Practical Design and Implementation of Virtual Chatbot Assistants for Bioinformatics Based on a NLU Open Framework

Abstract

In this work, we describe the implementation of an infrastructure of conversational chatbots by using natural language processing and training within the Rasa framework. We use this infrastructure to create a chatbot assistant for the users of a bioinformatics suite. This suite provides a customized interface solution for omic pipelines and workflows, and it is named GPRO. The infrastructure has also been used to build another chatbot for a Laboratory Information Management System (LIMS). The two chatbots (namely, Genie and Abu) have been built on an open framework that uses natural language understanding (NLU) and machine learning techniques to understand user queries and respond to them. Users can seamlessly interact with the chatbot to receive support on navigating the GPRO pipelines and workflows. The chatbot provides a bridge between users and the wealth of bioinformatics knowledge available online.

1. Introduction

In recent years, the integration of artificial intelligence (AI) has emerged as a transformative interdisciplinary force across various domains with multiple distinct utilities. For example, conversational chatbots built using natural language understanding (NLU) have revolutionized human–computer interactions across various fields. Chatbots are intelligent conversational computer systems designed to mimic human conversation and thinking to enable automated online guidance and support [1,2,3,4] such as healthcare [5,6], e-commerce [7,8], industry, art, and more. This revolution has been supported by the availability and versatility of platforms like Rasa [9] and Dialogflow [10], which allow developers to create sophisticated chatbots regardless of their experience levels.

NLU approaches and conversational chatbots constitute an outstanding opportunity to explore new avenues as tools for user support in more complex scenarios like bioinformatics. Bioinformatics is an interdisciplinary field at the intersection of biological, computational, and data sciences, playing a pivotal role in processing and analyzing complex biological big datasets from omic approaches such as genomics, proteomics, transcriptomics, etc. [11,12,13,14]. The need for sophisticated tools and methods to interpret these datasets has made bioinformatics essential to modern biological research. Data processing in bioinformatics is often performed using protocols referred to as pipelines or workflows, which automate sequences of computational tasks [15,16]. These workflows are complex, frequently requiring numerous computational steps and integrating various command-line interface (CLI) tools [17].

Managing these pipelines is often a challenging task [18], particularly for researchers that lack proficiency managing operating systems or have novice skills in informatics. This complexity arises because bioinformatics pipelines are not based on a single analysis but more properly on the sequential or parallel execution of complex and multiple combination of different steps to extract and process biologically relevant information from the raw omic data. As pipelines vary depending on the omic approach, researchers need to be continuously consulting a wide variety or sources such as biological databases, software manuals, social networks, and online forums like SEQanswers [19], the GATK [20] forum, and BioStar [21], among others. As a result, there is an increasing need for more intuitive, automated tools to assist researchers in managing such workflows [15,22,23].

AI, particularly in the form of machine learning and deep learning, plays a crucial role in bioinformatics by automating the analysis of large and complex datasets [24,25]. AI can uncover patterns in genomic sequences, predict protein structures, and assist in drug discovery, significantly improving the efficiency of bioinformatics tasks. In this context, virtual assistants powered by AI are emerging as a valuable tool in bioinformatics. They provide a user-friendly interface for interacting with bioinformatics tools, simplifying the complex workflows and reducing the need for constant manual intervention. Traditional methods for managing bioinformatics workflows involve manual searches through databases and forums or relying on CLI-based software, which is often time-consuming and prone to errors. These methods require specialized knowledge, making them inaccessible to researchers without advanced computational skills. Virtual assistants such as chatbots offer a solution by providing instant, automated support, reducing the complexity of bioinformatics tools, and improving efficiency by guiding users through workflows and troubleshooting in real-time.

In a previous work [23], we introduced an AI-based infrastructure, primarily composed of expert systems and chatbots developed with the Rasa framework [9]. This infrastructure was tailored to support bioinformatics researchers using the GPRO suite [26], a tool designed to simplify bioinformatics pipelines. Our chatbot assistant significantly enhances the user experience with GPRO by offering intuitive guidance through the complex workflows, thus improving efficiency and understanding. Furthermore, we implemented another chatbot to assist users of our Laboratory Information Management System (LIMS) and its website.

In this work, we present a novel virtual assistant in bioinformatics, built using the Rasa framework, offering insights into the development process and its utilization within bioinformatics infrastructures such as the GPRO project. By focusing on clarifying GPRO pipelines, interfaces, and analytical methodologies, our chatbot empowers users to make informed decisions. It not only simplifies the user experience but also fosters a deeper understanding of the tools available. As the bioinformatics community becomes more diverse, our chatbot serves as a catalyst for collaboration and knowledge sharing, pushing the boundaries of expertise. Furthermore, we delve into the step-by-step technical details, methods, and strategies applied to build this infrastructure of bioinformatics chatbots. Our proposal can be considered as a know-how guide for addressing new projects involving the design and development of chatbot assistants in specialized interaction environments.

This study introduces the application of conversational AI within bioinformatics infrastructures, particularly for platforms like GPRO, where managing complex pipelines is crucial. By expanding chatbot capabilities to interact with bioinformatics tools, this work fills a gap in literature on the intersection of AI and bioinformatics tool management. Our key contributions can be summarized as the following:

• Developed two virtual assistants, Genie (for GPRO platform users) and Abu (for decision-making tasks in LIMS), with integrated API interactions that expand the chatbot knowledge base.

• Provided a bilingual (English and Spanish) interface, allowing users to seamlessly ask questions and receive insightful answers.

• Simplified bioinformatics workflows enabling better management of pipelines and decision-making tasks.

• Empowered users to understand and utilize bioinformatics tools more effectively by offering real-time support and guidance.

The structure of this paper is as follows: In Section 2 (Methods), we detail the approaches and strategies applied to develop and implement the bioinformatics chatbot infrastructure. Section 3 (Results) presents the outcomes of integrating the chatbot with the GPRO and LIMS platforms, including the functionalities and performance of the Genie and Abu assistants. Finally, Section 4 (Conclusions) summarizes the contributions of this study and outlines potential future directions for improving chatbot-based solutions in bioinformatics.

2. Methods

In this section, we describe the main steps and tools used for the design and development of the virtual assistants.

2.1. Data Collection

As the foundation of the chatbot rests upon the domain-specific knowledge and linguistic nuances for effective communication, we collected our data from different sources. The steps to achieve this were as follows:

• Domain expert collaboration: Key intents, entities, and domain-specific vocabulary that form the basis of chatbot interactions were stated based on expertise knowledge in the field.

• Data mining: Knowledge and expert data were obtained by mining the manuals of the tools constituting the bioinformatics pipelines managed by distinct GPRO applications, and the following public databases and repositories: SEQanswers [19], BioStar [21], NCBI [27] and GATK forum [20]. The questions and answers within these databases and repositories cover a spectrum of topics, including bioinformatics, where users can find multiple distinct threads with questions and answers for a wide variety of issues, usage modes, protocols, and command functionalities from the most common tools used in bioinformatics.

  The process of data mining was conducted using the tools for data download provided by each resource or using methodologies respectful and consistent with policies of good practices and data retrieval of each source. More specifically, these sources were mined by using distinct several Python scripts that first went through the different sections of the forums collecting the links of all the posts or threads. Once all the links to the forums were obtained, each one was accessed, and the website html page was downloaded and stored for further processing.

  Selenium WebDriver and Selenium Python API [28] were used to crawl the forums. The interesting data from the websites were extracted and stored in JSON format [29].

2.2. Data Corpus Compilation

Knowledge and expert data were compiled and structured into the following knowledge databases:

1. Usage database: this database collects information from the manuals of the command-line tools implemented in the GPRO pipelines and workflows. This database is organized in two disjoint databases:

   • A command-line database that stores an XML file for each CLI tool used by GPRO pipelines. In GPRO, every tool is executed through the command line in the server-side pipelines, and it provides information about the type of input that each program receives and the output that it generates, as well as the different parameters and customization options. This database was built by using a Java plugging in the GPRO source code to extract information that defines the forms of each of the interfaces that implement all GPRO applications. The XML dataset contains over one hundred different tools that are used on different pipelines in GPRO applications.

   • A contextual database that contains the relationships between the steps and tools used in any pipeline and workflow. In GPRO, this database is needed to generate the pipeline flows. The collections of flows are distributed in image files.

Figure 1 shows a summary of the GPRO pipelines and workflows and Figure 2 shows an example of user interaction flows with the pipeline of one of the GPRO tools [23].

Support database: This database was created by extracting information on questions and answers from the user manuals and FAQ information from the websites of every tool in GPRO. This database was manually curated and classified into different intents based on user expectation. The database contains about 8097 questions/statements classified into 338 intents, divided into 2920 questions classified into 120 intents for Genie and 5177 questions classified into 218 intents for Abu.

Bioinformatics QAs database: This database focused on domain-specific questions and answers provided by sources like SEQanswers, BioStar, and the GATK forum about different bioinformatics tools. The inclusion of this domain-specific QAs database serves a dual purpose: it does not only bolster the depth and accuracy of the chatbot answers, but it also aligns the chatbot closely with the real-world challenges and queries encountered by bioinformatics professionals. GATK has 1904 posts with 7728 comments, SEQanswers has 21,686 posts with 80,355 comments, and BioStar has 99,730 posts and 469,700 comments. All data were provided as sentences in the form of questions and the corresponding answers or comments.

2.3. Data Processing

The data processing involved several steps to prepare the dataset for model training. First, noise and irrelevant information were removed through text cleaning. Techniques implemented in Rasa included lowercasing, punctuation removal, and stop word removal. The XML source files were transformed into JSON format for easier understanding by the chatbot. The XML source files were imported to the chatbot system structured database (mySQL database). As the chatbot dialog manager requests information from the database via an API request, it receives a response in JSON format. This serves as a means of data interchange between different components of the system. Additionally, image files from the contextual database were transformed into map images for integration into the knowledge base.

In the NLP pipelines of Rasa, multiple levels of analysis were applied to ensure effective processing of user inputs. One of the strengths of the Rasa framework is its modular architecture, which allows for the integration of custom components alongside the default ones, enabling more tailored solutions for specific tasks. For instance, we implemented a custom language detection component, which allowed the chatbot to dynamically detect the user’s language, further enhancing its adaptability to diverse user groups.

• Both chatbots Genie and Abu shared a common foundational pipeline consisting of the following key components:

  • -. WhitespaceTokenizer: This tokenized the input text by splitting it based on whitespace.

  • -. RegexFeaturizer: This component extracted structured data using predefined regular expressions, useful for handling numbers, dates, or specific patterns.

  • -. LexicalSyntacticFeaturizer: This component provided additional syntactic and lexical information, such as prefixes and suffixes, enhancing the token-level features.

  • -. CountVectorsFeaturizer: It converted text into numerical representations using character-level n-grams, enabling the chatbot to handle complex text patterns.

  • -. DIETClassifier: Used for both intent classification and entity recognition, this transformer-based classifier allowed for more accurate identification of user intents and entities.

  • -. EntitySynonymMapper: This mapped different user inputs to a single entity, improving consistency in entity recognition.

  • -. ResponseSelector: This selected the most appropriate response based on the context and user intent.

  • -. FallbackClassifier: A fallback mechanism that handled uncertain or ambiguous inputs, ensuring that the chatbot responded gracefully when it was not confident in its prediction.

• While both chatbots shared the above components, the Abu chatbot employed additional features to handle more advanced tasks:

  • -. DucklingEntityExtractor: This component handled the extraction of time-related and date entities, using different configurations for English and Spanish locales to support bilingual functionality.

  • -. LanguageDetector is a custom component used to automatically detect the language of user input. It allows the chatbot to dynamically adjust its processing pipeline and response to handle different languages seamlessly.

  • -. CustomEntityExtractor: This component handled the extraction of custom entities from users’ queries. Such entities are used to generate dynamic API request to the back-end LIMS system to retrieve custom data on behalf of the user.

QA database embeddings use a universal sentence encoder to encode QA data into high-dimensional vectors that can be used later for text classification and information retrieval. The text embedding was created using the Google Universal Sentence Encoder model which was trained with a deep averaging network (DAN) encoder [30].

Furthermore, the interfaces of GPRO applications were selected from a collection of flows distributed in image files and integrated into the knowledge database. These images serve as visual representations of specific bioinformatics tools and their functionalities.

2.4. Data Annotation

All the collected datasets were annotated to intent tagging and entity identification. Annotations were performed by classifying the intent of each question in the database and annotating entities for keywords that distinguish the questions that have the same intent.

• Image annotation: Each region in the image of interest within the map images was semantically annotated to correspond to specific features or functionalities of GPRO applications. This process was performed by mapping and giving semantic links to specific image coordinates containing relevant information, such as the definition of the element that the parameter needs for the analysis for the chatbot, or by creating the link between visual elements and questions stored in the databases.

• Entity recognition: Named entity recognition (NER) was employed to identify and categorize domain-specific entities within user inputs by assigning the values of entities.

• Synonyms: A file was created to enhance the chatbot functionality. This file helps the chatbot understand and handle different terms used in questions or for specific entities, even if they mean the same thing. By recognizing synonyms, the chatbot can better grasp what users are asking or referring to, improving its ability to provide accurate responses. This approach expands the chatbot language comprehension, making it more versatile in interacting with users.

2.5. Model Architecture

The architecture of our chatbot is underpinned by the Rasa framework. Rasa has been chosen due to its flexibility, open-source nature, and strong support for dialogue management. The data for training the chatbot in Rasa were processed through Rasa NLU pipeline, which involves several stages: tokenization, intent classification, entity recognition, and response generation. We used the supervised_embeddings pipeline, which allows us to train custom machine learning models on our dataset. The dataset included annotated examples of user inputs linked to specific intents and entities, which were used to teach the chatbot how to interpret and respond to various inputs. The training process involved feeding the chatbot with labeled examples of user interactions, including diverse variations of queries and responses. These labeled examples were divided into training and test sets to ensure the model generalizes well to new inputs. The model learned to classify intents (such as greetings, questions, or commands) and extract entities (such as names, dates, or locations) from the training data. The key components of the chatbot model architecture are the following:

• NLU training data: Rasa uses YAML [31] as a unified and scalable format for organizing all types of training data that include:

  • -. NLU components to train the chatbot to parse and understand users’ inputs. These files contain examples of user messages categorized and classified by intents, which represent the user’s goals. They also contain entities, which are specific pieces of information within those messages.

  • -. Narrative flows including rules and stories, where rules describe short pieces of conversations that should always follow the same path, while stories allow for the representation of more complex and variable conversational sequences, enabling the chatbot to understand and adapt to dynamic user interactions.

  • -. Synonyms files.

  • -. Domain files which act as operational guidelines for the chatbot. These files list all possible user intents, bot responses, entities, and actions, essentially mapping out every interaction that the chatbot is able of handling. This setup ensures that the chatbot operates within its defined scope, by providing structured responses and actions based on the user’s needs and the context of the conversation.

• Dialogue management: Rasa dialogue management components, including dialogue policies, action selection, and custom actions. Custom actions use used to enable chatbot query data index or/and the XML files, the external database and the application programming interface (API). Custom actions allow a dynamic response generation for crafting coherent, contextually relevant, and customized responses to user queries.

• New features and components:

  • -. Bilingual component which facilitates seamless interactions in multiple languages, enhancing the accessibility and inclusivity of the chatbot. Leveraging Rasa’s extensibility, we created a language model that enables the chatbot to understand and respond to queries in both English and Spanish. The methodology involves training the model using bilingual datasets, adjusting language-specific nuances, and integrating it seamlessly into the existing Rasa architecture.

  • -. Custom API integration with LIMS for personalized interactions and streamlines access to user’s information. The API is meticulously designed to ensure secure and efficient communication between the chatbot and the external website.

• NLU pipeline: It comprises tokenization, intent recognition, and entity recognition stages. We employed the supervised machine learning capabilities of Rasa to train custom NLU models using the annotated dataset.

2.6. Challenges and Adaptations in Applying NLP and Rasa to Bioinformatics

In developing a chatbot tailored for bioinformatics, several unique challenges emerged that required specific adaptations of natural language processing (NLP) techniques and the Rasa framework to suit the complexity and specificity of the domain.

• Domain-Specific Terminology: Bioinformatics involves highly technical and specialized terminology, which is often outside the scope of standard NLP models. To address this, we had to extend the Rasa framework by customizing its intent classification and entity recognition systems. This involved integrating domain-specific knowledge from resources such as SEQanswers, BioStar, and the GATK forum into the training data, ensuring the chatbot could accurately understand and respond to queries involving specialized bioinformatics tools and workflows.

• Real-Time Updates and Novel Content: Bioinformatics is an ever-evolving field, with new tools and methods constantly emerging. Ensuring the chatbot remained current with the latest developments required incorporating up-to-date content from these platforms. To address this, we adapted the Rasa-X interface for continuous retraining, enabling users and domain experts to provide real-time feedback and updates, which were integrated back into the training pipeline without manual intervention.

• Handling Ambiguity in User Queries: Bioinformaticians often phrase their questions in complex or ambiguous ways, especially when dealing with multifaceted datasets. To tackle this, we used advanced entity recognition and developed custom fallback mechanisms to ensure the chatbot could gracefully handle unclear inputs by either requesting clarification or escalating to a human expert when necessary. This improvement allowed the chatbot to adapt dynamically, enhancing its reliability in real-world bioinformatics applications.

• Multi-Language Support and Flexibility: Given the global nature of bioinformatics, the chatbot needed to support multiple languages. By leveraging Rasa’s modular pipeline, we incorporated the Duckling Entity Extractor for date and time extraction in both English and Spanish, alongside a custom language detection component, enabling the chatbot to automatically detect and switch between languages based on the user’s input.

  Novel Techniques and Adaptations:

  • -. Custom Entity Extractors: To enhance the chatbot’s ability to handle bioinformatics-specific entities, we integrated a custom entity extraction system to recognize key terms and commands specific to bioinformatics workflows.

  • -. Flexible API Integration: We extended the Rasa framework to connect the chatbot with external bioinformatics tools and databases, allowing users to query multiple resources without leaving the chat interface.

2.7. Implementation and Development

The implementation and development of the chatbot was carried out in a controlled and well-defined environment to ensure consistency and reproducibility. The programming language was Python 3.9, and the integrated development environment (IDE) used was Visual Studio Code, which has the benefits of robust code editing capabilities, version control integration, and support for Python development. The version control was Git, facilitating collaborative development and code management among team members.

We chose Rasa as the framework for building the chatbot due to several key advantages that align with the specific needs of our project. First, Rasa-X offers a user-friendly interface for managing the chatbot, making it ideal for team members or partners who may not have a programming background. This interface allows supervisors to monitor conversations in real-time, identify when the chatbot misclassifies intents or cannot answer questions, and address those issues interactively. The ability to update the chatbot through an interface, rather than directly editing JSON files, makes it accessible for non-developers to refine and retrain the model, ensuring continual improvement. Additionally, Rasa’s modular architecture allows for easy integration of external libraries and features, such as bilingual support and widget integration for graphical interfaces. This flexibility was crucial in our project, where we required seamless integration of bilingual capabilities and ease of adding new functionality over time. Moreover, Rasa’s open-source nature enabled us to tailor both the NLU pipeline and the dialogue management policies to our specific needs, giving us more control over the chatbot’s behavior and making it easier to integrate with the GPRO and LIMS platforms. The supervised_embeddings pipeline for NLU allowed us to train custom models, while the Rasa core ensured that conversations were context-aware, offering dynamic and adaptive responses based on user interaction. We evaluated Rasa against other frameworks and chose it not at random, but because of its robust combination of modularity, customizability, and its interface through Rasa-X. These features provided significant advantages in chatbot training, deployment, and continuous improvement, especially for users who may lack technical expertise. This makes Rasa an effective choice for projects where technical and non-technical stakeholders collaborate.

Training and deployment were performed on a ThinkSystem SR530 running a Linux operating system, equipped with two Intel Xeon Silver processors, each featuring 12 cores at 2.2 GHz, and approximately 700 GB of RAM. Although we did not utilize all available resources, we observed that the training process used approximately six cores and no more than 150 GB of RAM.

2.8. Data Integration and Training

The chatbot intelligence is developed through a strict data integration and training process:

• Image integration: Map images that illustrate the interfaces of GPRO applications were created and integrated into the chatbot knowledge database. These images serve as visual representations of GPRO applications and their functionalities.

• Data integration: Annotated data created in the pre-processing phase were integrated into the Rasa framework, including training data for intent recognition, entity recognition, and dialogue management.

• Model training: We leveraged Rasa’s training capabilities to develop and fine-tune the machine learning models. This involved iterative training cycles to optimize intent classification, entity recognition, and dialogue policies.

2.9. Custom Actions and Responses

To enhance the chatbot conversational capabilities and domain relevance, custom actions and response generation techniques were implemented:

• Custom actions: Rasa’s custom action functionality was used to enable specific actions such as the generation of dynamic responses or fetching data from external sources such as XML documents, contextual database, interaction with other databases, and the performance of system-level operations.

• Custom components: This includes custom entity extraction components and a bilingual component. A bilingual component was introduced to enable the chatbot to handle users’ queries in different languages. Our current focus shifted towards ensuring the chatbot’s proficiency in both Spanish and English, leveraging the NLTK UDHR library [32] for this purpose. Our selection criteria for translation libraries have been heavily influenced by the need for high translation accuracy and robust language recognition to prevent the chatbot from misinterpreting the user’s intended language.

• Integration with other systems: To enable data exchange between the chatbot and other databases, such as LIMS databases, and external services, custom API integrations were developed to ensures real-time access to relevant information.

A user interface (UI) to interact with the chatbot was implemented in three alternative ways:

• A web-based interface.

• A user interaction mechanism: The chatbot interface is designed to allow users to interact with map images. Through an easy-to-use mechanism, users can easily click on areas of interest within an image, indicating their intent to inquire about specific functions depicted in GPRO applications.

• A GPRO interface: An interactive user interface implemented in each client application to query the chatbot directly from the application. The chatbot allows an API to modulate the chatbot engines, developed by using the Rasa framework.

Figure 3 summarizes the process of model training and user interaction throw the dialog manager. On the top of the figure, a description of the training process, starting from importing and preparing the data corpus for training, and a feedback loop to refine and re-annotate the data based on model evaluation and evaluator feedback. On the bottom of the figure, a general flow diagram of the dialog manager and how it retrieves user responses from the model and the databases is shown.

3. Results

In this work, we have introduced and described a Rasa-powered chatbot infrastructure specifically designed for the complex landscape of bioinformatics in both management and analytical tasks. The chatbot infrastructure offers a bilingual interface to seamlessly transition between English and Spanish. The architecture of the chatbots allows users to explore the complexities of the applications of the GPRO platform and LIMS, as well as to ask questions and receive insightful answers regarding the most common pipelines and workflows. Users can collect relevant information in real-time, enriching their bioinformatics endeavors. A notable step up in the chatbots capabilities include its interaction with map images. Users can effortlessly click on regions of interest within an image, prompting the GPRO chatbot to intuitively respond to questions regarding the specific interface of GPRO applications described. This feature not only enhances user engagement but also simplifies the exploration of complex bioinformatics tools. Bilingual proficiency, database connectivity, and image interaction capabilities lay the foundation for a more comprehensive, user-centered bioinformatics experience. The integration of the chatbot with Laboratory Information Management Systems (LIMS) represents a significant leap forward in streamlining laboratory operations and enhancing user engagement. In Figure 4, we show the integration of the chatbots within the bioinformatics suites.

Language understanding and response accuracy: The bilingual performance of the Rasa chatbot was evaluated with respect to language understanding across a set of queries in both English and Spanish. The language model, trained on bilingual datasets, showed strong capabilities in understanding query input. Response accuracy was consistently high, showing the effectiveness of the bilingual component implemented in facilitating smooth interactions in multiple languages.

Database connectivity and real-time responses: Integration with the databases have proven effective in enriching the knowledge base of the chatbot. Real-time communication allowed the chatbot to fetch the latest information from the database, ensuring that users received contextually relevant answers to their bioinformatics queries. API integration has facilitated efficient data flow, and the chatbot’s responsiveness in fetching information from the database of LIMS has been improved for a seamless user experience.

User interaction and feedback: The conversational interface of the chatbot facilitates intuitive interactions allowing users to have accurate and timely information. In this regard, Rasa-X is a valuable tool for gathering insights into questions that the chatbot could not answer. This feedback mechanism provides an accurate understanding of user queries and helps the identification of areas for further training and improvement.

Diversity and inclusion: Bilingual functionality has greatly increased the diversity and breadth of the chatbot. What sets this capability apart is the seamless language switching within the same session. Users can easily switch between languages by simply formulating their questions in either English or Spanish. The chatbot is able to handle and respond to queries in real-time, allowing for a smooth and dynamic chat experience.

Direct access to user-specific information: One of the key results of integrating the chatbot with LIMS is its newfound ability to fetch and display personalized information for users. Authorized users can perform a wide range of queries directly from the chatbot to retrieve or access specific information associated with their accounts, such as quotations or projects. This personalized interaction not only saves time but also increases user satisfaction by providing immediate access to relevant data.

Advanced filtering capabilities: Our results show the chatbot’s enhanced capability to understand and process complex queries, such as retrieving projects from a specific timeframe. The chatbot effectively interprets user requests for historical project data, by applying dynamic filters to return a concise list of relevant projects. This feature was particularly appreciated by users seeking quick overviews of their past activities without the need to manually sift through extensive databases.

Automated report generation: Another significant outcome of this integration is the chatbot ability to assist in report generation. Upon request, the chatbot presents users with a selection of similar projects, guiding them to their project of interest. It then provides a link to a detailed report of the selected project, which users can choose to save for future reference. This functionality not only enhances the decision-making process but also shows the chatbot’s role in facilitating a more efficient workflow.

Ensuring Accuracy and Reliability: The accuracy and reliability of the model were evaluated through several methods:

• Evaluation metrics: We monitored key metrics such as precision, recall, and F1-score for both intent classification and entity recognition [33]. These metrics helped measure the model ability to correctly identify user intents and extract entities while minimizing false positives and false negatives. The following metrics were used:

  • -. Precision: Measures how many times the chatbot correctly predicted the intent out of all the times it made a prediction for that intent.

  • -. Recall: Measures how often the chatbot correctly predicted the intent out of all possible true cases for that intent.

  • -. F1-Score: Provides a balance between precision and recall, offering a single metric for overall performance.

• Manual evaluation: Chatbot were manually tested by different individuals as manual curation step. Multiple cycles of testing and feedback were used to evaluate where the model struggled in classifying intents, allowing us to identify misclassifications. This feedback cycle was instrumental in fine-tuning the model, particularly by adjusting the data and re-labeling ambiguous examples.

• Rasa-X Feedback Loop: After deployment, we used Rasa-X to collect real-time feedback on the chatbot’s performance. When the chatbot failed to respond correctly, these instances were logged. This allowed us to continuously improve the model by retraining it with new examples that addressed the misclassified intents or entities.

Security and privacy measures: The security measures implemented in the chatbot, especially in the API connection with the LIMS database, ensured the privacy and confidentiality of user interactions.

Both Abu and Genie were tested by our partner interns, who are experienced bioinformatics researchers, frequently utilizing the GPRO tools. As active users with extensive bioinformatics expertise, they posed various domain-specific questions during the testing phase. The feedback they provided was highly positive, and they found the chatbot responses satisfactory for their needs. In particular, the performance metrics for Genie and Abu, which are displayed in Table 1 and Table 2, demonstrate the chatbot’s effectiveness in intent classification, meeting the expectations for both accuracy and relevance in handling bioinformatics-related queries.

The chatbots are available for use at the following web addresses:

• Genie: https://gpro.biotechvana.com/genie (accessed on 17 November 2024)

• Abu: https://biotechvana.com/

4. Conclusions

In this work, we have introduced a Rasa-powered oriented know-how for creation of chatbots, specifically designed to give support in the field of bioinformatics. Based on this knowledge, we have built two chatbots: one (called Genie) oriented to give support to the users of the GPRO platform and the other (Abu) to manage decision-making tasks in LIMS. The API integration facilitates direct interactions with both the GPRO applications and the LIMS, thus expanding the chatbot knowledge base. The chatbots offer a bilingual interface to seamlessly transition between English and Spanish and a robust architecture allowing users to ask questions and receive insightful answers with ease.

While we have demonstrated a successful implementation of the chatbots, ongoing efforts in analyzing user feedback, including insights from Rasa-X, and ongoing training iterations remain integral to future improvements. Our roadmap for the improvement and optimization of the chatbots includes integration of tools for voice interaction to enable users to interact with the chatbots through spoken queries, thus providing a more intuitive and accessible method of communication. The chatbot’s ability to respond by voice increases its versatility, catering to a wide range of user preferences, and enhancing the overall user experience. On the other hand, chatbots are constantly evolving. User feedback, along with the insights provided by Rasa-X, will guide us in refining and improving its performance.

Author Contributions

Conceptualization, A.A.E., A.I.H., C.L. and J.M.S.; methodology, A.A.E., A.I.H., A.G., C.L. and J.M.S.; software, A.A.E., A.I.H. and A.G.; validation, A.A.E., A.I.H., R.C., G.M., B.S. and C.L.; formal analysis, A.A.E., A.I.H. and J.M.S.; investigation, A.A.E., A.I.H., A.G. and J.M.S.; resources, G.M., B.S. and C.L.; data curation, A.A.E., A.I.H., R.C., G.M., B.S. and C.L.; writing—original draft preparation, A.A.E. and J.M.S.; writing—review and editing, A.A.E., A.I.H., C.L. and J.M.S.; visualization, A.A.E. and J.M.S.; supervision, A.I.H., C.L. and J.M.S.; project administration, C.L.; funding acquisition, C.L. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

Dataset available on request from the authors.

Conflicts of Interest

The authors declare no conflict of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Table 1

Performance metrics of Genie.

	Precision	Recall	F1-Score
Macro average	0.875	0.854	0.849
Weighted average	0.940	0.938	0.937

Table 2

Performance metrics of Abu.

	Precision	Recall	F1-Score
Macro average	0.875	0.85	0.846
Weighted average	0.936	0.935	0.934

References

1. Caldarini, G.; Jaf, S.; McGarry, K. A literature survey of recent advances in chatbots. Information; 2022; 13, 41. [DOI: https://dx.doi.org/10.3390/info13010041]

2. Bhirud, N.; Tataale, S.; Randive, S.; Nahar, S. A literature review on chatbots in healthcare domain. Int. J. Sci. Technol. Res.; 2019; 8, pp. 225-231.

3. Almansor, E.H.; Hussain, F.K. Survey on intelligent chatbots: State-of-the-art and future research directions. Complex, Intelligent, and Software Intensive Systems: Proceedings of the 13th International Conference on Complex, Intelligent, and Software Intensive Systems (CISIS-2019), Sydney, Australia, 3–5 July 2019; Springer: Berlin/Heidelberg, Germany, 2020; pp. 534-543.

4. Wollny, S.; Schneider, J.; Mitri, D.D.; Weidlich, J.; Rittberger, M.; Drachsler, H. Are we there yet?—A systematic literature review on chatbots in education. Front. Artif. Intell.; 2021; 4, 654924. [DOI: https://dx.doi.org/10.3389/frai.2021.654924] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34337392]

5. Goel, R.; Goswami, R.P.; Totlani, S.; Arora, P.; Bansal, R.; Vij, D. Machine learning based healthcare chatbot. Proceedings of the 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE); Greater Noida, India, 28–29 April 2022; pp. 188-192.

6. Ayanouz, S.; Abdelhakim, B.; Benhmed, M. A smart chatbot architecture based NLP and machine learning for health care assistance. Proceedings of the 3rd International Conference on Networking, Information Systems & Security; Marrakech, Morocco, 31 March–2 April 2020; Volume 78, pp. 1-6.

7. Chung, M.; Ko, E.; Joung, H.; Kim, S.J. Chatbot e-service and customer satisfaction regarding luxury brands. J. Bus. Res.; 2020; 117, pp. 587-595. [DOI: https://dx.doi.org/10.1016/j.jbusres.2018.10.004]

8. Jansom, A.; Srisangkhajorn, T.; Limarunothai, W. How chatbot e-services motivate communication credibility and lead to customer satisfaction: The perspective of Thai consumers in the apparel retailing context. Innov. Mark.; 2022; 18, 13. [DOI: https://dx.doi.org/10.21511/im.18(3).2022.02]

9. Bocklisch, T.; Faulkner, J.; Pawlowski, N.; Nichol, A. Rasa: Open source language understanding and dialogue management. arXiv; 2017; arXiv: 1712.05181

10. Boonstra, L. Definitive Guide to Conversational AI with Dialogflow and Google Cloud; Springer: Berlin/Heidelberg, Germany, 2021.

11. Luscombe, N.M.; Greenbaum, D.; Gerstein, M. What is bioinformatics? A proposed definition and overview of the field. Methods Inf. Med.; 2001; 40, pp. 346-358. [DOI: https://dx.doi.org/10.1055/s-0038-1634431]

12. Yadav, D.; Tanveer, A.; Malviya, N.; Yadav, S. Overview and Principles of Bioengineering: The Drivers of Omics Technologies. Omics Technologies and Bio-Engineering; Barh, D.; Azevedo, V. Academic Press: Cambridge, MA, USA, 2018; pp. 3-23.

13. De Sá, P.H.C.G.; Guimarães, L.C.; das Graças, D.A.; de Oliveira Veras, A.A.; Barh, D.; Azevedo, V.; da Silva, A.L.D.C.; Ramos, R.T. Next-Generation Sequencing and Data Analysis: Strategies, Tools, Pipelines and Protocols. Omics Technologies and Bio-Engineering; Barh, D.; Azevedo, V. Academic Press: Cambridge, MA, USA, 2018; pp. 191-207.

14. Ouzounis, C.A. Rise and Demise of Bioinformatics? Promise and Progress. PLoS Comput. Biol.; 2012; 8, e1002487. [DOI: https://dx.doi.org/10.1371/journal.pcbi.1002487]

15. Köster, J.; Rahmann, S. Snakemake - a scalable bioinformatics workflow engine. Bioinformatics; 2012; 28, pp. 2520-2522. [DOI: https://dx.doi.org/10.1093/bioinformatics/bts480]

16. De Oliveira Veras, A.A.; De Sá, P.H.C.G.; da Costa Pinheiro, K.; Barh, D.; Azevedo, V.; Ramos, R.T.J.; da Silva, A.L.D.C. Computational Techniques in Data Integration and Big Data Handling in Omics. Omics Technologies and Bio-Engineering; Barh, D.; Azevedo, V. Academic Press: Cambridge, MA, USA, 2018; pp. 209-222.

17. Dudley, J.T.; Butte, A.J. A Quick Guide for Developing Effective Bioinformatics Programming Skills. PLoS Comput. Biol.; 2009; 5, e1000589. [DOI: https://dx.doi.org/10.1371/journal.pcbi.1000589]

18. Kulkarni, P.; Frommolt, P. Challenges in the Setup of Large-scale Next-Generation Sequencing Analysis Workflows. Comput. Struct. Biotechnol. J.; 2017; 15, pp. 471-477. [DOI: https://dx.doi.org/10.1016/j.csbj.2017.10.001] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/29158876]

19. Li, J.; Schmieder, R.; Ward, R.M.; Delenick, J.; Olivares, E.C.; Mittelman, D. SEQanswers: An open access community for collaboratively decoding genomes. Bioinformatics; 2012; 28, pp. 1272-1273. [DOI: https://dx.doi.org/10.1093/bioinformatics/bts128]

20. McKenna, A.; Hanna, M.; Banks, E.; Sivachenko, A.; Cibulskis, K.; Kernytsky, A.; Garimella, K.; Altshuler, D.; Gabriel, S.; Daly, M. et al. The Genome Analysis Toolkit: A MapReduce framework for analyzing next-generation DNA sequencing data. Genome Res.; 2010; 20, pp. 1297-1303. [DOI: https://dx.doi.org/10.1101/gr.107524.110] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/20644199]

21. Parnell, L.D.; Lindenbaum, P.; Shameer, K.; Dall’Olio, G.M.; Swan, D.C.; Jensen, L.J.; Cockell, S.J.; Pedersen, B.S.; Mangan, M.E.; Miller, C.A. et al. BioStar: An online question & answer resource for the bioinformatics community. PLoS Comput. Biol.; 2011; 7.

22. Tommaso, P.D.; Chatzou, M.; Floden, E.W.; Barja, P.P.; Palumbo, E.; Notredame, C. Nextflow enables reproducible computational workflows. Nat. Biotechnol.; 2017; 35, pp. 316-319. [DOI: https://dx.doi.org/10.1038/nbt.3820] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/28398311]

23. Hafez, A.I.; Soriano, B.; Elsayed, A.A.; Futami, R.; Ceprian, R.; Ramos-Ruiz, R.; Martinez, G.; Roig, F.J.; Torres-Font, M.A.; Naya-Catala, F. et al. Client Applications and Server-Side Docker for Management of RNASeq and/or VariantSeq Workflows and Pipelines of the GPRO Suite. Genes; 2023; 14, 267. [DOI: https://dx.doi.org/10.3390/genes14020267]

24. Ezziane, Z. Applications of artificial intelligence in bioinformatics: A review. Expert Syst. Appl.; 2006; 30, pp. 2-10. [DOI: https://dx.doi.org/10.1016/j.eswa.2005.09.042]

25. Tang, B.; Pan, Z.; Yin, K.; Khateeb, A. Recent Advances of Deep Learning in Bioinformatics and Computational Biology. Front. Genet.; 2019; 10, 214. [DOI: https://dx.doi.org/10.3389/fgene.2019.00214]

26. Futami, R.; Muñoz-Pomer, A.; Viu, J.M.; Dominguez-Escriba, L.; Covelli, L.; Bernet, G.P.; Sempere, J.M.; Moya, A.; Llorens, C. GPRO: The professional tool for annotation, management and functional analysis of omic sequences and databases. Biotechvana Bioinform.; 2011; SOFT3.

27. Geer, L.Y.; Marchler-Bauer, A.; Geer, R.C.; Han, L.; He, J.; He, S.; Liu, C.; Shi, W.; Bryant, S.H. The NCBI BioSystems database. Nucleic Acids Res.; 2010; 38, pp. D492-D496. [DOI: https://dx.doi.org/10.1093/nar/gkp858]

28. Bruns, A.; Kornstadt, A.; Wichmann, D. Web Application Tests with Selenium. IEEE Softw.; 2009; 26, pp. 88-91. [DOI: https://dx.doi.org/10.1109/MS.2009.144]

29. Peng, D.; Cao, L.; Xu, W. Using JSON for Data Exchanging in Web Service Applications. J. Comput. Inf. Syst.; 2011; 7, pp. 5883-5890.

30. Cer, D.; Yang, Y.; Kong, S.; Hua, N.; Limtiaco, N.; John, R.S.; Constant, N.; Guajardo-Cespedes, M.; Yuan, S.; Tar, C. et al. Universal Sentence Encoder. arXiv; 2018; arXiv: 1803.11175

31. YAML. Available online: https://ymlthis.r-lib.org/ (accessed on 29 July 2024).

32. NLTK UDHR Corpus Reader. Available online: https://www.nltk.org/api/nltk.corpus.reader.udhr.html (accessed on 29 July 2024).

33. Powers, D.M.W. Evaluation: From precision, recall and F-measure to ROC, informedness, markedness and correlation. arXiv; 2020; arXiv: 2010.16061