different effects on the effect of the model. When the word vector of Word2Vec is used for initialization, the accuracy of the model is 89.20%, the precision is 88.00%, the recall is 87.30%, and the F1 score is 87.60%. When the GloVe word vector is used for initialization, the performance of the model is slightly improved, with the accuracy of 90.50%, precision of 89.10%, recall of 89.90% and F1 score of 89.50%. When randomly initialized word vectors are used, the performance of the model is relatively poor, with accuracy of 86.70%, precision of 85.30%, recall of 85.00% and F1 score of 85.10%. It shows that the initialization of GloVe word vector can provide more effective semantic information, thus improving the classification performance of the model.

Model performance under patent type subdivision

The experiment evaluates the performance of the model in patent text classification, including agriculture, medicine and biotechnology, and tests the adaptability of the model under diverse text data. The results of model performance analysis under patent type subdivision are shown in Fig. 7.

Fig. 7.

Results of model performance analysis under patent type subdivision.

The experiment evaluates the performance of the optimized model in the task of patent text classification, covering agriculture, medicine and biotechnology. In the classification task of agricultural patents, the model performs well, with the accuracy of 92.10%, precision of 90.50%, recall of 91.70% and F1 score of 91.10%. In the classification of medical patents, the accuracy of the model is 91.80%, precision is 90.30%, recall rate is 91.00%, and F1 score is 90.60%. In the classification task of biotechnology patents, the performance is slightly inferior, with the accuracy rate of 90.40%, precision of 88.90%, recall rate of 89.50% and F1 score of 89.20%. Overall, the model has shown strong adaptability in each subdivision category and can handle diversified patent text data well. The findings of the experiment demonstrate that the optimized RCNN model outperforms the conventional model in every test situation. Furthermore, pre-training word vectors and the Top-K maximum pooling method can greatly enhance the model’s performance, confirming the significance of sophisticated pooling technology and word vector initialization in patent text processing of biological genetic resources.

Robustness verification: cross verification and ablation experiment

To evaluate the robustness of the model, this study conducts 5-fold cross-validation and designs ablation experiments to verify the contribution of each optimization component. The cross-validation results show that the standard deviation of accuracy of the optimized RCNN model across different folds is ± 0.35%, and the standard deviation of F1-score is ± 0.28%, indicating that the model performs stably on different data subsets. The results of the ablation experiments are shown in Table 4:

Table 4.

Ablation experimental results.

Model configuration	Accuracy	F1 score	Note
RCNN (basic)	87.6%	86.2%	No Top-K pooling, random initialization word vector
RCNN + GloVe word vector	89.5%	88.1%	The initialization of word vectors is obviously improved
RCNN + GloVe word vector + Top-K pooling	90.2%	89.0%	Further improve the ability to retain key features

The basic RCNN model, which only used a bidirectional RNN structure, randomly initialized word vectors, and traditional max-pooling, achieved an accuracy of 87.6% and an F1-score of 86.2%. This result has outperformed some traditional deep learning models (e.g., CNN, LSTM), demonstrating the inherent advantage of the RCNN architecture in capturing contextual sequence information of patent texts. When randomly initialized word vectors were replaced with GloVe word vectors pre-trained on large-scale corpora, the model performance improved significantly—both accuracy and F1-score increased by 1.9%. This improvement confirms that high-quality semantic feature initialization is crucial for the model to understand patent texts with dense professional terms. Pre-trained word vectors provide the model with rich prior semantic knowledge, effectively accelerating model convergence and enhancing the depth of feature extraction, which serves as the foundation for the model to achieve superior performance. On the basis of GloVe word vectors, after introducing the Top-K max-pooling strategy (K = 5), the model performance reached its peak: accuracy further increased to 90.2% and F1-score reached 89.0%. This result is mutually corroborated with the conclusion in Sect. 4.2.2. The advantage of this strategy lies in overcoming the defect that traditional max-pooling may discard multiple important features. By retaining the top K most significant features and maintaining their relative position information, the model can more comprehensively capture the expressions of multiple key technical points in long patent sentences, thereby significantly enhancing the model’s ability to retain features of complex text structures. The results of the ablation experiments show that GloVe word vectors and the Top-K max-pooling strategy contribute significantly to the improvement of model performance, and all components of the model structure have practical utility.

Discussion

The paper proposes an optimized RCNN model for the classification of patent texts, with a focus on biological genetic resources patent protection. This paper’s experimental section validates various pooling tactics, word vector initialization techniques, and the model’s flexibility in subdividing patents into several categories. According to the experimental findings, the optimized RCNN model outperforms the conventional text classification model in several indicators, including LR, SVM, the standard CNN model, etc. The combination of Top-K maximum pooling and GloVe word vector initialization shows the best performance in the selection of different pooling strategies and word vector initialization methods. Additionally, the model’s performance in subdivided patents (such as those in biotechnology, medicine, and agriculture) confirms its adaptability to a variety of text data. The optimized model’s accuracy and recall are particularly good in patent protection for biological genetic resources. Compared with the latest research using the same dataset, it is found that the accuracy and F1 score of the proposed method are significantly higher than those of Li et al. (2023) and Haghighian Roudsari et al. (2022). For example, the highest accuracy of this proposed method reaches 92.10%, while the highest accuracy of Li et al.‘s BERT combined with vocabulary network is 83.66%, and the F1 value of Haghighian Roudsari et al.‘s XLNete (Xtreme Language Net) model is 84.72%. This paper specifically optimizes for patents related to biological genetic resources and can effectively handle complex terminology and contextual information in the biotechnology field. In comparison, Li et al.‘s method mainly targets general patent texts, and Haghighian Roudsari et al.‘s method focuses on multi-label classification tasks. Both show weaker adaptability in the classification of biological genetic resource patents.

These results not only prove the effectiveness of the optimized RCNN model in patent text classification, but also reveal how to further improve the classification performance through different pooling strategies and word vector initialization methods. At the same time, this paper also shows the potential application value of the model in specific fields (such as patent protection of biological genetic resources), which can provide support for the management and protection of intellectual property rights.

Conclusion

Research contribution

The main contributions of this paper are as follows: (1) An optimized RCNN model is proposed, achieving excellent performance in patent text classification, especially in the protection of biological genetic resource patents. Experimental results show that the optimized RCNN model outperforms traditional text classification models such as logistic regression, support vector machines, and standard CNN models in multiple metrics. Among different pooling strategies and word vector initialization methods, the combination of Top-K max pooling and GloVe word vector initialization demonstrates the best performance. (2) The impact of different pooling strategies and word vector initialization methods on patent text classification performance is systematically analyzed and verified. It is found that Top-K max pooling and GloVe word vector initialization are key factors in improving model performance. (3) Experiments on fine-grained patent texts in agriculture, medicine, and biotechnology validate the model’s adaptability to diverse text data, providing a reference for patent text classification in different fields. (4) The application potential of the optimized RCNN model in the field of biological genetic resources is experimentally demonstrated, offering new technical support for the patent protection of biological genetic resources.

Future works and research limitations

Globally, the protection and utilization of biological genetic resources is an important issue involving biodiversity conservation, sustainable development, and intellectual property management. The research findings of this paper not only provide technical support for the patent protection of biological genetic resources, but also offer a new perspective for global intellectual property management. Through the application of deep learning technology, innovative achievements of biological genetic resources can be identified and protected more efficiently, promoting technological exchange and cooperation worldwide.

However, despite the satisfactory experimental results achieved here, there are still some limitations and room for future research: (1) Dataset expansion: The current experiments are mainly based on a limited patent text dataset. In the future, the model’s universality and robustness in patent classification tasks across different languages, regions, and more fields can be further verified by expanding the scale of the dataset. (2) Domain adaptability: This paper mainly focuses on the field of biological genetic resource patents. In the future, the optimized model can be applied to patent classification in more fields, such as environmental protection and artificial intelligence, to explore its cross-domain adaptability. (3) Multilingual support: The current research is mainly based on English patent texts. In the future, patent text classification tasks in multiple languages can be considered to meet the diverse needs of global intellectual property management. (4) Dynamic updating capability: The field of biological genetic resources is rapidly evolving. Future research can explore dynamic updating mechanisms for the model to adapt to the changing characteristics of patent texts and technological backgrounds.

Practical and managerial implications

The deep learning-based method for protecting biological genetic resource patents proposed in this paper not only demonstrates significant innovation and superiority at the technical level but also provides important insights and guidance for practice and management in related fields.

1. Improving Patent Protection Efficiency and Accuracy

   In the field of biological genetic resources, the complexity and professionalism of patent protection demand strong technical support. The proposed optimized RCNN model can efficiently extract key technical information from patent texts, significantly enhancing the accuracy and efficiency of patent application, authorization, and protection. By automating analysis and classification, it reduces the time and cost of manual review while improving the quality and reliability of patent protection.

2. Supporting Intellectual Property Strategy Planning

   For companies and research institutions, patent layout in biological genetic resources is a crucial part of their intellectual property strategy. This method can quickly identify and analyze potential patent opportunities, helping companies and research institutions optimize their patent layout and enhance their position in global intellectual property competition. Through precise patent mining and analysis, companies can better define their technology development direction and market strategy, maximizing the commercial value of intellectual property.

3. Informing Policy Making and Regulation

   At the policy level, the protection of biological genetic resources involves global issues such as biodiversity conservation and sustainable development. The research findings of this paper provide policymakers with a technical basis, helping them better understand and manage intellectual property issues related to biological genetic resources. The optimized deep learning model can be used to monitor and assess the quality of patent applications, preventing potential threats to biodiversity and public interests from improper patent applications.

4. Promoting Interdisciplinary Collaboration and Innovation

   Patent protection of biological genetic resources involves multiple disciplines, including biology, genetics, law, and information technology. The deep learning method proposed in this paper offers new ideas and technical tools for interdisciplinary collaboration. By integrating natural language processing, image recognition, and deep learning technologies, it can promote synergistic innovation across disciplines and drive technological progress in the field of biological genetic resource protection.

5. Enhancing Corporate Competitiveness

   In market competition, intellectual property protection is an important manifestation of a company’s core competitiveness. This method can help companies quickly identify and protect their innovative achievements related to biological genetic resources, strengthening their technological barriers and market competitiveness. Through precise patent analysis, companies can better respond to intellectual property challenges from competitors, thus gaining a favourable position in the market.

The deep learning-based method for protecting biological genetic resource patents proposed in this paper is not only innovative in technology but also significant in practice and management. The research findings provide strong technical support and practical guidance for the intellectual property protection of biological genetic resources. This is achieved by improving patent protection efficiency. It also supports intellectual property strategy planning. The findings inform policy making and regulation. They promote interdisciplinary collaboration. Corporate competitiveness is enhanced and sustainable development is driven.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Zichen Liu: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparationLu Liu: writing—review and editing, visualization, supervision, project administration, funding acquisition.

Funding

This research received no external funding.

Data availability

The datasets used and/or analyzed during the current paper are available from the corresponding author Lu Liu on reasonable request via e-mail 13791569188@163.com.

Declarations

Ethics statement

This article does not contain any studies with human participants or animals performed by any of the authors. All methods were performed in accordance with relevant guidelines and regulations.

Competing interests

The authors declare no competing interests.